_is_real_user_message no longer accepts a user-role compaction summary as
the human anchor, so _compress_context restores the original user turn
after the summary; update the lifecycle-status test's expectation.
Direct-Anthropic requests used a single shared _anthropic_client, and the
stale/interrupt watchdog closed + rebuilt it from the poll (stranger) thread
at four sites (non-streaming stale/interrupt, streaming stale/interrupt).
Closing a client whose TLS socket a worker thread was still reading released
the FD from a stranger thread; the kernel recycled it under a live SSL BIO,
which then wrote a 24-byte TLS record into an unrelated SQLite header
(cron/executions.db), bricking every cron on the profile. Same shape as the
OpenAI-only #29507 fix, but the Anthropic path never got the owner-thread
contract.
Extend the #29507 ownership contract to Anthropic: build a per-request client
(_create_request_anthropic_client), register it with the request-client
holder tagged by kind, and route _close_request_client_once by kind — a
stranger thread only shuts the request client's sockets down
(_abort_request_anthropic_client), while the owning worker performs the SDK
close (_close_request_anthropic_client). The shared _anthropic_client is now
never closed from inside a request (streaming or non-streaming), including the
worker retry-cleanup sites, since each attempt builds a fresh request client.
The #28161 no-hang guarantee is preserved: the poll-thread socket abort
unblocks the worker immediately.
Salvages the approach from #51688 (@raymondyan-zhijie), reimplemented onto
current main (non-streaming dispatch was refactored into
_dispatch_nonstreaming_api_request; streaming grew _cancel_current_stream_attempt
and worker retry-cleanup sites). Tests updated to the request-local mechanism
(incl. replacing a banned source-reading test with a behavior test) plus new
regression coverage proving the watchdog aborts the request client and never
touches the shared client.
Co-authored-by: raymondyan-zhijie <32435458+raymondyan-zhijie@users.noreply.github.com>
Second call site (non-streaming / gateway path) now flattens list-type
content with flatten_message_text before the inline <think> regex and the
surrogate sanitizer, matching the interim-text fix from the prior commit.
Adds regression tests (tests/run_agent/test_66267_multimodal_interim.py)
covering:
- build_assistant_message with list content does not raise TypeError
- inline <think> inside list content is extracted + stripped correctly
- _interim_assistant_visible_text is safe for tool messages (list content)
- duplicate_previous_interim dedup guards against tool messages
Verified the tests fail without the fix (TypeError: expected string... got
'list') and pass with it.
_bedrock_reasoning_stale_floor only matched dashed floor-table keys (opus
'claude-opus-4-6'), so sonnet reasoning models keyed with a dotted version in the
shared table ('claude-sonnet-4.5'/'4.6') got no reasoning floor from the dashed
Bedrock inference-profile id — premature stale-abort if the base timeout is set
below 180s. Generate both version-separator forms (digit-dash-digit <-> digit-dot-
digit, via lookbehind/lookahead so only version numbers flip) and try all
candidates; matches the table however each model is keyed, no edit to the shared
table. Verified: opus->240 (unchanged), sonnet-4.5/4.6->180, haiku->None. New
TestBedrockReasoningStaleFloor (8 cases).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Re-applied against v0.18.2. Upstream now covers the OpenAI-path stall watchdog
(cross-turn give-up breaker #58962) and the tool-batch deadline, so those are
dropped. Two gaps remain:
- Bedrock streaming had NO liveness watchdog and was excluded from the #58962
breaker (it returns before _check_stale_giveup). Add an on_event hook to
stream_converse_with_callbacks (fires per yielded event = true wire-level
liveness), drive a stale timer from it, and wire Bedrock INTO the existing
breaker: entry _check_stale_giveup, _bump_stale_streak on stall, raise to end
the call (invalidate_runtime_client can't abort the in-flight botocore stream,
so the streak escalates across turns like the OpenAI path), and reset the
streak on success.
- local providers still got float('inf') stale timeout (watchdog disabled) — give
a finite ceiling (HERMES_LOCAL_STREAM_STALE_TIMEOUT, default 900s).
- tests: on_event per-event + swallow; Bedrock stall bumps streak + aborts;
pre-elevated streak aborts at entry; success resets streak.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
strip_think_blocks() ran re.sub() directly on content that could be a
list of blocks (Anthropic via OpenRouter returns assistant content as
[{type:text,...},{type:thinking,...}]). A list reaching re.sub raised
'TypeError: expected string or bytes-like object, got list', which the
outer conversation loop swallowed and retried forever — the observed
infinite 'preparing terminal...' loop that re-emitted the same
assistant text every iteration.
The live-turn path normalized list content to a string, but
_interim_assistant_visible_text reads a *stored* history message whose
content was persisted as a list and passes it straight into the shared
strip_think_blocks helper. Fix at the shared choke point: coerce
list/dict content to visible text (dropping reasoning blocks, which is
the function's job) before any regex runs, so every caller is safe.
Fold #62349's broader provider-boundary handling into the header fix: bound top-level and xAI override keys again at preflight after middleware, preserve unrelated headers, and cover boundaries and collisions.
Co-authored-by: Nick Taylor <nicktaylor@TheWorldofNick-Lappy.local>
Widen the #65991 single-writer fence to run_codex_stream: each codex
attempt claims the delta sink before consuming events, and the consume
loop's interrupt_check now also stops the instant a newer attempt
supersedes this one. Parity with the chat_completions / anthropic /
bedrock paths from the salvaged fix.
Two regression tests: superseded codex stream is fenced mid-stream;
sole-writer codex stream delivers unchanged.
Assert that a superseded stream (older writer token, other thread) is fenced
from the delta sink, the active writer is never fenced, a non-claiming thread
is never treated as a writer, and the real consume loop stops the instant it is
superseded — so two streams can never interleave into one turn (#65991).
Port from anomalyco/opencode#36130: the Responses spec carries streaming
error details at the top level of the error frame, but the official OpenAI
SDK and several OpenAI-compatible proxies wrap them in an HTTP-style nested
envelope ({"type": "error", "error": {code, message, param}}).
_raise_stream_error only read top-level fields, so nested-envelope frames
collapsed to the generic 'stream emitted error event' placeholder with
code=None — the error classifier never saw the provider's real failure
reason, misrouting rate-limit / context-overflow / entitlement errors into
the generic retry path.
Top-level fields keep precedence; the envelope is a fallback. Null-tolerant
for spec-compliant frames with explicit nulls.
Follow-ups on top of @xxxigm's salvaged bridge (#33294):
- Remove the now-dead narrow item/started-only mapper from #38835
(_codex_note_to_tool_progress) — the full bridge supersedes it and
keeps the same tool-name contract; its tests are repointed at the
bridge helpers.
- Preserve main's request_routing/approval-bypass wiring on the
CodexAppServerSession constructor (landed after the PR was filed).
- Gate agentMessage interim delivery on display.show_commentary so the
app-server runtime honors the same toggle as the codex_responses
commentary channel (tool progress is unaffected).
- Add json import (bridge helpers use json.dumps) and modernize the
wiring test's stub agent for main's usage-accounting attributes.
Commentary delivery is on by default; users who find the extra mid-turn
narration noisy can set display.show_commentary: false to restore the
previous behavior (commentary routed to the reasoning channel, visible
only with show_reasoning).
- hermes_cli/config.py: display.show_commentary default true
- agent/agent_init.py: wire config -> agent.show_commentary
- run_agent.py: gate structured commentary extraction on the flag
- agent/codex_runtime.py: gate live-stream commentary callback (falls
back to legacy reasoning-channel routing when off)
- docs + 2 tests (interim path off, live stream fallback)
Also adds AUTHOR_MAP entries for davidrobertson and 100yenadmin.
Keep each xAI OAuth auth-add login as an independent manual device-code pool entry and recognize xAI personal-team spending-limit 403 responses as billing exhaustion. Preserve the structured top-level error message so the failed credential is quarantined and the next healthy account is selected without attempting a pointless token refresh.
Route direct xAI HTTP consumers through the credential pool as well. Proactive and 401-reactive refreshes update the exact issuing manual entry, preserve validated xAI base URL overrides, and serialize single-use refresh-token rotation across concurrent pool instances.
Add current-main regression coverage for both the registered provider
profile and core GitHub Responses path while leaving live catalog loading
to the complementary catalog-resolution work in #51953.
_extract_parallel_scope_path used Path.cwd() (process cwd) instead of the
tool's actual execution cwd, and os.path.abspath() instead of os.path.realpath(),
so symlink aliases and relative/absolute path pairs that resolve to the same
physical file were treated as distinct targets and placed in the same parallel
segment. On case-insensitive platforms (Windows) os.path.normcase() was also
absent, allowing Foo.txt and foo.txt to race.
Changes:
- agent/tool_dispatch_helpers.py: introduce _canonical_path(raw_path,
execution_cwd) applying expanduser->abspath->realpath->normcase; thread
execution_cwd through _extract_parallel_scope_path and
_plan_tool_batch_segments
- agent/tool_executor.py: pass get_active_env(effective_task_id).cwd as
execution_cwd to _plan_tool_batch_segments; add pathlib.Path import
- run_agent.py: pass active env cwd to _plan_tool_batch_segments at the
second call site inside _execute_tool_calls
- tests/run_agent/test_tool_batch_segmentation.py: add 5 regression tests
covering relative/absolute same target, symlink alias, execution_cwd vs
process cwd, symlink parent + nonexistent write target, and Windows
case-insensitive alias (skipped on non-Windows)
Fixes a file-corruption / lost-update race introduced by the mixed
tool-batch segmentation feature (perf commit #64460).
* feat(analytics): record auxiliary model usage per task in session accounting
Auxiliary LLM calls (vision, compression, title_generation, web_extract,
session_search, ...) discarded their token usage, leaving dashboard
analytics blind to aux model spend (issue #23270).
- hermes_state.py: session_model_usage gains a task PK dimension
(''=main loop) via v22 table-rebuild migration (SQLite can't alter a
PK); record_auxiliary_usage() writes per-(model,provider,task) deltas
WITHOUT touching sessions counters (gateway overwrites those with
absolute main-loop totals — folding aux in would double-count or be
clobbered). Aux rows never inherit the session's main-loop route.
- agent/aux_accounting.py: ContextVar ambient accounting context
(mirrors the portal_tags conversation context); record_aux_usage()
normalizes usage via usage_pricing.normalize_usage, estimates cost,
and is strictly best-effort. moa_reference/moa_aggregator excluded —
conversation_loop already folds MoA usage+cost into the main delta.
- agent/auxiliary_client.py: _validate_llm_response is the recording
chokepoint — every successful non-streaming aux response passes
through it exactly once, sync and async, including fallback paths
(model read from the response itself stays accurate across
fallbacks).
- run_agent.py: run_conversation publishes/resets the accounting
context; agent/title_generator.py republishes on its bare thread.
- hermes_cli/web_server.py: /api/analytics/usage folds aux rows into
by_model (aux-only models finally appear) and adds a by_task
summary; /api/analytics/models surfaces aux rows on the Models page.
Design per review of PR #62850 by @eeksock (thread-local + separate
auxiliary_usage table): rebuilt on ContextVar (async-safe — thread-local
cross-attributes concurrent coroutines on one event loop) and the
existing session_model_usage table instead of a parallel accounting
path, extended beyond vision to every aux task, and wired the analytics
endpoints so the dashboard actually shows it. Credit to @eeksock for
the approach and @tboatman for the detailed root-cause analysis.
* test(moa): match _validate_llm_response mock to new accounting-hint signature
* test(aux): accept accounting-hint kwargs in remaining _validate_llm_response mocks
The end-to-end _build_api_kwargs parity test asserted the Nous Portal
tags exactly equal the base two-tag list. With the per-session
conversation tag, a real agent (which has a session_id) now emits a
third `conversation=<session_id>` tag. Assert against
nous_portal_tags(session_id=agent.session_id) so the check stays exact.
Two consecutive incomplete assistant interims with identical visible
content (content + reasoning) are collapsed even when opaque provider
state (encrypted reasoning item ids, message item phases) drifts per
continuation — previously that drift defeated dedup and caused message
storms (#52711). The latest opaque payload is written onto the existing
message in place, so continuation replay still uses fresh provider
state.
Salvage note: the original PR also made the 3-retry continuation cap
cumulative per turn (no reset on progress). That half is intentionally
dropped — a legitimate long turn alternating incomplete/progress would
hard-fail at 3 cumulative, changing semantics beyond the reported bug.
Map Codex Responses status=incomplete with incomplete_details.reason=content_filter to finish_reason=content_filter so the existing refusal/fallback path runs instead of burning incomplete continuation attempts.
Community reports of GPT 'infinitely thinking' are usually a slow or
overloaded provider plus silent retry machinery: the CLI/TUI/Desktop
spinner shows a generic 'cogitating...' verb for the whole wait and the
gateway heartbeat says only 'Working — N min'.
Add AIAgent._emit_wait_notice(): rewrites the live spinner/status line
(thinking_callback → CLI prompt_toolkit widget, thinking.delta → TUI +
Desktop) and updates the activity tracker (included in the gateway's
'⏳ Working — N min' heartbeat). Wired at the four wait points:
- non-streaming wait loop: after 30s with no response, the line becomes
'⏳ waiting on <model> — Ns with no response yet (provider may be slow
or overloaded; auto-reconnect at Ns)'
- streaming wait loop: same explanation after 30s with no chunks,
including the long-thinking case
- TTFB / stale-stream kills: '⚠ no response from provider in Ns —
reconnecting...'
- Codex continuation retries: '↻ model returned reasoning with no final
answer — asking it to continue (n/3)' instead of silence
Notices are fail-open (display errors never break the wait loop) and
gateway sessions without a display callback still get the improved
activity description.
grok-4.x on the xAI /v1/responses surface sometimes ends a turn with only
reasoning items — no message output item, no tool calls — and those
reasoning items carry no encrypted_content. Two compounding problems:
1. The model occasionally emits its final answer INSIDE the reasoning
channel, delimited by grok's internal "<response>" tag. The answer
exists but is classified reasoning-only → finish_reason=incomplete.
2. An interim assistant message holding only plain-text reasoning replays
as nothing in _chat_messages_to_responses_input, so every continuation
request is byte-identical to the one that just failed. The model
deterministically repeats the reasoning-only response until the retry
budget is exhausted and the turn dies with "Codex response remained
incomplete after 3 continuation attempts".
Fixes:
- _normalize_codex_response (xai_responses only): salvage the
<response>-delimited tail from the reasoning text and promote it to
assistant content; the untagged prefix stays as thinking text.
- Codex-incomplete continuation path: when the interim message has
nothing the input converter will replay (no content, no encrypted
reasoning items, no message items), append a user-role nudge so the
retry actually differs and explicitly asks for the final answer /
pending tool call. Mirrors the existing _get_continuation_prompt
pattern used for length truncation.
Observed live with grok-4.20 on xai-oauth (2026-07-13); sibling of the
grok-composer web_search incomplete-loop fix in transports/codex.py.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Follow-up to the salvaged #64449: the status-trusting branch flipped
github_responses to 'stop' alongside unknown relays. Copilot fronts the
same OpenAI model family as codex_backend and shows the same
reasoning-only 'still thinking' degeneration, so it stays on the
continuation path. Only unrecognized (other:*) backends trust
response.status='completed' as terminal.
The chat_completions path raises EmptyStreamError when a stream yields
no chunks and no finish_reason (#64420). _call_anthropic() had no
equivalent, and the eventless failure surfaces differently by client:
- Real Anthropic SDK: no message_start means no final-message snapshot,
so get_final_message() raises a bare AssertionError — not in the
retry loop's transient set, so it burned no retries and surfaced raw.
- OpenAI-compat shims: may fabricate a contentless Message with no
stop_reason (or return None), flowing out as a 'successful' empty turn.
Track whether any stream event arrived and normalize both shapes to
EmptyStreamError, giving the anthropic path the same transient retry
budget in the shared _call() retry loop. A real completed response
always carries a stop_reason, so the guard cannot fire on legitimate
turns (including eventless mocks with stop_reason='end_turn').
Follow-up to #64420.
Follow-up to @michaelHMK's cherry-picked fix for #50892:
- agent/tool_dispatch_helpers.py: guard the segment planner's except-path
debug log against non-string arguments (the planner replaced the old
_should_parallelize_tool_batch body after #64460 and inherited the same
latent arguments[:200] slice).
- tests: the PR's executor-coercion hunks were dropped as redundant —
both executors now route through _parse_tool_arguments, which already
rejects null/non-object args with a structured error result instead of
coercing to {} (the 'we do not repair bad model outputs' contract).
Reworked the salvaged tests to pin the current behavior: None args are
rejected without dispatch, valid siblings still run, the planner treats
them as a barrier without raising, and the mainline run_conversation
path (which normalizes None to '{}' before dispatch) stays crash-free
under verbose logging.
A model response containing several parallel-safe reads plus one unsafe
tool used to lose ALL concurrency: _should_parallelize_tool_batch was
all-or-nothing, so a single barrier call (terminal, clarify, unknown
tool, malformed args) forced the entire batch onto the sequential path.
_plan_tool_batch_segments now splits the batch into ordered segments:
maximal contiguous runs of parallel-safe calls execute on the existing
concurrent path, barrier calls on the sequential path, strictly in the
model's emission order. Invariants preserved:
- one tool result per call, appended in emission order (segments are
contiguous, so no result reordering across a barrier)
- side-effect boundaries: no call starts before an earlier barrier ends
- overlapping file targets split into separate ordered parallel runs
- turn-end budget enforcement + /steer injection run exactly once per
batch (segment executors run with finalize=False; the segmented
dispatcher owns the whole-turn finalize)
- interrupt during segment k drains segments k+1..n with cancelled
results, keeping one result per tool_call_id
Homogeneous batches keep their original single-path dispatch (zero
behavior delta); _should_parallelize_tool_batch remains as a thin view
over the planner for existing callers and tests.
Add agent.reasoning_overrides dict to config.yaml. Users can now set
a reasoning_effort per model, overriding the global agent.reasoning_effort.
Example:
agent:
reasoning_effort: "medium" # global default
reasoning_overrides:
"openrouter/anthropic/claude-opus-4.5": "xhigh"
"openai/gpt-5": "low"
"claude-sonnet-4.6": "high" # bare model name also works
The helper is spelling-tolerant: override keys match regardless of
provider prefix or dots-vs-dashes normalization, so users can write
keys in any sensible form and they'll match.
Resolution priority:
1. Session-scoped /reasoning --session override (gateway only; unchanged)
2. Per-model override from agent.reasoning_overrides (spelling-tolerant)
3. Global agent.reasoning_effort (existing)
4. Provider default (unchanged)
Wired into:
- CLI startup (cli.py)
- Messaging gateway agent construction (gateway/run.py)
- Desktop/TUI _load_reasoning_config (tui_gateway/server.py)
- Cron job scheduler (cron/scheduler.py)
- /model mid-session switch (agent/agent_runtime_helpers.py)
+ _primary_runtime now tracks reasoning_config for correct fallback recovery
- Fallback activation (agent/chat_completion_helpers.py::try_activate_fallback)
+ Re-resolves reasoning_config for the fallback model (best-effort)
Closes#21256 (per-model reasoning_effort defaults).
Note: no hermes config set agent.reasoning_overrides.<model> support;
users edit the YAML directly. _set_nested splits on "." and would
corrupt model keys containing version dots.
Follow-up to the cherry-picked empty-user-turn drop: the placeholder
introduced in 8582f35d9 fired for whitespace-only STRING turns too
(content=' ' flattens to non-stripping text but isn't in the
(None, '', []) exclusion set), fabricating an attachment note for a turn
that carried nothing. Gate the placeholder on isinstance(content, list)
so only genuinely structured (e.g. image-only) turns get it; empty and
whitespace-only string turns now fall through to the drop path.
Edge cases verified: trailing empty user turn still ends the view on the
synthetic advisory marker; an all-empty transcript degenerates to [].
MoA's _reference_messages() unconditionally appended every user-role
message to the advisory view sent to reference models, even when the
message content was an empty string or a non-string/multimodal payload
that the text-extraction step flattens to "".
Strict providers (Kimi/Moonshot, and others that enforce non-empty user
content) reject such a message with:
400 Invalid request: the message at position N with role 'user'
must not be empty
Lenient providers (DeepSeek) accept it, so an identical rendered view
passes on one reference and 400s on another within the same fan-out —
the user sees "kimi doesn't support MoA" when the real cause is an empty
user turn leaking into the advisory transcript.
Skip empty user turns, mirroring the existing behavior for empty
assistant turns (which are already dropped when they carry no parts).
The end-on-user invariant is preserved: the synthetic advisory-request
user turn is still appended when the view would otherwise end on an
assistant turn.
Adds a regression test asserting the advisory view contains no empty
user turn and still ends on a user turn.
Review follow-up to the reasoning_config cache-parity fix:
- Only inherit the parent's reasoning_config when the fork runs on the
parent's model (not routed). On the routed aux path
(auxiliary.background_review.{provider,model}) the cache is cold
regardless, so parity buys nothing, and the parent's effort vocabulary
can be invalid for the routed model/provider: OpenRouter
extra_body.reasoning.effort is forwarded unclamped
(chat_completions.py) and codex_responses only maps max/ultra for
gpt-5.6 — an exotic parent effort routed to a strict provider could
400 the review. Mirrors the existing 'not _routed' gate on
_cached_system_prompt / session_start three lines below.
- Add a routed-path regression test asserting reasoning_config is
omitted from the fork kwargs when _resolve_review_runtime returns
routed=True.
- Extract the four copy-pasted recorder stubs in
test_background_review_cache_parity.py into a single
_make_recorder_class() factory so a new fork attribute needs one stub
edit, not four.
PR #17276 painstakingly pinned `_cached_system_prompt`, `session_start`,
`session_id`, and the toolset config on the background-review fork so its
outbound request body would byte-match the parent's and hit Anthropic's
exact-prefix cache. The contributor measured a ~26% end-to-end cost
reduction on Sonnet 4.5.
That optimization is currently being silently undone by a missing
`reasoning_config` kwarg. The fork's `AIAgent(...)` call omits it, so the
fork's `reasoning_config` defaults to `None`. `anthropic_adapter.build_anthropic_kwargs`
(line ~2165) then short-circuits the `thinking` / `output_config` block,
and the fork's request body lands in a DIFFERENT Anthropic cache namespace
from the parent's.
Result on the wire: 0 `cache_read_input_tokens`, full `cache_creation_input_tokens`
of the entire parent prefix — every single background review.
7 days of midagent.db traffic from one host running stock Hermes against
Anthropic Sonnet:
```
Background-review FIRST calls (the moment a review fork is born):
count = 68
cache_write tokens = 7,004,297
cache_read tokens = 1,016,335
Cost on Sonnet ($3.75/M write vs $0.30/M read):
Spent on these writes: $26.27
Cost if they had hit parent cache instead: $2.10
WASTED: $24.16 / week / user
```
That is from one user. Multiply by Hermes's installed base for the full
impact.
Tested against api.anthropic.com directly (see refs/api-tests/ in the
attached investigation repo if needed):
| pair | cache_r | cache_w |
|---------------------------------------------|---------|---------|
| parent fresh | 0 | 24,047 |
| parent same again | 24,047 | 0 |
| fork: appends 2 new tail msgs, thinking ON | 24,047 | 22 |
| fork: appends 2 new tail msgs, thinking OFF | 0 | 24,047 |
Same fork-shape request, only difference is `thinking`. With the fix,
the fork hits the parent's full prefix and only writes the delta
(the `Review the conversation above…` prompt block, ~3-5K tokens).
One line in `agent/background_review.py`: pass
`reasoning_config=getattr(agent, "reasoning_config", None)` to the
`AIAgent(...)` constructor of the review fork. A short comment block
above it explains why so the next person who reads this code doesn't
re-introduce the regression.
`tests/run_agent/test_background_review_cache_parity.py` already covers
the system-prompt / session-id / toolset-config parity contracts that
PR #17276 introduced. I added:
* a `reasoning_config` attribute to `_make_agent_stub` so the stub has
a non-None parent value the test can verify is propagated.
* `test_review_fork_inherits_parent_reasoning_config()` — asserts the
fork's `AIAgent(...)` kwargs carry the parent's `reasoning_config`.
Pre-fix this test fails with `None vs expected {'enabled': True, 'effort': 'medium'}`;
post-fix all 4 tests in the file pass.
```
$ python -m pytest tests/run_agent/test_background_review_cache_parity.py -v
test_review_fork_inherits_parent_cached_system_prompt PASSED
test_review_fork_pins_session_start_and_session_id PASSED
test_review_fork_inherits_parent_toolset_config PASSED
test_review_fork_inherits_parent_reasoning_config PASSED ← new
```
Also runs against the broader background-review test suite:
`test_background_review.py` (4), `test_background_review_summary.py` (8),
`test_background_review_toolset_restriction.py` (3) — 19/19 pass.
`agent/curator.py:1691` has the same omission for the umbrella-curation
fork, but curator's prompt is "curate all skills" — it shares no prefix
with any user conversation, so cache-parity is a non-issue there. Worth
auditing if the curator ever takes a parent conversation as input, but
not part of this PR.
The `agent/auxiliary_client.py:1006` `reasoning_config=None` hardcode is
intentional (title/summary one-shots on short prompts — per-call cost
of namespace flip is negligible) and is also out of scope.
Provider auto-detection (URL-based inference for Anthropic, OpenAI Codex,
and xAI endpoints) runs before credential-pool validation in AIAgent init,
but #63048 placed the pool validation before auto-detection. When the agent
is constructed with provider=None and a recognized endpoint URL, the pool
is validated against an empty provider identity and discarded, even though
auto-detection correctly resolves the provider moments later.
Fix: move the credential-pool validation block to after the URL-based
auto-detection chain. The pool is stored on the agent before
auto-detection; validation now checks the resolved provider and only
nullifies agent._credential_pool when the pool's scoped provider genuinely
doesn't match.
Regression test covers all three auto-detection paths:
- Anthropic (api.anthropic.com)
- OpenAI Codex (chatgpt.com/backend-api/codex)
- xAI (api.x.ai)
Fixes#63425.
Salvage of #63888. The original fix clears stale _last_content_with_tools
on substantive tool-only turns but doesn't clear _mute_post_response, which
a prior housekeeping turn may have set. This suppresses tool progress
output via _vprint until the no-tool-call branch resets it at line ~4834
— after all tools have finished executing.
Fix: also reset _mute_post_response = False when clearing stale fallback.
Added test: verify pure housekeeping turns (content + only housekeeping
tools) still set the fallback correctly — the original use case the
fallback was designed for.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
A cached _last_content_with_tools response from a housekeeping-only turn
could survive a later substantive tool-only turn. When the model returned
an empty response, Hermes incorrectly finalized the older housekeeping
narration instead of invoking the post-tool empty-response nudge.
Production impact: scheduled cron jobs could return early without completing
their actual work (e.g., daily report job returning a housekeeping message
instead of producing the report artifact).
Root cause: The fallback state was only updated when a turn had both
content AND tool_calls. A turn with tool_calls but empty visible content
would skip state updates entirely, leaving stale fallback state intact.
Fix: Classify tools in every tool-call turn (regardless of visible content).
When any tool is substantive (non-housekeeping), clear the older fallback state
before processing later empty responses. This prevents two-turn-old housekeeping
narration from being treated as if it belonged to the immediately preceding
substantive tool turn.
Regression test added: tests/run_agent/test_conversation_fallback_state.py
Fixes#63860
Cache-decorated turns (apply_anthropic_cache_control converts string
content to [{type: text, ..., cache_control}] lists — applied BEFORE the
MoA facade since the #57675 cache-cold fix) and multimodal turns
(text + image_url parts) flattened to empty strings in
_reference_messages, which only read str content. On turn 1 of a
provider:moa session with a Claude aggregator the references received a
single EMPTY user message: Anthropic-side providers 400'd ('messages: at
least one message is required') while tolerant models answered 'no user
request is present' (live incident Jul 14 2026, preset 'closed').
Fixes, in totality:
- _reference_messages: extract visible text via
agent/message_content.flatten_message_text for user/assistant/tool
turns (skips image parts, so no base64 leaks into the advisory view);
decorated and undecorated transcripts now produce a byte-identical
advisory view (advisor cache prefix stays stable).
- image-only user turns get a placeholder instead of an empty message
(Anthropic rejects empty text blocks) or a silently dropped turn
(would break user/assistant alternation).
- degenerate-case fallback flattens structured content too.
- _attach_reference_guidance: a decorated/multimodal trailing user turn
now receives the guidance as a NEW text part appended AFTER the
cache_control-marked part (cached prefix byte-stable) instead of
falling through to a second consecutive user message (strict providers
reject user/user).
- conversation_loop MoA injection: multimodal user turns get the MoA
context appended as a trailing text part instead of being dropped;
user_prompt for the one-shot path flattens content lists instead of
str()-ing them (which leaked base64 payloads into the prompt).
Live-verified on the 'closed' preset (real OpenRouter wire, 2 user
turns, tool loop): all 4 reference calls carry the full document +
rendered tool state, end on user, zero tool-role/tool_calls; advisor
cache_write 7968 then cache_read 5909+; aggregator cache_read
14880-15237 on iterations 2+.
Co-authored-by: bo.fu <bo.fu@meituan.com>
Salvage of #63862. is_output_cap_error() returns False for vLLM/LM Studio
error messages that contain 'prompt contains ... input tokens' (treated as
input-overflow signal). But parse_available_output_tokens_from_error() CAN
extract a valid available_tokens from those same messages. The
compression-disabled guard only checked is_output_cap_error(), so vLLM/LM
Studio users with compression off still got a terminal failure instead of
the max-tokens retry.
Fix: also exempt when parse_available_output_tokens_from_error() returns a
value — that function determines whether the retry path can actually handle
the error, so it's the right predicate for the exemption.
Added test: verify vLLM-format error with compression_disabled=False still
triggers the max-tokens retry path.
Co-authored-by: dmabry <dmabry@users.noreply.github.com>
Bundle Fireworks AI as a first-class BYOK provider across the CLI, web/TUI,
and desktop onboarding.
- New model-provider plugin with attribution headers (HTTP-Referer / X-Title)
so Fireworks can attribute Hermes traffic; PAYG-safe default aux + fallback
models (accounts/fireworks/models/...), IDs tracking fw-ai/fireconnect.
- Registered in CANONICAL_PROVIDERS so it appears in the CLI/web/TUI pickers.
- Alias wiring (fireworks-ai, fw) into both CLI resolvers.
- First-class wiring: OPTIONAL_ENV_VARS, HERMES_OVERLAYS (FIREWORKS_BASE_URL
override), doctor env hints. Live catalog + model_metadata are auto-derived.
- doctor: treat Fireworks' native slash-form IDs (accounts/fireworks/...) as
valid, not aggregator vendor prefixes, so it no longer tells Fireworks users
to switch to openrouter or drop the prefix.
- picker: plugin providers with no static curated list now lead with their
profile fallback_models, so the default is an agentic chat model instead of
whatever the live catalog returns first (Fireworks listed an image model,
flux-*, ahead of its chat models).
- Desktop onboarding: Fireworks as a RECOMMENDED hero card with the official
Fireworks logomark and a brand-purple badge, routing to the BYOK key form;
i18n in en/ja/zh/zh-hant.
- Tests: profile contract, first-class wiring (both resolvers, overlay, config,
doctor incl. the slash-form regression, aux headers, credentials), discovery
spot-check, and a live smoke test driven through the Hermes runtime.
Fire Pass (fpk_) support is coming soon; the future wiring is kept as a
commented-out scaffold in the plugin.
force_close_tcp_sockets stayed shutdown-only after #29507 to avoid
cross-thread FD recycle. That left CLOSED sockets unreclaimed when
httpx.close() skipped already-shutdown sockets under long-lived
gateways (~1 CLOSED fd / 6 min via proxy).
Add release_fds= for the owning-thread dispose path only; abort still
defaults to shutdown-only.
Only restore held verification text when the loop genuinely ends through budget exhaustion. Preserve later interrupts and failures, keep generated-summary fragment explanations, and add regression coverage for both contracts.
Treat intent-ack continuation text as non-final so last-turn exhaustion requests a real summary instead of surfacing a premature promise. Keep iteration-limit fallback text free of the abnormal-fragment explainer.
Track held-back verification responses explicitly so budget exhaustion returns the composed report without a second model call. Keep unrelated error and recovery exit reasons intact, preserve Kanban timeout accounting, and cover the real run_conversation paths.
* test: deflake CI and dev-machine flaky tests in bulk
Fixes ten distinct flake sources found by mining recent CI failures and
running the full suite on a dev machine with real user state:
CI-observed races:
- tests/conftest.py live-system guard: allow signal 0 (pure liveness
probe) through _guarded_kill/_guarded_killpg. psutil.pid_exists()
probes a just-killed grandchild reparented to init; the subtree check
fails for it and the guard RuntimeError'd
test_entire_tree_is_sigkilled_not_just_parent intermittently on
unrelated PRs.
Hermeticity flakes (fail on dev machines with real state, pass on CI):
- agent/coding_context.py: _marker_root() now skips the shared temp
root (tempfile.gettempdir()) like it skips $HOME — a stray
/tmp/package.json flipped every tmp_path test into the coding
posture (9 failures in test_coding_context.py).
- test_agent_guardrails.py: pin MAX_CONCURRENT_CHILDREN=3 via autouse
monkeypatch instead of freezing the user's real config value at
import time (import-time vs call-time config mismatch).
- test_web_tools_config.py: TestCheckWebApiKey now neutralizes the
ddgs package probe and registry providers — the optional ddgs
package in a dev venv lit up the fallback backend.
- test_credential_pool.py: block claude_code/hermes-oauth credential
autodiscovery in the two pool-merge tests that assert exact id
lists (a real ~/.claude/.credentials.json seeded an extra entry).
- test_modal_sandbox_fixes.py: clear _permanent_approved /
_session_approved — the user's real command_allowlist silently
approved the guard-escalation commands under test.
- test_setup_irc.py: stub prompt_checklist to select only the IRC row;
the non-TTY cancel fallback re-ran the real configured platforms'
interactive setup_fn, which hit input() under captured stdin.
- test_doctor.py: TestGitHubTokenCheck now patches the module-level
HERMES_HOME constant (the file's established pattern) instead of
only setenv — doctor was running PRAGMA integrity_check against the
real multi-GB state.db and blowing the 300s per-file budget.
Latent atexit-duplication (same _enter_buffered_busy class as #34217):
- test_undo_command.py: drop importlib.reload(tui_gateway.server) in
fixture teardown; reload re-registers the module's atexit hooks.
- test_session_platform_resolution.py: drop per-test reload of
tui_gateway.server; every resolver reads env at call time.
* test: sentinel model value in ignore-user-config fallback assertion
With HERMES_IGNORE_USER_CONFIG=1, load_cli_config() falls back to the
repo-root cli-config.yaml (untracked, gitignored). On a dev machine that
file can legitimately set the same popular model the test hardcoded
(anthropic/claude-sonnet-4.6), flipping the != assertion locally while
CI (no cli-config.yaml) stayed green. Use an impossible sentinel model
name instead.
Three tests for the #61099 salvage: OpenRouter attribution headers
present after switching to openrouter.ai, Kimi User-Agent sentinel
present after switching to api.kimi.com, and stale headers cleared
when switching to a provider with no URL-specific headers.
2/3 fail on unpatched main (DID NOT ATTACH), confirming the bug.
switch_model() unconditionally set agent.provider but only set
agent.base_url when the resolved value was truthy. When a real
provider change resolved an empty base_url (e.g. minimax after
copilot), the agent ended up with provider="minimax" but
base_url still pointing at api.githubcopilot.com. That incoherent
pair then got snapshotted into agent._primary_runtime, so it kept
re-applying on every subsequent turn via restore_primary_runtime()
until the process restarted.
try_activate_fallback() and _swap_credential() were audited and
confirmed unaffected: both always derive base_url from an actually
constructed client, never from a possibly-empty resolver hint.
Fix: when base_url is empty AND the provider is genuinely changing,
raise ValueError instead of silently keeping the old provider's URL.
This routes through switch_model()'s existing snapshot/rollback
path, and callers (tui_gateway/server.py's _apply_model_switch)
already catch and surface a clean "switch failed, staying on X"
message. Re-selecting the SAME provider with an empty base_url
(credential-only refresh) still keeps the current URL, unchanged.
Fixes#47828
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Trim the salvaged commit to its two still-valid conversions:
- ChatCompletionsTransport.convert_messages (copy-on-write sanitize)
- QwenProfile.prepare_messages (copy-on-write normalize + cache_control)
Dropped from the original PR:
- agent/prompt_caching.py selective-copy: superseded by #57229 which
already rewrites apply_anthropic_cache_control on current main.
- Gemma extra_content narrowing (_model_consumes_thought_signature
'gemini or gemma' -> 'gemini' only) + its two tests: unrelated
behavior change reverting deliberate e8c3ac2f5; belongs in its own
PR with its own justification if pursued.
Conflict resolution: preserved main's newer timestamp-stripping
(#47868) inside the copy-on-write path.
Review finding: the substring check ('kimi' or 'moonshot' in model)
under-matches bare release slugs like k2-thinking that the repo's
canonical _model_name_is_kimi_family matcher (anthropic_adapter.py)
already covers. Reuse it instead of a second ad-hoc matcher; add a
regression test for the bare-slug case.
Adapted from PR #26014's test file to the canonical seam
(tests/run_agent/test_anthropic_prompt_cache_policy.py _make_agent
helper) instead of a new top-level file with sys.path manipulation.
Covers: kimi-k2.6 + moonshot-v1 on OpenRouter (envelope layout),
kimi via Nous Portal, and the non-OpenRouter negative case.
- Return the boundary snapshot from
_launch_session_boundary_memory_flush as a local value instead of
staging it on self._session_boundary_snapshot. The instance-attr
handoff could leak (no memory manager configured) or mis-fire a
stale snapshot on a later /new if an exception hit between staging
and consumption. A local variable eliminates the class; the helper
also returns None when no memory manager is configured so
new_session takes the inline-switch path.
- Drop the now-dead session_id kwarg from commit_memory_session:
after the redesign no production caller passes it (gateway, TUI,
compression all use the default), and speculative params are
rejected per AGENTS.md. The explicit-old-session need is served by
cli.py's direct engine call + commit_session_boundary_async.
- Drop the dead providers snapshot in commit_session_boundary_async
(only the emptiness check used it).
- Tests updated accordingly (dead-kwarg test removed, snapshot
assertion now covered by return-value contract).
Phase-2 gates: 2a tests/cli 1048 passed + 6 memory files 137 passed;
2b programmatic live smoke 0.38ms non-blocking caller, end→switch→sync
ordering verified; 2c structured 4-angle review — no Criticals, these
warnings fixed.
When an API error carries an httpx.Response whose body was consumed via
iter_bytes() during streaming error handling (e.g. GeminiAPIError from
agent/gemini_native_adapter.py), accessing .text raises
httpx.ResponseNotRead. The secondary exception replaced the real,
already-computed provider error (429 free-tier quota guidance) with the
generic 'Attempted to access streaming response content' message on
every turn.
Guard the .text access so it degrades to an empty snippet and falls
through to the str(error) fallback, which carries the full original
message. Mirrors the existing guards in
agent/error_classifier.py::_extract_error_body() and
agent/gemini_native_adapter.py::gemini_http_error().
Fixes#59769
Salvaged from PR #59868 (guard + regression test); the unrelated
desktop Ctrl-C fix bundled in that PR was intentionally dropped and is
triaged separately.
Sessions on sub-512K-context models were spending most of their wall-clock
re-summarizing: the 50% trigger left too little post-compaction headroom
(the incompressible floor — system prompt, tool schemas, protected tail,
rolling summary — ate most of the reclaimed space), so compaction re-fired
every 1-2 turns. Three compounding defects fixed:
- Threshold floor: models with context windows below 512K now trigger at
>=75% of the window (raise-only — a higher configured value or per-model
autoraise like Codex gpt-5.5's 85% always wins). Re-derived on
update_model() in both directions.
- No max_tokens on the summary call: the summary budget is prompt guidance
only ("Target ~N tokens"). The wire cap truncated summaries mid-section
on the Anthropic Messages / NVIDIA NIM paths (thinking models burn the
cap on reasoning first), yielding truncated or thinking-only summaries
and compaction loops. Summary token ceiling lowered 12K -> 10K to keep
the guidance within the intended 1K-10K envelope.
- Reasoning traces excluded end-to-end: inline <think>/<reasoning> blocks
are now stripped from assistant content before serialization to the
summarizer, and from the summarizer's own output before the summary is
stored (previously a thinking summarizer model's trace was persisted in
_previous_summary and re-fed into every iterative update, compounding
bloat). Native reasoning fields were already excluded.
Verified E2E with real imports against a temp HERMES_HOME: threshold table
across 64K-1M windows, override interactions (user 0.85 wins, spark 0.70
raised, gpt-5.5 0.85 kept), full compress() round-trip with a thinking
summarizer, and wire-kwargs capture proving no max_tokens is sent.
Review findings on the salvaged #60332 breaker, fixed as follow-ups:
- restore_primary_runtime() now resets the streak (third provider-swap
path; without it a recovered primary was short-circuited before a
single attempt and could never be re-proven healthy except via /model).
- interruptible_api_call (non-streaming) now carries the same breaker
(guard at entry, bump on stale_call_kill, reset on success). Quiet-mode
/ subagent / headless sessions — the profile most like #58962's
unattended 494-failure session — take this path and had the identical
infinite stale-retry class.
- Partial-stream stub return now resets the streak (chunks were received,
provider demonstrably responsive).
- Consolidated the triple-duplicated counter arithmetic into shared
helpers (_stale_streak/_bump_stale_streak/_reset_stale_streak/
_check_stale_giveup) with one canonical comment block; error message
now says 'consecutive stale attempts' (the counter counts kills, not
turns — a single turn can produce several).
4 new tests (restore resets / no-op restore keeps latch / non-streaming
short-circuit / non-streaming success reset).
Follow-up for the salvaged #60332 circuit breaker. The breaker latches:
once the streak trips, interruptible_streaming_api_call raises before any
stream is attempted, so the on-success reset can never run again. The
error text tells the user to switch models and retry — but neither
switch_model() nor try_activate_fallback() cleared the streak, so a
freshly selected healthy provider kept short-circuiting forever (only
/new recovered), and the automatic fallback chain was wedged the same way.
Reset the streak at both swap sites (after a successful rebuild only;
rollback/exhaustion paths keep the latch). 4 tests.
A session wedged against an unresponsive OpenAI-compatible provider can hit the stale-stream detector on every turn and loop forever, burning the full 180s x retries each turn with no response. Issue #58962 reports 494 consecutive failures over 3+ days on a single session.
The streaming retry path already caps retries WITHIN a turn (HERMES_STREAM_RETRIES, default 2) but has no cross-turn cap. Once a session's conversation state makes every turn stale, it retries indefinitely across turns and never notifies the user.
Add a per-session consecutive-stale-stream counter on the agent:
- incremented on every stale-stream kill in the outer poll loop;
- reset to 0 only when a stream actually completes;
- when it reaches HERMES_STREAM_STALE_GIVEUP (default 5), the next turn aborts immediately with a clear, actionable RuntimeError instead of spending 180s x retries again.
This is distinct from the existing stale-stream work (local-provider hard ceiling #44938, backoff/parse-error #60031): those bound a single hung stream, while this bounds repeated cross-turn staleness and surfaces a user-visible error.
Adds tests/run_agent/test_stream_stale_circuit_breaker.py covering the short-circuit, the success-reset, and the increment.
Drop the Responses-API native compaction path and its opt-in umbrella
flag from the salvaged feature. On the Codex OAuth chat route Hermes
owns the message list and the summary compressor works (and stays
provider-portable — encrypted compaction items would lock the session
history to chatgpt.com and break /model switches and provider
fallback). On the app-server runtime (codex CLI/agent) the codex agent
owns the real thread context, so thread/compact/start is the only
mechanism that can actually shrink it (#36801) — that path is now the
default behavior for codex_app_server sessions, controlled by
compression.codex_app_server_auto (native|hermes|off), no umbrella
flag.
Removed: responses.compact() call path, codex_compaction_items replay/
persistence plumbing, codex_native_compaction + codex_responses_threshold
config keys, desktop settings fields, and their tests. Kept: everything
app-server (compact_thread(), compaction notifications, bookkeeping,
docs, tests) plus cache-busting keys for the surviving knobs.
The provider-mismatch guard now checks pool_provider and
current_provider != pool_provider. MagicMock.provider returns
a truthy child mock by default, which would trigger the guard
and skip the pool recovery tests. Set pool.provider='' explicitly.
When context.engine selects a plugin engine (e.g. LCM), the host
compression threshold — including the Codex gpt-5.5 50% -> 85%
autoraise — only configures the built-in ContextCompressor and never
reaches the plugin. The autoraise notice still fired, telling the user
auto-compaction was raised when nothing actually changed, and the
startup context-limit line printed the host percent next to the
engine's own threshold_tokens, contradicting itself.
- Clear _compression_threshold_autoraised when a plugin engine is
selected, suppressing both the CLI startup notice and the gateway
turn-1 replay via _compression_warning.
- Print the active engine's own threshold_percent in the startup
context-limit line so percent and token count agree.
- Built-in behavior is preserved, including the fallback path where a
configured engine fails to load and the built-in compressor takes
over.
Fixes#44439
gpt-5.3-codex-spark has a native 128K context window but the default
50% compaction trigger fires at ~64K, wasting half the usable window
before the session has accumulated enough turns to summarize
meaningfully. This raises the trigger to 70% (~90K) on the Codex OAuth
route only, leaving ~38K headroom for the summary and continued
conversation before the 128K hard limit.
The override is not gated by allow_codex_gpt55_autoraise because 128K
is the model's native window (unlike gpt-5.5's artificial 272K Codex
cap). Non-Codex routes are unaffected.
Also adds a boundary regression test verifying the short-session
scenario from the issue always yields a non-empty compressible window
(no silent context wipe).
ruff check --fix --select F541 . on current main. Pure prefix removals;
adjacent-string concatenations keep the f only on interpolating fragments.
No string content or live placeholder altered.
TestMcpParallelToolBatch seeded provenance under old-style
mcp_<server>_<tool> names, which no longer pass the
is_mcp_tool_parallel_safe() prefix gate after the naming change.
coerce_tool_args only repaired the outermost value, so JSON-encoded
*elements* of array properties (and nested object sub-fields) were left
as strings. Three core tools have array<object> schemas — todo.todos,
delegate_task.tasks, memory.operations — so a model emitting
{"todos": ["{...}"]} would pass raw JSON strings into the tool and fail
downstream on item["id"]/item["goal"] access.
Adds a schema-guided recursive pass (_normalize_json_strings_for_schema)
that parses JSON-string array items and nested object fields only when
the matching schema position expects an array/object, preserving
legitimate JSON-looking string fields (type: string).
Adapted from cline/cline#11803 to hermes-agent's existing coercion layer.
Follow-up to the salvaged #58696 (devatnull) + #41343 (annguyenNous)
commits: instead of fully suppressing commentary/analysis-phase stream
deltas, fire on_reasoning_delta so the CLI/gateway display them like
thinking text. Matches Codex CLI semantics where commentary is never
the turn's final answer, while keeping the narration visible in the
reasoning display. Adds devatnull to AUTHOR_MAP.
The custom ``httpx.HTTPTransport(socket_options=[SO_KEEPALIVE, ...])``
in ``_build_keepalive_http_client()`` was introduced to fix CLOSE-WAIT
socket accumulation on long-lived connections (#10324).
That approach broke streaming for providers behind reverse proxies
(OpenResty, Cloudflare, etc.) because the custom socket options
conflict with the proxy's chunked-transfer handling (#54049, #12952).
It also stripped TCP_NODELAY, stalling TLS handshakes and SSE encoding.
Narrow per-provider bypasses were added for Copilot (#50298), Codex
(#36623, #12953), but the root cause remained.
The fix moves connection lifecycle management from the socket layer to
the HTTP pool layer:
- ``httpx.Limits(keepalive_expiry=20.0)`` tells httpx to close idle
pooled connections at 20 s, before a reverse proxy's typical 30-60 s
timeout drops them and causes CLOSE-WAIT accumulation.
- The default httpx transport preserves OS TCP defaults (including
TCP_NODELAY), so TLS handshakes and SSE chunked encoding work
correctly.
- ``trust_env=False`` prevents httpx from double-dipping on env vars
(we handle proxy detection ourselves via ``_get_proxy_for_base_url``
which respects NO_PROXY).
- The Copilot host bypass (line 3632) is no longer needed since all
providers now use the same standard httpx.Client.
Closes#54049. Supersedes #12010, #36623, #12953, #50298.
Extend the pre_tool_call plugin hook return contract with a new directive:
{"action": "approve", "message": "why this needs human confirmation"}
Previously a pre_tool_call hook could only veto a tool call (action: block)
or allow it silently. It could not escalate to the existing human-approval
flow. This unlocks user-defined runtime approval rules on ANY tool (HTTP
writes, file writes to sensitive paths, email sends), enforced at runtime —
resolving #51221 as a pure plugin, with no core approval.py rule schema.
Mechanism:
- get_pre_tool_call_directive() returns (action, message) for block|approve;
get_pre_tool_call_block_message() kept as a block-only back-compat shim.
- resolve_pre_tool_block() is the single dispatch-site chokepoint: fetches
the directive and, for approve, invokes the human gate; fail-closed to a
block on denial, timeout, or gate exception. ALL FOUR tool-dispatch sites
now call it: tool_executor (concurrent + sequential), agent_runtime_helpers,
and model_tools.handle_function_call.
- request_tool_approval() escalates via the SAME machinery as Tier-2
dangerous commands: session/permanent allowlist, prompt_dangerous_approval
(CLI) / submit_pending (gateway), [o]nce/[s]ession/[a]lways/[d]eny,
timeout fail-closed, approvals.cron_mode for cron contexts.
Architecture: extracted the shared decision core into _run_approval_gate(),
called by BOTH check_dangerous_command() and request_tool_approval() so the
fail-closed / cron / gateway / yolo / persist policy lives in ONE place and
cannot drift. Fixed a latent divergence — the plugin path now honors --yolo.
Approval grain: [a]lways is keyed on tool_name + a hash of the reason (an
explicit plugin rule_key overrides), so distinct reasons on the same tool
persist independently instead of one 'always' blanketing the whole tool.
Non-interactive: cron honors approvals.cron_mode (parity with commands); any
other non-interactive non-gateway context fails CLOSED for the plugin path
(the command path keeps its historical fail-open default, unchanged).
No new config schema, no new env vars, no new hook events.
Strict providers (DeepSeek) reject a payload where the same tool_call_id
appears more than once with HTTP 400 'Duplicate value for tool_call_id'.
The issue was filed as an 'orphaned tool message' compression bug, but the
pasted error is a DUPLICATE tool_call_id — orphans are already handled on
main; duplicates were not. Reproduced live on main: both shapes leaked
through repair_message_sequence and sanitize_api_messages.
Two chokepoints, two shapes:
- repair_message_sequence: consume the id from known_tool_ids on first
match so a SECOND tool result reusing it falls into the drop branch
(duplicate tool-result shape). This is @Robinlovelace's kernel from
#55436 (applied manually — that PR was ~800 commits stale and bundled
an unrelated duplicate-DB-write change for #860, which is dropped here).
- sanitize_api_messages (final pre-API pass): add a dedup pass covering
BOTH (a) duplicate tool_calls sharing an id WITHIN one assistant message
(the message[6] shape) and (b) later tool result messages reusing an
already-seen id. #55436 covered neither of these at this chokepoint.
Tests: duplicate-tool-result dedup at both functions, duplicate-assistant-
tool_call-id collapse, and a negative control proving distinct ids are
never dropped (no over-dedup).
Credit: @Robinlovelace (#55436) for the repair_message_sequence dedup kernel.
Closes#58327.
repair_message_sequence Pass 1 registered only tc.get("id") when building
the set of known assistant tool_call ids, then matched tool results against
it by tool_call_id. In the Codex Responses format an assistant tool_call
carries both id (fc_...) and a distinct call_id (call_...); a tool result's
tool_call_id may be keyed on either depending on which builder produced it.
Registering only id made a valid tool result whose tool_call_id matched
call_id look orphaned, so the pass dropped it and left the assistant
tool_call unanswered -- producing HTTP 400 on strict providers (DeepSeek,
Kimi): 'Messages with role tool must be a response to a preceding message
with tool_calls'. Long-running sessions that persisted such a sequence were
permanently broken, re-sending the orphan every turn.
Register both id and call_id for each assistant tool_call so a result
matching either key is recognized, consistent with
AIAgent._get_tool_call_id_static and the compressor's _sanitize_tool_pairs.
Apply the same call_id||id precedence to the corrupted-args sanitizer's
existing-result scan / stub insertion, which had the identical mismatch.
Adds 3 regression tests covering the codex id!=call_id case (match on
call_id, match on only call_id, match on id when both present).
The salvaged max-output/pressure fix set conversation_history=None after
the new pre-API compaction. That is only correct for legacy session-
rotation. Under the default in-place compaction (compression.in_place:
True), archive_and_compact inserts the compacted rows into the session DB
directly without stamping them with the intrinsic persisted-marker, so a
subsequent flush with conversation_history=None re-appends them — doubling
the active context and retriggering compression (the early-persist
duplicate-row trap).
Use conversation_history_after_compression(agent, messages), matching the
two existing compaction sites (post-response should_compress and the
turn-prologue preflight), which returns None for rotation and
list(messages) for in-place so the compacted dicts are skipped by identity.
Adds a regression test with a real SessionDB + real archive_and_compact
that asserts the compacted summary row is persisted exactly once (fails
with 2 copies on the None variant).
Follow-up to the per-site strips from the review gate. The two copy-site
strips are correct but positional — a copy site added after the assembly
loops would re-leak _db_persisted into the child-session flush. Add a single
terminal sweep (_strip_persistence_markers) run once on the fully-assembled
compressed list so the invariant 'no compacted message leaves compress()
carrying a persistence marker' is structural, not dependent on copy-site order.
- agent/context_compressor.py: _strip_persistence_markers() called before
compress() returns; helper docstring notes the sweep is the authoritative guard
- tests/agent/test_context_compressor.py: structural regression — neuter the
per-site helper to a leaking copy, assert the terminal sweep still strips
- tests/run_agent/test_compression_persistence.py: pin the fixture assumption
behind the exact-equality row-count assertion
Shallow messages[i].copy() during context compression propagated the
_db_persisted marker from cached gateway incremental flushes into the
post-rotation compressed list. _flush_messages_to_session_db then skipped
every row when writing to the new child session, so gateway restarts
lost the compacted transcript (severe amnesia).
Strip the marker in _fresh_compaction_message_copy() and add regression
tests for rotation flush + compressor assembly.
Fixes#57491
Replace the loopback/PKCE-callback server and manual-paste fallback with
the RFC 8628 device-code flow as the only xAI Grok OAuth login path. The
flow works in headless/SSH/container sessions with no 127.0.0.1 listener,
shrinking the local attack surface.
- Poll the token endpoint with server-provided interval, honoring
slow_down and expires_in; store tokens with auth_mode
oauth_device_code.
- Adaptive proactive refresh skew for short-lived device-code JWTs;
rotated tokens sync back to auth.json, the global root store, and the
credential pool (no refresh-token replay).
- Clear source suppression on successful re-login (CLI + dashboard) and
drop the duplicate dashboard pool entry so exactly one seeded
device_code entry exists.
- Use the shared device_code source name for consistency with the
nous/codex device-code providers.
- Desktop: remove the loopback OAuth flow states and dead type variants;
pkce providers' sign-in URL selection is unchanged.
- Docs (EN + zh-Hans) rewritten for device-code login; drop the deleted
--manual-paste flag from documented commands.
Named providers / custom_providers entries in config.yaml now accept an
extra_headers dict scoped to that endpoint — for reverse proxies, API
gateways, and custom auth schemes (e.g. Cloudflare Access service tokens).
- hermes_cli/config.py: normalize extra_headers on provider entries
(_normalize_custom_provider_entry + providers-dict translation), add
get_custom_provider_extra_headers /
apply_custom_provider_extra_headers_to_client_kwargs helpers keyed on
base_url (case/trailing-slash insensitive, no substring bypass —
mirrors the TLS helpers)
- hermes_cli/runtime_provider.py: surface extra_headers in the resolved
runtime for named custom providers (providers dict, legacy
custom_providers list, and the credential-pool path)
- run_agent.py / agent/agent_init.py: merge per-provider extra_headers
onto the OpenAI client default_headers at construction and on every
_apply_client_headers_for_base_url re-application (credential swaps,
rebuilds), most-specific level wins; OpenAI-wire only (native
Anthropic/Bedrock scoped out)
- agent/auxiliary_client.py: accept model.extra_headers as an alias of
model.default_headers for the global variant
- cli-config.yaml.example: documented commented example
- Header values are treated as secrets and never logged
Salvaged from PR #3526 by @jneeee, reimplemented against current main.
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
Three CLI reliability fixes:
1. Interrupt reliability: chat() only re-queued the user's interrupt
message when the turn result carried interrupted=True. When the agent
thread raced past its last interrupt check (or finished) before the
interrupt landed, the message was silently dropped — and the stale
_interrupt_requested flag left on the agent instantly aborted the
NEXT turn. Un-acknowledged interrupt messages are now re-queued as
the next turn and the stale flag is cleared (only when the agent
thread actually exited). The clarify-race path also parks the message
in _pending_input instead of dropping it.
2. Slow exit (5+ min): stdlib ThreadPoolExecutor workers are non-daemon
and joined unconditionally by concurrent.futures' atexit hook — even
after shutdown(wait=False). One wedged tool worker (abandoned after
interrupt/timeout) held the process open forever. Promoted
async_delegation's daemon executor to a shared tools/daemon_pool
module and adopted it in tool_executor (concurrent tool batches),
memory_manager (background sync), delegate_tool (child timeout wrapper
+ batch fan-out), and skills_hub (source fan-out). Added a 30s exit
watchdog (HERMES_EXIT_WATCHDOG_S) armed at _run_cleanup start as a
backstop for wedged cleanup steps.
3. Exit jank: after prompt_toolkit tears down the input/status bars the
terminal sat silent for the whole cleanup window, looking hung. Print
'Shutting down… (finalizing session)' immediately at exit start.
E2E: live PTY interrupt of a foreground 'sleep 120' terminal tool now
aborts in ~1s and the typed message runs as the next turn; wedged-worker
+ wedged-cleanup subprocess exits in 5.8s (watchdog) instead of hanging.
Follow-up on the salvaged #56392 guard. The cherry-picked change matched
custom:<name> pool entries against the primary by raw base_url string
equality, which (a) can't disambiguate two named custom providers sharing
one gateway base_url and (b) left a latent bare-"custom" entry bypass.
Route the match through get_custom_provider_pool_key(rt[base_url]) compared
against the entry's custom:<name> key, mirroring the sibling guard in
recover_with_credential_pool. Use CUSTOM_POOL_PREFIX instead of the literal.
Add regression tests for the custom same-endpoint (swap) and cross-endpoint
(skip) branches, plus the plain-provider fallback-pool case from #56885.
* fix(streaming): handle completed responses with empty/None choices
The streaming fallback guard added in #55932 recognized a completed
response object only when its `choices` was a non-empty list. But an
adapter can return a completed response whose `choices` is `None` or an
empty list (an error / content-filter / terminal frame) — still a whole,
non-iterable response, not a token stream. Those shapes fell through to
`for chunk in stream` and crashed with
'types.SimpleNamespace' object is not iterable
which is exactly issue #55933 (MoA `openai-codex` aggregator on
TUI/Desktop, where a stream consumer forces the streaming path).
Broaden the guard to discriminate on the PRESENCE of a `choices`
attribute (a genuine provider Stream object exposes none), disable
streaming for the session, and return the completed object so the outer
loop's normal invalid-response validation handles empty/None choices via
its retry path instead of iterating.
Based on the diagnosis in #56525 by @spiky02plateau (that PR normalized
the MoA aggregator return with a one-shot chunk iterator; the common
text/tool-call crash was already fixed at this seam by #55932, so this
extends the existing guard to cover only the remaining empty/None-choices
gap).
Fixes#55933
* refactor(streaming): simplify empty-choices guard body and parametrize tests
Post-review cleanup (no behavior change):
- Inline the single-use `response_choices` local and drop the redundant
`if first_choice is not None else None` guard (getattr(None, ...) already
returns the default safely).
- Collapse the two near-identical empty/None-choices regression tests into
one `@pytest.mark.parametrize` case.
Mutation-verified: reverting the guard to the old non-empty-list condition
still makes both parametrized cases fail with the historical
'types.SimpleNamespace' object is not iterable.
---------
Co-authored-by: spiky02plateau <155588579+spiky02plateau@users.noreply.github.com>
Wire ssl_ca_cert and ssl_verify through custom_providers config and env
vars into the keepalive httpx client, fixing APIConnectionError against
mkcert/self-signed Ollama proxies behind HTTPS.
Self-review follow-up on the salvaged approval-routing fix.
The initial adaptation re-read os.getenv("HERMES_YOLO_MODE") at session-build
time. That diverges from the repo's security invariant: HERMES_YOLO_MODE is
frozen into tools.approval._YOLO_MODE_FROZEN at import time precisely so a skill
running mid-process cannot set the env var and instantly flip the approval
bypass (a prompt-injection escalation path). A live re-read re-opened that hole
for the codex routing path.
- Add tools.approval.is_approval_bypass_active() — the canonical three-source
bypass check (frozen --yolo/HERMES_YOLO_MODE + session /yolo + approvals.mode
off) in one place. This is the 4th inline copy of that OR-chain (the three
sites in approval.py and tui_gateway/server.py:3121 all use the same idiom);
the helper is the shared chokepoint they can collapse onto.
- codex_runtime.py now calls is_approval_bypass_active() instead of the
hand-rolled mode-or-session check plus a runtime env re-read.
- Update the env-yolo test to patch _YOLO_MODE_FROZEN (the canonical test
pattern, e.g. tests/tools/test_yolo_mode.py) rather than setenv, which is
dead-on-arrival against the frozen constant.
Fail-closed default preserved on every branch; 28 integration + 77 session/yolo
tests pass; E2E confirms the real exec decision flips decline->accept only when
bypass is active.
On gateway/cron/non-CLI contexts the codex app-server runtime has no UI to
surface codex's exec/apply_patch approval requests, so they fail closed
(silently decline) — the bot appears responsive but cannot write files, with
no approval prompt anywhere ("patch rejected by user").
When the user has explicitly opted out of Hermes approvals (approvals.mode: off,
the /yolo session toggle, or HERMES_YOLO_MODE=1), collapse to codex's own
sandbox permission profile (~/.codex/config.toml) as the policy gate by passing
_ServerRequestRouting(auto_approve_exec=True, auto_approve_apply_patch=True) to
the session. Defaults (manual/smart/unset) preserve the current fail-closed
behavior — a no-op for users who have not opted out.
Reads the mode via the canonical tools.approval._get_approval_mode() (which
already normalizes the YAML-1.1 bare-'off'->False case) at session-build time,
so a mid-session /yolo toggle is honored too.
5 integration tests: each opt-out mechanism (config off, YAML False, env var,
session yolo) plus the default fail-closed regression guard.
Closes#26530
Co-authored-by: snav <jake@nousresearch.com>
test_references_run_in_parallel asserted elapsed < 0.9 for two 0.5s
sleeps that run concurrently. On a loaded CI runner, thread-pool
startup pushed the wall time to 0.9001s — a 0.14ms miss — flaking the
shard. Loosen to < 0.95, which still sits well below the 1.0s serial
floor, so a genuine serialization regression (>=1.0s) still fails hard.
- Track auth store source path on Nous state reads and write rotated
OAuth refresh tokens back to the same store, preventing stale-token
replays when Hermes falls back to a global/root auth.json.
- Skip Nous fallback entries locally when no access/refresh token is
present, suppressing repeated failed resolution attempts within a
session.
- Sync session model metadata after fallback switches so the gateway
DB reflects the backend that actually served the latest turn.
The persist user-message override was applied in place to the live messages
list. On the early crash-resilience persist (which runs BEFORE api_messages is
built), that stripped observed group-chat context off the live user message and
silently dropped it when observe_unmentioned_group_messages was enabled.
Fix at the single chokepoint: _flush_messages_to_session_db resolves the
override (idx/content/timestamp) locally and applies it ONLY to the row written
to the DB — the live dict is never mutated, so EVERY persist caller (early
persist, mid tool-loop flush, /resume, /branch) is protected uniformly. This
supersedes the earlier shallow-copy approach, which broke the intrinsic
_DB_PERSISTED_MARKER idempotency (copies never propagated the marker back to
the live dicts → duplicate rows) and closes the sibling class tracked in #56303.
Trailing empty-response scaffolding is still dropped from the live list in
_persist_session (unchanged behavior).
Salvaged from #48817; chokepoint reworked to coexist with the marker-based
dedup (#50372).
Co-authored-by: kyssta-exe <kyssta-exe@users.noreply.github.com>
Phase 2c review follow-up on the id()-reuse persistence fix:
- test_recycled_id_in_dedup_set_still_persists_new_message seeded an EMPTY
dedup set, so it never injected a collision and passed under id-based dedup
too (couldn't distinguish the designs). Replace with
test_stale_seed_id_from_prior_flush_cannot_suppress_new_message, which asserts
the durable invariant: the seed is empty after every flush (mutation-checked:
removing the post-flush reset now fails BOTH id-reuse tests).
- Refresh the _flush_messages_to_session_db docstring: it still described the
old per-session identity tracking; document the intrinsic-marker mechanism,
that _flushed_db_message_ids is now a one-shot seed, and the shared-dict
mutation safety note.
_flush_messages_to_session_db deduped persisted messages with a retained
{id(msg)} set (_flushed_db_message_ids) kept across turns. Once a flushed dict
is dropped from the live list (scaffolding rewind / in-place compaction) and
GC'd, CPython recycles its address onto a new assistant/tool dict whose id()
collides with the stale entry — so the real turn is silently never written to
state.db.
Replace the retained id-set with an intrinsic _DB_PERSISTED_MARKER stamped on
each dict. The id-set is demoted to a one-shot seed (valid only while the
caller's objects are alive) that is translated to markers and cleared after
every flush, so no id() outlives a flush to alias a future message. The marker
is _-prefixed so the wire sanitizers strip it before any request leaves.
Preserves the existing _is_ephemeral_scaffolding skip. Salvaged from #50372.
Co-authored-by: rrevenanttt <290873280+rrevenanttt@users.noreply.github.com>
When text compression can't reduce a 413 request further, evict base64
image parts from tool messages and retry once instead of dead-ending
with 'Payload too large and cannot compress further.'
A 413 is a request-body byte-size limit, not a token limit. browser_vision
screenshots (2-5MB base64 each) keep the HTTP body oversized even after
aggressive summarization. The strip pass passes remember_model=False so a
413 does not poison _no_list_tool_content_models — that set is for providers
that reject list-type tool content, a distinct failure mode.
Cherry-picked from #47397 by Tranquil-Flow; placed onto main's current
token-aware 413 recovery else branch.
Review follow-up on the concurrent-tool deadline salvage. timed_out_indices is
snapshotted from not_done at the deadline; a worker can still finish and write
results[i] in the window before the post-execution result loop reads it. The
loop unconditionally replaced results[i] with a fabricated 'timed out' message
for any snapshotted index, discarding a genuinely-successful (just-late) result.
Gate the timeout message on 'and r is None' so a real result always wins. Add a
regression test that forces the snapshot-vs-result-loop race deterministically
(mutation-checked: reverting the guard fails it). Also document the intentional
detached-worker leak at the executor abandon site.
A tool with no internal interrupt check (read_file, web_search, or a wedged
terminal backend) that never returns keeps the concurrent-tool poll loop alive
forever: the loop only breaks when all futures finish or an interrupt is
requested, and the 30s heartbeat resets the gateway idle monitor so idle-kill
never fires. The ThreadPoolExecutor was also used as a context manager, so its
__exit__ joined the hung worker with wait=True.
Add a wall-clock batch deadline (HERMES_CONCURRENT_TOOL_TIMEOUT_S, default 420s
— above the 360s web_extract timeout; 0/negative disables). When it fires:
cancel pending futures, signal an interrupt to the worker threads, abandon the
executor (shutdown wait=False, cancel_futures=True) so hung threads aren't
joined, and return a per-tool 'timed out' result for the unfinished calls while
still surfacing the finished ones. Also fixes the latent futures.index(f)
lookup (ambiguous with duplicate futures) by tracking a future->index map.
Salvaged from #54562.
Co-authored-by: Gustavo Mendes <87918773+gustavosmendes@users.noreply.github.com>
AIAgent.run_conversation() promises a dict with final_response, but 16
terminal-failure branches returned dicts that either omitted the key or
set it to None. Callers that index result['final_response'] directly
(run_agent.py chat() + the __main__ printer) turn a real provider/context
failure into an opaque KeyError instead of surfacing the actionable error.
Every offending branch already carried usable 'error' text, so this
mirrors that text into final_response for all 16 sites (8 that omitted the
key, 8 that returned None). Adds an AST regression test that fails if any
run_conversation() dict return omits final_response or sets it to a literal
None, and tightens the invalid-response test to assert final_response == error.
The MoA aggregator received the per-turn reference block merged into the most
recent `user` message. In an agentic tool loop that message is the original
task near the top of the context (everything after it is assistant/tool turns),
so injecting text that changes every iteration diverges the prompt prefix early.
The server's KV cache then cannot be reused and the entire conversation
re-prefills on every tool-loop step — full prefill each step, which dominates
latency on long contexts.
Append the reference block at the end of the prompt instead (merging into the
last message only when it is already a trailing user turn, i.e. plain chat).
This keeps the [system][task][tool-history] prefix stable and cache-reusable so
only the new block re-prefills, and gives the aggregator the references with
recency. Extracted as `_attach_reference_guidance` with unit tests.
Measured on a local llama.cpp aggregator over a long agentic task: KV-cache
reuse on follow-up steps went from ~0.3% to ~93-95% and per-step prefill on an
~80k-token context dropped from ~44s to <1s, with no change to output.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Ephemeral empty-response/prefill recovery scaffolding (the synthetic
assistant "(empty)" turn, the user nudge, the terminal "(empty)"
sentinel, and the thinking-only prefill placeholder) exists only to
drive the next API retry; the in-memory loop pops it before appending
the real response. The append-only flush did not mirror that, so a
mid-turn persist could commit scaffolding to the SQLite session store
(and JSON log), and a resumed session would replay synthetic
"(empty)"/nudge turns as genuine context — re-poisoning the empty-retry
boundary forever.
Filter ephemeral scaffolding at both durable-write sites
(_flush_messages_to_session_db + _save_session_log), by flag not
position, so buried scaffolding (an answered nudge leaves the synthetic
pair mid-list) is skipped too. Covers all three flags including
_thinking_prefill.
Adapted onto current main's identity-tracking flush.
Cherry-picked from #41281 by petrichor-op.
Whole-bug-class follow-up to the tui_gateway fix: the same -1
last_prompt_tokens sentinel (parked by conversation_compression after a
compression) leaked into other status readers, producing a raw -1 or a
NEGATIVE usage_percent on the transitional turn:
- agent/context_engine.py get_status() (the ABC default every external
context engine inherits) — highest blast radius
- gateway/slash_commands.py /usage context line
- cli.py session usage printout
All clamped to >=0, mirroring cli.py _get_status_bar_snapshot and the
tui_gateway fix. Adds an ABC get_status sentinel-clamp regression test.
The gateway/API server rebuilds the in-memory TodoStore by replaying
caller-supplied conversation_history. _hydrate_todo_store previously
accepted any role:tool message containing a "todos" array, so a forged
bare tool result could seed arbitrary todo state and re-inflate context
every turn (GHSA-5g4g-6jrg-mw3g).
Restrict hydration to tool results paired with an earlier assistant
todo tool call (matching tool_call_id, function name == todo, no
user/system boundary between). Reuse the existing _get_tool_call_id/
name_static helpers so dict- and object-shaped tool calls both work.
Add a generous MAX_TODO_RESULT_CHARS payload guard to drop absurd
forged results before parsing; item/content caps already exist on main.
Co-authored-by: Hermes Agent <agent@nousresearch.com>
@janrenz's PR #35862 added prompt_caching.enabled=false at init only. But
_anthropic_prompt_cache_policy re-derives _use_prompt_caching on every /model
switch (agent_runtime_helpers) and fallback-model swap (chat_completion_helpers),
which re-enabled markers and re-broke the strict proxy the toggle was meant to fix.
Move the kill switch into anthropic_prompt_cache_policy so it returns (False, False)
on every path. Drop the now-redundant init-time override (kept @janrenz's isinstance
hardening on the cache_ttl read). Add policy-level tests + docs for the toggle.
Follow-up to salvaged PR #35862.
Adds moa.save_traces (default off). When on, every MoA turn that runs the
reference fan-out appends one JSON line to
<hermes_home>/moa-traces/<session_id>.jsonl capturing the TRUE FULL turn:
each reference model's exact input messages (system advisory prompt + full
advisory view, not the truncated display preview) + full output + usage +
per-advisor cost, and the aggregator's exact input (including the injected
reference-context guidance block) + output. Lets MoA runs be audited and
improved offline — what every model saw, said, and cost.
- agent/moa_trace.py: config-gated JSONL writer, profile-aware path via
get_hermes_home(), best-effort (never breaks a turn), moa.trace_dir override.
- agent/moa_loop.py: _RefAccounting now carries full input/output/model/
provider/temperature; create() stashes the full turn on a cache MISS
(once per turn, never on the cache-HIT repeat iterations); non-streaming
aggregator output captured inline, streaming marked + pointed at the
session assistant message. consume_and_save_trace(session_id) flushes it.
- agent/conversation_loop.py: flushes the trace with the live session_id
right after MoA usage consumption. No-op for non-MoA clients.
- hermes_cli/config.py: moa.save_traces + moa.trace_dir defaults.
Traces are a side channel — NOT the messages table, never in replay, safe
to delete. Off by default; only overhead when off is one config read on a
MoA cache-MISS turn.
Tests: full-trace-when-enabled (per-ref input+output+cost, aggregator
input-with-guidance + output), nothing-when-disabled. Live E2E through
run_conversation confirmed the loop wiring writes the file.
MoA ran the reference models before the aggregator but returned only the
aggregator's usage to the loop — _run_reference discarded each advisor
response's .usage entirely. Session accounting (state.db, /insights, cost)
therefore undercounted every MoA turn by the whole reference fan-out, which
is usually the bulk of the spend and scales with advisor count.
- _run_reference normalizes each advisor's usage with ITS OWN resolved
provider/api_mode and prices it at ITS OWN model rate (correct cache-read/
cache-write split), returning a _RefAccounting(usage, cost).
- create() sums advisor usage + cost once per turn (cache MISS only, so a
repeat tool-iteration reusing cached advice does not double-charge) and
exposes it via MoAClient.consume_reference_usage().
- conversation_loop folds advisor tokens into the reported/persisted token
counts and adds advisor cost (priced per-advisor) on top of the
aggregator cost, in both the in-memory session totals and the state.db
per-call delta. Aggregator cost is still priced on aggregator-only usage
so advisor tokens are never repriced at the aggregator rate.
- CanonicalUsage gains __add__ for per-bucket summing.
Tests: advisor usage/cost capture, per-turn sum + consume-clears +
cache-hit no-double-charge, CanonicalUsage.__add__.
_slot_runtime maintained a hand-listed name-preservation set
({nous, anthropic, openai-codex, xai-oauth, bedrock}) that returned bare
provider+model to avoid call_llm collapsing an explicit base_url to the generic
'custom' route. That duplicated _resolve_task_provider_model's
_preserve_provider_with_base_url guard (a provider-catalog capability check)
and had to be extended by hand for every provider with custom auth/signing —
the exact drift that produced the anthropic (#54609) and bedrock (#54912) 429/
empty-response bugs.
Removes the whitelist: _slot_runtime now forwards the resolved base_url/api_key/
api_mode for every slot, and the single chokepoint
(_resolve_task_provider_model -> _preserve_provider_with_base_url) decides
identity preservation. Behavior is unchanged for the five providers — their
provider branches (codex Responses+Cloudflare, xai-oauth, bedrock SigV4,
anthropic OAuth Bearer+anthropic-beta, nous Portal tags) re-resolve their own
credentials by name and ignore a forwarded base_url/api_key, so forwarding is
safe even for bedrock's placeholder 'aws-sdk' key.
Verified via real-import E2E: _slot_runtime -> _resolve_task_provider_model
preserves openai-codex/xai-oauth/bedrock/anthropic/nous (+openrouter control) —
none collapse to custom. Tests updated to assert the pipeline invariant against
the real resolver instead of the removed whitelist's bare-return shape.
#54609 moves anthropic into the _slot_runtime name-preservation set (it must
NOT forward base_url/api_key — OAuth sk-ant-oat* needs the provider branch's
Bearer + anthropic-beta header). The pre-existing parametrized
test_moa_provider_backed_slot_survives_aux_resolution still listed anthropic
asserting the forward path, contradicting the new behavior. anthropic is now
covered by test_slot_runtime_anthropic_oauth_routes_through_provider_branch;
drop it from the forward-path parametrize (minimax-oauth/qwen-oauth remain).
MoA's _slot_runtime() whitelists providers that must keep their provider
identity (so call_llm runs their provider branch) instead of being treated
as a plain custom endpoint via forwarded base_url/api_key. Native anthropic
was missing from this set.
Native anthropic subscription OAuth setup-tokens (sk-ant-oat*) require Bearer
auth plus the 'anthropic-beta: oauth-*' header, which only the anthropic
provider branch adds. Without the whitelist entry, the slot's base_url/api_key
were forwarded and call_llm sent the OAuth token as x-api-key, which Anthropic
rejects with a bare 429 (rate_limit_error with no quota details). This made
anthropic references in MoA presets fail every time.
Add 'anthropic' to the whitelist so native anthropic reference/aggregator
slots route through the provider branch. Extends upstream 9229d0db1 which
added 'nous' for the same reason.
When the summary LLM hits a 429/transient failure, _generate_summary() sets
a cooldown and returns None; compress() inserts a static fallback marker and
returns. Tokens stay above threshold, so should_compress() kept returning
True and every subsequent agent turn re-fired _compress_context() — the CLI
appeared frozen until the cooldown expired.
Add a cooldown guard to should_compress(): return False while
_summary_failure_cooldown_until is in the future. Reuses the existing float;
no new state. Manual /compress (force=True) still clears the cooldown first.
Fixes#11529
Generic provider:custom relays were force-routed to the OpenAI Responses
API whenever the model matched gpt-5*, and a stale persisted
model.api_mode=codex_responses survived /reset and upgrades. Some
OpenAI-compatible relays do not implement Responses semantics, which
surfaced as malformed function_call.name replay errors in gateway sessions.
- runtime_provider: route custom-provider api_mode through
_resolve_plain_custom_api_mode(), which drops a stale codex_responses
unless the URL is direct OpenAI/xAI
- run_agent: _provider_model_requires_responses_api returns False for
custom; direct api.openai.com / api.x.ai URLs still upgrade via
_is_direct_openai_url() / URL detection
- regression coverage for plain relays vs direct OpenAI/xAI URLs
Co-authored-by: HiddenPuppy <HiddenPuppy@users.noreply.github.com>
* fix(agent): drop tool_calls with empty function.name to prevent orphan 400
Salvage of #12807 by @melonboy312 — rebased onto current main (sanitizer
moved to agent_runtime_helpers), scoped to the sanitizer fix, with a
regression test that fails without it.
* fix(agent): repair (not drop) empty-name tool_calls to preserve anti-priming + prevent 400
Dropping empty-name tool_calls in the pre-call sanitizer collided with #47967,
which intentionally keeps an empty-name call paired with a synthesized
'tool name was empty' anti-priming result so weak models self-correct without
a full catalog dump. Dropping the call orphaned that result and stripped the
signal (breaking tests/agent/test_empty_tool_name_loop_dampening.py).
The actual HTTP 400 cause is an ORPHANED function_call_output (adapter drops
the empty-name function_call but keeps its output). Rename the blank name to a
non-empty sentinel instead: the call and its result stay paired, the adapter
no longer drops the function_call, no orphan, no 400 — and the anti-priming
result content the model needs is preserved.
---------
Co-authored-by: Bartok9 <danielrpike9@gmail.com>
background_review hardcoded enabled_toolsets=["memory", "skills"] in the
review fork's whitelist, so a skill-review fork on a profile with
memory_enabled: false still granted the LLM the built-in MEMORY.md read/write
tool — contaminating a profile that opted out of built-in memory. The flag was
already in scope (review_agent._memory_enabled). Include "memory" only when
_memory_enabled or _user_profile_enabled (USER.md also needs the tool).
Layer 1 of #54937 (the path leak) is fixed by this PR's thread-context
propagation: get_memory_dir() is already per-call on main, so once the
bg-review thread inherits the profile override its writes land in the right
profile (verified). This commit closes the remaining whitelist layer.
MoA sessions could not stream: the gateway streaming toggle was a no-op for
provider "moa", so users saw nothing until the entire response finished — minutes
of silence on long turns. The aggregator's reply was always fetched whole.
Root cause was twofold:
1. conversation_loop hard-disabled streaming for provider in {"copilot-acp",
"moa"} (MoA grouped with the ACP client, whose facade isn't a stream).
2. MoAChatCompletions.create() fetched the aggregator response whole via
call_llm(), which had no streaming mode.
For provider "moa", _create_request_openai_client() returns the MoAClient facade
itself, so the existing streaming consumer already calls
MoAChatCompletions.create(stream=True). We reuse that battle-tested consumer
(text-delta delivery, tool_call reassembly, stale-stream detection, non-streaming
fallback) instead of adding a parallel streaming path.
Changes:
- call_llm() gains stream/stream_options. When streaming it returns the raw SDK
stream iterator directly, bypassing _validate_llm_response and the
temperature/max_tokens/payment fallback chain (which assume a complete
response). The caller owns reassembly and fallback.
- MoAChatCompletions.create() runs the references first (unchanged), then when
stream=True returns the aggregator's raw stream, forwarding stream_options and
the consumer's per-request read timeout. stream=False is byte-identical to
before (no stream/stream_options/timeout forwarded).
- conversation_loop streams MoA only when a display/TTS consumer is present;
quiet/subagent/health-check paths keep the complete-response path.
Tests: tests/run_agent/test_moa_streaming.py — create() stream/non-stream
branches, stream_options + timeout forwarding, call_llm raw-stream return vs
validated non-stream. Existing MoA tests unchanged (20 passed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
_resolve_task_provider_model() flattened any explicit base_url to
provider=custom. Correct for bare/custom endpoints, but wrong for
provider-backed routes (anthropic, qwen-oauth, minimax-oauth,
openai-codex, etc.) whose provider branch adds auth refresh, transport,
or request shaping. MoA reference slots resolved through those providers
lost their identity before the aux call, so e.g. a Codex reference hit
chatgpt.com/backend-api/codex without its Cloudflare headers and got
HTML back (surfacing as a spurious rate-limit).
Keep first-class providers intact when paired with a resolved base_url
via _preserve_provider_with_base_url(); bare/custom/auto/unknown and the
direct openai alias still route through custom.
Co-authored-by: Hermes Agent <127238744+teknium1@users.noreply.github.com>
* fix(agent): merge consecutive assistant messages in repair_message_sequence
Strict OpenAI-compatible providers (DeepSeek v4, Moonshot/Kimi) reject a
replayed history where an assistant message carrying tool_calls is
immediately followed by another assistant message instead of its tool
results — HTTP 400 'An assistant message with tool_calls must be
followed by tool messages...'.
repair_message_sequence (the defensive belt run before every API call)
fixed orphan-tool and consecutive-user shapes but never merged
consecutive assistant messages. Adds a Pass 0 that collapses adjacent
assistant turns into one — union of tool_calls, concatenated content,
carried reasoning_content — covering both reported shapes:
- parallel tool calls split across two assistant turns (#29148)
- content-only assistant followed by tool_calls-only assistant (#49147)
A tool result or user turn between two assistants blocks the merge
(distinct, valid rounds). Runs before Pass 1 so the merged union of
tool_call ids is known to the orphan-tool filter.
Closes#29148, #49147.
Co-authored-by: Bartok9 <danielrpike9@gmail.com>
Co-authored-by: woaini30050 <woaini30050@users.noreply.github.com>
Co-authored-by: weidzhou <weidzhou@users.noreply.github.com>
* fix(agent): exempt codex Responses interim turns from assistant merge
The Pass 0 consecutive-assistant merge collapsed codex_responses interim
turns, which legitimately stay separate — each carries its own encrypted
continuation state (codex_reasoning_items / codex_message_items) that
must replay verbatim. Skip the merge when either side is a codex interim
(has codex_reasoning_items / codex_message_items / finish_reason=='incomplete').
Fixes the slice-2 regression in test_run_agent_codex_responses.py
(test_duplicate_detection_distinguishes_different_codex_{reasoning,message_items}).
---------
Co-authored-by: Bartok9 <danielrpike9@gmail.com>
Co-authored-by: woaini30050 <woaini30050@users.noreply.github.com>
Co-authored-by: weidzhou <weidzhou@users.noreply.github.com>
The earlier enterprise base URL change (proxy-ep parsing) gave us URLs
like `api.enterprise.githubcopilot.com`, but ~15 host-matching call
sites still hard-coded `api.githubcopilot.com`. Enterprise users would
therefore drop the `Copilot-Integration-Id: vscode-chat` header at
client-build time, and upstream rejected requests with:
The requested model is not available for integrator "zed"
(or "copilot-language-server") — verify the correct
Copilot-Integration-Id header is being sent.
The header was correct in copilot_default_headers(); it just never
made it into default_headers for non-default hostnames because every
detector compared against the exact string "api.githubcopilot.com".
This commit broadens all those checks to "githubcopilot.com" via
base_url_host_matches (which already does proper subdomain matching),
so api.enterprise.githubcopilot.com, api.business.githubcopilot.com,
etc. all share the same headers, vision routing, max_completion_tokens
selection, and reasoning-effort detection as the default endpoint.
Also adds ".githubcopilot.com" to _URL_TO_PROVIDER so context-window
resolution via models.dev works for enterprise base URLs, and tightens
_is_github_copilot_url to use suffix matching instead of strict equality.
Tests:
- New: enterprise Copilot endpoint preserves Copilot-Integration-Id
- New: enterprise endpoint returns max_completion_tokens (not max_tokens)
- Existing 333 base_url / copilot / aux-client / credential-pool tests pass
Parts 5 of #7731.
_sanitize_api_messages() compared raw tool_call_id strings without
stripping whitespace. When assistant-side IDs and tool-result IDs
diverged due to surrounding whitespace, valid tool results were treated
as orphaned and replaced with [Result unavailable] stub placeholders.
Strip whitespace in _get_tool_call_id_static() (both call_id/id paths,
dict and object) and at the two result_call_id comparison sites in
sanitize_api_messages(). Adds regression tests for preserved-whitespace
results and orphaned-whitespace removal.
Closes#9999
Defense-in-depth on top of _safe_session_filename_component (#5958):
Sink (makes the bad write impossible regardless of entry point):
- run_agent._save_session_log: sanitize session_id before building the
session_{sid}.json snapshot path.
- agent_runtime_helpers.dump_api_request_debug: sanitize before building
the request_dump_{sid}_{ts}.json path.
Boundary (clean 400 instead of a silently-hashed filename):
- api_server rejects path-traversal-shaped X-Hermes-Session-Id on the
session-continuation path and the explicit /api/sessions create path,
reusing gateway.session._is_path_unsafe (mirrors the native gateway's
entry-boundary guard). Also enforces the session-header length cap on
the continuation path.
Tests: traversal session_id stays contained at the write site; sanitizer
always yields a traversal-free segment; the API header rejects
../, absolute, and Windows-traversal IDs with 400.
Google's native Gemini REST endpoint (generativelanguage.googleapis.com,
non-/openai) rejects OpenAI-only stream_options={"include_usage": true},
crashing every streaming chat-completions call with TypeError. Omit it for
that endpoint while keeping it for the Gemini OpenAI-compat shim and all
OpenAI-compatible aggregators (OpenRouter, etc.) so usage accounting is
preserved.
Reuses is_native_gemini_base_url() so the compat shim (.../openai), which
accepts stream_options, is correctly excluded from the omission.
Fixes#14387
Co-authored-by: Hermes Agent <127238744+teknium1@users.noreply.github.com>
When an LLM API call returns HTTP 4xx with an empty parsed SDK `body` ({}),
`_summarize_api_error` fell through to a bare `str(error)`, so users saw only
"HTTP 400" with no provider detail (reported on Windows in #36109). The SDK
leaves `body` empty in this case, but the httpx `response` still carries the
payload in `.text`.
- run_agent.py `_summarize_api_error`: when `body` is empty, fall back to
`response.text` — parse a JSON `error.message`/`message` when present, else
surface the raw (truncated) body. Platform-agnostic diagnostics.
- hermes_cli/oneshot.py: `hermes -z` now runs via `run_conversation` and returns
exit code 2 when the run is failed/partial with no usable final response, so
scripts can detect LLM failures (still 0 when a response — incl. an error
summary as output — is produced).
Tests: new tests/run_agent/test_summarize_api_error.py (empty-body JSON + raw
text, RED/GREEN verified) + oneshot exit-code/`run_conversation` wiring tests.
NOTE: #36109's original root cause (Windows "all providers return empty 400")
is not reproducible on current main (heavy provider-transport churn since
v0.15.1). This change does not claim to fix that root cause — it makes any
empty-body API error LEGIBLE so a future occurrence shows the real provider
message instead of a bare HTTP 400. Relates to #36109 (does not close it).
When a provider's output-layer safety filter (MiniMax "output new_sensitive
(1027)", Azure content_filter, etc.) kills a streaming response after deltas
were already sent, interruptible_streaming_api_call swallows the raw error
into a finish_reason=length partial-stream stub. The conversation loop then
burned 3 continuation retries against the SAME primary — re-hitting the
content-deterministic filter every time — and gave up with "Response remained
truncated after 3 continuation attempts", never consulting fallback_providers.
Builds on @595650661's classifier change (cherry-picked) so error_classifier
recognizes the filter; then:
- chat_completion_helpers: run the swallowed error through error_classifier at
the stub-creation point and stamp _content_filter_terminated on the stub
(single source of truth — no parallel pattern list).
- conversation_loop: read the tag and activate the fallback chain BEFORE
burning any continuation retries; roll partial content back to the last
clean turn and re-issue against the new provider (restart_with_rebuilt_messages).
Plain network stalls are unaffected (only content_policy_blocked is tagged).
Credits #32479 (@sweetcornna) and #33845 (@Tranquil-Flow) which fixed the
same issue via the stub-tag and loop-escalation approaches respectively.
Live E2E confirmed: before, _try_activate_fallback called 0x; after, fallback
fires on the first stub and the fallback provider completes the turn.
The MiniMax output-layer safety filter surfaces the error verbatim as
`output new_sensitive (1027)` (sometimes with additional provider
wrapping like 'Stream stalled mid tool-call: output new_sensitive (1027)').
When the model emits a large tool-call argument block, the upstream
filter trips and the SSE stream is truncated mid-flight, producing
'stream stalled mid tool-call' errors. Until now this case was
misclassified and retried 3x on the same provider, reproducing the same
refusal and burning paid attempts.
Adding `new_sensitive` to `_CONTENT_POLICY_BLOCKED_PATTERNS` routes
it through the existing is_client_error path: skip 3x retry, activate
configured fallback model immediately, surface a clear provider-safety
message to the user.
Refs #32421
The advisory reference view stripped all tool calls and tool results, so
reference models judged a task whose actions and results they never saw — and
references only fired once per user turn, never re-running as the agent's
state advanced through the tool loop.
Two fixes:
- _reference_messages() now PRESERVES the agent's tool calls and tool results,
rendering them inline as text ([called tool: ...] / [tool result: ...]) so a
reference gives an informed judgement on the real current state. Still emits
zero tool-role messages and zero tool_calls arrays (strict providers reject
those), and large tool results are previewed head+tail (4000-char budget).
The required end-on-user shape is met by APPENDING a synthetic advisory user
turn — not by deleting the agent's latest context (which the prior fix did).
- References now re-run on every state change — each new user message AND each
new tool result — instead of once per user turn. The state-sensitive advisory
signature drives the cache: new tool result = miss (re-run), identical-state
re-call = hit (no re-run, no re-emit).
The acting aggregator still receives the full, untrimmed transcript.
* fix(moa): reference advisory view must end with a user turn
MoA reference calls failed with Anthropic models that don't support
assistant prefill (e.g. Claude Opus 4.8): '400 ... must end with a user
message'. The advisory view built by _reference_messages() kept the last
assistant turn's text while dropping the following tool result, leaving a
trailing assistant turn — which Anthropic (and OpenRouter->Anthropic)
interpret as an assistant prefill to continue. References are advisory and
must end on the user turn they answer.
Strip trailing assistant turns from the advisory view (preserving
intervening ones). Update the existing test that encoded the buggy shape
and add a mid-tool-loop regression test.
* feat(moa): give reference models an advisory-role system prompt
Reference models received the bare trimmed conversation with no role
framing, so they assumed they were the acting agent and refused ("I can't
access repositories/URLs from here") or tried to call tools they don't have.
Prepend a dedicated advisory system prompt to every reference call: the
model is an analyst, not the actor — it cannot execute, should not
apologize for lacking tools, and should reason about the presented state to
advise the aggregator/orchestrator on approach, next steps, tool-use
strategy, risks, and anything the acting agent missed. Its output is private
guidance for the aggregator, not a user-facing answer.
The exhaustion-cooldown timing assertions relied on a wall-clock budget
(before + window + 1.0s). On loaded CI runners the activation calls could
exceed the 1s slack, flaking 'Run tests slice 4/8'. Freeze
chat_completion_helpers.time.monotonic so the cooldown math is exact and
load-independent across all four tests.
Three CI flakes hit while landing the credential-pool restore fix; all three
were timing/wall-clock races in the tests, not product bugs (each passes
locally and the assertions are correct):
- test_entire_tree_is_sigkilled_not_just_parent: _terminate_host_pid SIGKILLs
synchronously, but the test's 4s budget after a 1s in-function SIGTERM grace
left almost no slack for the kernel to tear down 3 processes + reparent the
children to zombies under loaded-CI scheduling. Widen the wait to 15s and
make the liveness predicate tolerant of vanished-pid / zombie races. The
assertion never weakens: every tree member must end up dead or zombie.
- test_session_resume_follows_compression_tip: appended messages got
time.time() timestamps (~now) while the test forced session started_at into
the past, so the get_compression_tip MAX(m.timestamp) tiebreaker depended on
wall-clock ordering. Pass explicit, well-separated message timestamps so the
chain resolution is deterministic by construction.
- test_non_retryable_exhaustion_arms_cooldown: asserted the short (5s)
exhaustion cooldown with a tight +1.0s slack, which false-fails when
wall-clock jitter between the 'before' snapshot and the cooldown computation
exceeds a second on a loaded runner. Widen to +30s — still cleanly below the
60s rate-limit window it must distinguish from.
The existing test_anthropic_stream_parser_valueerror_retries_before_delivery
asserted mock_replace.call_count == 1 — i.e. it passed precisely because the
buggy OpenAI rebuild was invoked on the Anthropic path. Repoint it to assert
the corrected close+rebuild-Anthropic behavior (#28161).
interruptible_streaming_api_call() has three connection-pool cleanup
sites that called _replace_primary_openai_client() unconditionally.
For api_mode=anthropic_messages this has two consequences:
1. _replace_primary_openai_client() fails (OPENAI_API_KEY unset on
Anthropic-only configs), so dead connections are never purged.
2. The stale-stream detector's outer-poll site (L1977) is the only
mechanism that can interrupt the worker thread while it blocks in
for event in stream:. Because the Anthropic client is never closed,
the thread stays blocked until the 900 s httpx read-timeout fires,
producing a visible 15-minute hang for Telegram/gateway users on
claude-opus-4-7.
Fix: mirror the existing interrupt-path pattern (L1989-1997) at all
three cleanup sites — if api_mode == "anthropic_messages", call
_anthropic_client.close() + _rebuild_anthropic_client() instead of
_replace_primary_openai_client(). _rebuild_anthropic_client() handles
both direct Anthropic and Bedrock-hosted Claude correctly, unlike the
inline build_anthropic_client() calls in open PR #14430.
PR #14430 (open) covers only the outer stale-detector site (L1977).
PR #23678 (open) covers only the inner retry sites (L1774, L1833).
This PR covers all three sites and uses _rebuild_anthropic_client()
for Bedrock parity.
Fixes#28161
Same bug class as the Anthropic fix (#26293): the OpenAI/aggregator client is
built without max_retries, so the SDK default of 2 applies. The SDK's own 1-2s
backoff ignores Retry-After and retries inside hermes's outer conversation loop,
burning request slots against a rate-limited bucket. Set max_retries=0 at the
single create_openai_client chokepoint (covers init, switch_model, recovery,
restore, request-scoped). auxiliary_client builds its own clients and is not
wrapped by the loop, so it keeps SDK retries.
The Anthropic SDK clients were built without max_retries, so the SDK
default (max_retries=2) retried 429/5xx with its own backoff that ignores
Retry-After — double-retrying inside hermes's outer loop and burning
request slots against a bucket that won't refill for minutes. Set
max_retries=0 on all Anthropic/AnthropicBedrock client constructions so
the outer conversation loop (which already honors Retry-After) owns retry.
Also raise the Retry-After cap in the conversation loop from 120s to 600s.
Anthropic Tier 1 input-token buckets reset in ~171s, so the 120s cap made
hermes retry before the reset window and re-trip the limit.
Refs #26293
A persistent upstream 401 on a single-entry OAuth pool (common for Claude
Max subscribers) made the credential-pool recovery spin forever:
try_refresh_current() re-mints a fresh token and reports success on every
401, so recover_with_credential_pool returned True and the retry loop
continue'd without ever incrementing retry_count or reaching the
auth-failover block. The configured fallback_model never activated and the
agent appeared to hang.
Cap consecutive successful same-entry refreshes (keyed by provider +
pool-entry id) at 2; once exceeded, treat the credential as unrecoverable
and return not-recovered so the loop falls through to
_try_activate_fallback. The 429/billing paths already rotate-or-fall-through
correctly (mark_exhausted_and_rotate returns None on a single entry), so
only the auth-refresh branch needed the cap.
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
When every provider in the fallback chain fails non-retryably back-to-back
(e.g. HTTP 400/402/429 across distinct providers), the within-turn walk is
already bounded — _fallback_index advances monotonically and the loop aborts
when the chain exhausts. The damaging mode is cross-turn: restore_primary_
runtime resets _fallback_index=0 every turn, so a client that re-submits
immediately replays the entire chain, re-marshaling the full (potentially
80k-token) context once per provider every turn with no throttle on the
non-rate-limit path. On constrained hosts this exhausts memory/swap.
Rate-limit/billing failures already arm a 60s cooldown via _rate_limited_until;
the gap was the non-rate-limit case. Now, when the chain exhausts on a non-
rate-limit failure with a non-empty chain, arm a short (5s) cooldown on the
same _rate_limited_until gate (max(), never shrinking an existing window).
The next turn's restore stays gated and does NOT reset the index, so the
chain isn't replayed until the cooldown clears. No new state, no thread sleep,
no false-trip on legitimately long chains (those walk normally within a turn).
Tests: tests/run_agent/test_24996_fallback_exhaustion_cooldown.py
The agent's image-rejection fallback strips images and retries text-only when
a provider rejects image content, which is what lets the gateway drain its
queued messages. The fallback only fires on a hardcoded phrase list, and the
OpenRouter wording — HTTP 404 'No endpoints found that support image input' —
was missing. For OpenRouter-routed non-vision models the fallback never fired,
the retry loop re-sent the same rejected request until exhaustion, and every
subsequent message (including plain text) stayed queued behind the stuck turn.
Add the phrase to _IMAGE_REJECTION_PHRASES (the 404 already passes the 4xx
gate). Add a positive test and a guard test so the sibling OpenRouter
'no endpoints ... data policy / guardrail' 404s do NOT get their images
stripped.
Fixes#21160. Reported by @liu14goal14-ux; PR #21198 by @ygd58.
When a MoA preset is selected, each reference model's answer now renders in the
CLI as a thinking-style block labelled with its source model, BEFORE the
aggregator responds — so the mixture-of-agents process is visible instead of a
silent pause. The aggregator's response (and its tool actions) follow as normal.
Mechanism (shared seam, all surfaces):
- MoAChatCompletions/MoAClient take an optional reference_callback and emit
'moa.reference' (index/count/label/text) per reference, then 'moa.aggregating'
(aggregator label) once. agent_init wires this to the agent's
tool_progress_callback, which every surface already consumes — so the events
reach CLI/TUI/desktop/gateway with no new plumbing.
- CLI _on_tool_progress renders 'moa.reference' as a labelled '┊ ◇ Reference
i/n — <model>' header + a thinking-style preview (reusing _emit_reasoning_
preview), and 'moa.aggregating' as a spinner transition. Display-only; never
touches message history (cache-safe).
Turn-scoped reference cache: the agent loop calls the facade once per tool-loop
iteration, but the advisory message view is identical across iterations within a
turn, so references are now run AND displayed once per user turn (keyed by the
advisory view's signature) instead of re-running/re-spamming on every iteration.
This also cuts reference API cost from O(iterations) back to O(turns).
Verified live via interactive PTY on the opus-gpt preset (gpt-5.5 + opus refs):
reference blocks render once per turn, labelled by model, before the aggregator;
fresh blocks on each new turn; aggregator tool actions still execute.
Follow-up: TUI/desktop rich rendering + gateway batched-summary already receive
the events via tool_progress_callback; their surface-specific renderers are a
separate change.
MoA was calling reference and aggregator models through a bare
call_llm(provider=slot["provider"], model=slot["model"]) with a forced
temperature and a forced max_tokens (the preset's hardcoded 4096). That left
base_url/api_key/api_mode unresolved — so the auxiliary auto-detector guessed
the API surface instead of using the provider's real runtime, and the 4096 cap
truncated long aggregator syntheses.
A MoA slot is just a model selection and must be called the same way any model
is called elsewhere. Each slot is now resolved through resolve_runtime_provider
(the canonical provider→api_mode/base_url/api_key resolver the CLI, gateway, and
delegate_task all use) via a new _slot_runtime() helper, and the resolved
endpoint is passed into call_llm. So a reference/aggregator gets its provider's
actual API surface — MiniMax → anthropic_messages, GPT-5/o-series →
max_completion_tokens, custom endpoints → their base_url — identical to how that
model is handled as the acting model.
MoA also no longer imposes its own output cap: max_tokens defaults to None
(omitted → the model's real maximum) for references and is passed through from
the caller for the aggregator. The preset's hardcoded 4096 is gone. The
max_tokens preset config field is left in place (config/web/desktop unchanged);
it is simply no longer applied as a forced cap.
Tests: slots route through resolve_runtime_provider with resolved base_url/
api_key; resolution errors fall back to bare provider/model; neither call
carries an output cap even when the preset config still contains max_tokens.
When automatic fallback activates a provider that differs from the
primary, try_activate_fallback() cleared the primary's pool (to avoid
cross-provider base_url contamination, #33163) but never loaded the
fallback provider's own pool. The fallback then ran with no pool, so
rate_limit/billing/auth recovery couldn't rotate its credentials.
After clearing a mismatched pool, load_pool(fb_provider) and attach it
when it has credentials, so provider-specific rotation continues to
work on the fallback target.
switch_model() swapped model/provider/base_url/api_key but never
refreshed agent._credential_pool, which stays bound to the original
provider. recover_with_credential_pool() then sees a pool.provider !=
agent.provider mismatch and short-circuits — so a 429/401 on the new
provider gets no rotation and falls through to fallback instead.
Reload load_pool(new_provider) inside switch_model when the provider
changes (or the pool is missing). The reload is inside the protected
swap block and the pool is added to the rollback snapshot, so a failed
client rebuild restores the original pool.
Fixes#16678, #52727.
Follow-up to the #13971 fix: a genuine native Ollama provider reached
through a reverse proxy carries no ollama/:11434 URL signature, so the
restricted detection would miss it. Add provider=="ollama" as an
explicit True case (idea from #14789, @Tranquil-Flow) and cover both it
and the #13971 LiteLLM-proxy-to-zai false-positive with E2E tests.
The _is_ollama_glm_backend() function was too broad: any local endpoint
running a GLM model was treated as Ollama, triggering the stop->length
misreport heuristic introduced in 8011aa3. This caused false truncation
detection on sglang, vLLM, LM Studio, and other non-Ollama servers that
correctly report finish_reason.
When a GLM model on sglang/vLLM returned finish_reason='stop', the agent
mistakenly reclassified it as 'length' if the response didn't end with
a whitelisted punctuation character (ASCII or CJK). This particularly
affected Chinese-language responses and Markdown-formatted text.
Root cause: the is_local_endpoint() fallback assumed any local GLM
endpoint = Ollama. But many non-Ollama servers also run on localhost.
Fix: remove the is_local_endpoint() catch-all. Only detect Ollama via
its distinctive signatures (port 11434, 'ollama' in URL). All other
local servers are assumed to report finish_reason correctly.
This is the correct tradeoff because:
- False negatives (Ollama at custom port, heuristic not triggered) only
mean the user sees a truncated response — same as having no heuristic
- False positives (non-Ollama server, heuristic wrongly triggered) inject
spurious continuation messages into the conversation — strictly worse
Adds two tests:
- sglang GLM response is NOT reclassified as truncated
- Ollama GLM on port 11434 still triggers the heuristic as before
Co-authored-by: Hermes Agent <hermes@nousresearch.com>
Force redact_sensitive_text(force=True) on the browser_type text arg so
recognized credentials (API keys, tokens, JWTs) are masked in tool
progress, previews, callbacks, and return payloads even when the global
security.redact_secrets opt-out is set — a typed credential reaching chat
history is a security boundary, not log hygiene. Normal typed text matches
no pattern and stays fully readable for debuggability.
Tests assert the API-key-shaped secret is masked across every surface and
that normal text passes through unchanged.
Stopping a turn while the model is streaming (stop/esc to redirect) raised
InterruptedError, set final_response to the throwaway "waiting for model
response" sentinel, and persisted messages WITHOUT the assistant text that
was already streamed to the screen. The next turn then had no record of the
half-finished reply, so the model appeared to "forget" what it just said.
Recover the on-screen text from _current_streamed_assistant_text in the
InterruptedError branch and append it as the assistant turn (and surface it
as final_response). The metadata sentinel is kept only when nothing was
streamed yet, preserving the ACP/client suppression behavior.
Completes the partial-stream recovery from 397eae5d9 (which wired the same
_current_streamed_assistant_text salvage into the connection-failure twin
but missed the user-interrupt path). The lossy handler dates to c98ee9852.
* feat(moa): expose MoA presets as selectable virtual models
Reconstructed onto current main (PR #46081's base had diverged with no common
ancestor, marking the PR dirty so CI never dispatched). MoA is now a virtual
provider: each named preset is a selectable model under provider 'moa', and the
preset's aggregator is the acting model that answers and calls tools.
Reference models fan out in parallel via a bounded ThreadPoolExecutor (the same
batch pattern delegate_task uses) — all references dispatched at once, collected
when every one finishes, then handed to the aggregator. Output order is
preserved, failures and the MoA-recursion guard stay isolated per reference.
- Removed the old mixture_of_agents model tool and moa toolset.
- Added moa as a virtual provider in the provider/model inventory.
- /moa is shortcut behavior over model selection (default preset / named preset
/ one-shot prompt).
- Dashboard + Desktop manage named presets; presets appear in model pickers.
- Parallel reference fan-out in agent/moa_loop.py with regression test.
* fix(moa): thread moa_config through _run_agent to _run_agent_inner
The reconstructed gateway MoA wiring declared moa_config on _run_agent (the
profile-scoping wrapper) and used it inside _run_agent_inner, but the wrapper
never forwarded it — _run_agent_inner had no such parameter, so the runtime hit
NameError: name 'moa_config' is not defined on the compression-failure session
sync path. Add moa_config to _run_agent_inner's signature and forward it from
both wrapper call sites (multiplex and non-multiplex). Caught by
tests/gateway/test_compression_failure_session_sync.py on CI shard test(4).
* fix(moa): classify moa as a virtual provider in the catalog
The moa virtual provider has no PROVIDER_REGISTRY/ProviderProfile entry, so
provider_catalog() fell through to the default auth_type="api_key" with no
env vars — tripping two catalog invariants:
- test_provider_catalog: api_key providers must expose a credential env var
- test_provider_parity: every hermes-model provider must be desktop-configurable
moa already declares auth_type="virtual" in HERMES_OVERLAYS; consult that
overlay as an auth_type fallback so the catalog reports moa as virtual (no real
credential, no network endpoint). Exempt virtual providers from the desktop
parity union check the same way 'custom' is exempt — derived from the catalog,
not a hardcoded slug, so future virtual providers are covered too.
When an explicit aux provider cannot build a client before any request is
sent (missing raw env key, exhausted/unavailable OAuth or credential-pool
auth, resolver returning (None, None)), call_llm raised a misleading
"no API key was found" error and bypassed the configured fallback_chain
entirely. A provider authenticated through Hermes auth / the credential
pool (e.g. ollama-cloud) whose pool entry is exhausted hit this path, so
compression failed instead of routing to the configured fallback.
Adds _try_configured_fallback_for_unavailable_client() and wires it into
both sync and async call_llm before the raise, and into the startup
compression feasibility check.
Salvaged from #51835 by @herbalizer404.
These 7 test sites assert rotation behavior (fork, child sessions, lock
contention, logging session-context follows id rotation, boundary hooks fire
on rotation). Pin each builder to in_place=False explicitly so they keep
exercising the retained rotation fallback regardless of the global default
(flipped to True in #38763). Rotation stays a working opt-out fallback and
deserves continued coverage — these are NOT deleted.
Pinned sites:
- test_compression_concurrent_fork._build_agent_with_db
- test_compression_logging_session_context._build_agent_with_db
- test_compression_rotation_state._build_agent_with_db
- test_compression_boundary_hook._make_agent (2 helpers: CompressionBoundaryHook + SessionCompressEvent)
- test_compression_concurrent_sessions._build_agent_with_db
Make verification closure the default coding behavior after landed file edits while keeping bounded retries and config/env switches for users who need to disable it.
Add compression.minimum_context_floor config key that allows users
to lower the compression threshold floor below the hardcoded 64K
default, preventing infinite tool-call loops on models whose
structured output degrades well before 64K tokens.
- agent/model_metadata.py: add get_configurable_minimum_context()
helper with 16K hard safety limit
- agent/context_compressor.py: accept minimum_context_floor param,
thread it through _compute_threshold_tokens
- agent/conversation_compression.py: use compressor's floor for
aux model context validation
- agent/agent_init.py: read compression.minimum_context_floor from
config and pass to ContextCompressor
- gateway/run.py: cache-busting includes new key
Salvaged from #31686 by @Tranquil-Flow onto current main.
Resolves conflicts with in-place compaction (#38763) and max_tokens
threshold computation (#43547) that landed after the original PR.
Closes#31600
Adds auxiliary.background_review.{provider,model} (default auto = main chat
model — unchanged). Set it to a different, cheaper model and the post-turn
self-improvement review runs there for ~3-5x lower cost.
Cache-aware by design: the main chat is warm in the prompt cache, so the
default full-history replay on the main model is cheap cache reads — left
exactly as-is. A different model can't reuse that cache (different key), so
when (and only when) routed to a different model the fork replays a compact
digest instead of the full transcript, minimising what it cold-writes on the
aux model. Same model -> full replay; different model -> digest.
Quality holds in benchmarks: memory capture identical, skill near-identical.
Nothing changes unless you opt in by naming a different model.
Co-authored-by: Hermes Agent <noreply@nousresearch.com>
The success/staged gating and op-expansion for mirroring built-in memory
writes to external providers lived in a standalone agent/memory_write_bridge.py
helper called inline from two core call sites (tool_executor.py,
agent_runtime_helpers.py). That left the mirror decision-making in the agent
loop, outside the memory-provider interface.
Fold it into a new MemoryManager.notify_memory_tool_write() entry point: the
loop now hands over the raw tool result + args and a metadata callback, and the
manager decides whether/what to mirror. Both core call sites collapse to a
single call; the orphan module is removed. No MemoryProvider ABC change.
Tests rewritten as behavior tests against the manager method.
Mirror built-in memory writes to external providers only after the native memory tool succeeds and is not staged for approval. Keep OpenViking's built-in memory mirroring add-only, since Hermes native memory entries do not yet have stable OpenViking file URIs for replace/remove.
Add a narrow viking_forget tool for exact user memory file deletion and document the current OpenViking write/delete behavior.
Adds test_413_retries_on_token_only_compression: same message count but
materially fewer tokens after compaction must count as progress and retry,
not abort. Fails on main without the salvaged fix, passes with it.
* fix(agent): strip stale reasoning_content when falling back to a strict provider
A reasoning primary (DeepSeek/Kimi/MiMo thinking mode) pins reasoning_content
on every assistant tool-call turn (a single space " " pad). api_messages is
built once under the primary; on a mid-session fallback to a strict
OpenAI-compatible provider (Mistral, Cerebras, Groq, SambaNova), those stale
pads were replayed verbatim and rejected with HTTP 400/422:
body.messages.2.assistant.reasoning_content: Extra inputs are not
permitted (input: ' ')
reapply_reasoning_echo_for_provider() only ever ADDED pads, so it never
reconciled history built under a reasoning primary against a strict fallback.
copy_reasoning_content_for_api() also leaked empty-string and 'reasoning'-only
shapes to non-pad providers.
Fix both sites: when the active provider does not enforce echo-back, strip
reasoning_content (empty, space-pad, or non-empty) entirely. Re-padding when
switching TO a reasoning provider is preserved. Covers the Cerebras 400 from
#45655 and the DeepSeek->Mistral 422 fallback report.
Refs #45655.
* test: update reasoning-replay tests for strict-provider stripping
test_explicit_reasoning_content_beats_normalized_reasoning_on_replay was
implicitly running on the OpenRouter fixture (non-pad); pin it to a reasoning
provider so the precedence it checks is observable. Add a positive
strict-provider test asserting reasoning_content is stripped on replay.
The background-review fork (fires ~every 10 turns) pins
review_agent.session_id = agent.session_id — the parent's LIVE id — for
prefix-cache parity, then calls close(). With session finalization now in
close(), that would end the still-active parent session mid-conversation.
Set _end_session_on_close = False on the fork so the real owner (CLI close /
gateway reset / cron) finalizes the session instead.
Follow-up to the #12029 fix.
When the main provider is the Codex app-server runtime (api_mode
codex_app_server), the gateway showed no verbose 'running X' tool-progress
breadcrumbs on Telegram while every other provider did. The app-server
session processes item/started notifications (command execution, file
changes, MCP/dynamic tool calls) but never surfaced them as Hermes
tool-progress events — the session was constructed without an on_event
hook, so the agent's tool_progress_callback was never invoked on this
route.
Add _codex_note_to_tool_progress() mapping item/started → (tool_name,
preview, args) for commandExecution / fileChange / mcpToolCall /
dynamicToolCall, and wire an on_event hook into CodexAppServerSession that
forwards mapped events to agent.tool_progress_callback('tool.started',
...) — the same signature the chat_completions path uses (tool_executor.py).
Non-tool items (agentMessage/reasoning) and non-item/started methods map
to None and are ignored.
Co-authored-by: jplew <462836+jplew@users.noreply.github.com>
When the active provider returns a 401/403 that survives its per-provider
credential-refresh attempt (revoked OAuth, blocked/expired key, or an
account pinned to a dead/staging inference endpoint), the conversation
loop now escalates to the configured fallback chain instead of dead-ending.
Before: the generic failover dispatch fired only for {rate_limit, billing};
auth/auth_permanent fell through to 'switch providers manually' advice and
never called _try_activate_fallback(). A user whose primary credential was
broken kept thrashing on the same dead credential every turn — the main
agent appeared 'stuck in fallback mode' while never actually failing over.
This also affected auxiliary tasks (compression, vision, title-gen), since
auto-resolved aux follows the main provider.
After: a persistent auth failure with a configured fallback chain switches
to the next provider (mirroring the rate-limit/billing failover path),
guarded one-shot per attempt by TurnRetryState.auth_failover_attempted.
When no fallback is configured the behavior is unchanged — it falls through
to the existing terminal handling and provider-specific troubleshooting
guidance.
Tests: test_auth_provider_failover.py — 401/403 classify as auth, the
gating condition fires only with a chain present + guard unset, the guard
blocks repeats, and non-auth (500) errors do not trigger auth failover.
Follow-up to the soft-archive durability fix. Reusing the rewind/undo active=0
flag for compaction-archived turns inherited the wrong search semantics: undo
rows are intentionally HIDDEN from session_search (the user took them back), but
compaction-archived turns must stay DISCOVERABLE — that is the whole point of
Teknium's "searchable / recoverable" requirement. As built, search_messages
defaulted to WHERE active=1, so after in-place compaction the pre-compaction
turns were in the FTS index but filtered out of the default search. (The earlier
"searchable" claim only held for a raw FTS query / include_inactive=True, not
the actual session_search tool.)
Empirically confirmed the gap: search 'HMAC' returned 2 hits before compaction,
1 after (only the summary's mention) — the originals were hidden.
Fix — a `compacted` flag distinct from `active`, giving a 3-way state:
- active=1, compacted=0 → live context (normal)
- active=0, compacted=1 → compaction-archived: OUT of live context, IN search
- active=0, compacted=0 → rewind/undo: OUT of live context, OUT of search
Changes:
- messages.compacted INTEGER NOT NULL DEFAULT 0 added to SCHEMA_SQL. Declarative
_reconcile_columns adds it on existing DBs — no version bump (plain column add).
- archive_and_compact: UPDATE … SET active=0, compacted=1 (was active=0 only).
- search_messages: default WHERE active=1 → (active=1 OR compacted=1), on BOTH
the main FTS5 path and the trigram CJK path. include_inactive=True still
returns everything. The short-CJK LIKE fallback already returns all rows
(no active filter) — unchanged.
- Docstrings on archive_and_compact + search_messages document the 3-way state.
Verified: after compaction, session_search default finds the archived originals
(ids 1 & 4); rewind/undo rows stay hidden by default (recoverable via
include_inactive); live context still excludes both. 322 in-place + hermes_state
tests and 46 session_search tests green; ruff clean. Mutation check: reverting
the search WHERE to active-only fails the new searchable test.
(Surfaced by the question "is search semantic or only FTS?" — answer: session
search is FTS5 keyword/BM25 only, no embeddings over the transcript; semantic
retrieval lives in the optional memory-provider layer. Tracing that confirmed
the active-only filter gap above.)
Teknium review: keeping one durable session id must NOT come at the cost of
destroying history. The prior in-place implementation used replace_messages,
which hard-DELETEs the pre-compaction turns (they also drop out of the FTS
index) — same id, but the original conversation is gone with no recovery path
and the summary becomes the only record. Rotation today is non-destructive
(the old session's full transcript survives under the old id); in-place must
match that durability contract, not weaken it.
Fix: compact in place by SOFT-ARCHIVING, reusing the existing messages.active
flag (the /undo soft-delete mechanic), instead of deleting:
- New SessionDB.archive_and_compact(session_id, compacted): in one atomic
write, UPDATE messages SET active=0 on the live turns, then insert the
compacted set as fresh active=1 rows. Nothing is deleted.
- The insert loop is extracted into a shared _insert_message_rows() helper so
archive_and_compact and replace_messages don't duplicate the 60-line
column/encoding block (extend-don't-duplicate).
- Agent in-place branch calls archive_and_compact instead of replace_messages.
Durability outcome (proven by test + E2E across repeated compactions):
- Live context load (get_messages_as_conversation / get_messages) filters
active=1, so a resume reloads ONLY the compacted set — compaction still
shrinks the live session.
- The pre-compaction turns stay on disk at active=0, recoverable via
get_messages(include_inactive=True) / restore_rewound.
- They remain FTS-searchable: the messages_fts* triggers index on INSERT and
remove on DELETE only — they do NOT key on active, and active=0 is a
content-preserving UPDATE. session_search still finds them.
- Verified across TWO successive compactions: the 1st compaction's originals
are still recoverable + searchable after the 2nd (answers the "no recovery
path after the next compaction" concern directly).
message_count now reflects the LIVE (active/compacted) count, matching the
live load. replace_messages keeps its DELETE semantics (still correct for
/retry, /undo) and gains a docstring note pointing compaction at the
non-destructive method.
Tests: test_in_place_keeps_same_session_id strengthened to assert the 8
seeded originals survive at active=0 alongside the 2 compacted rows AND stay
FTS-searchable. Mutation check: swapping archive_and_compact back to a hard
DELETE fails the test, so the non-destructive contract is bound. 285
hermes_state + in-place tests green; rotation/persistence/compress-command/cli
suites green; ruff clean.
Parallel 3-reviewer cleanup of the in-place compaction code. Findings applied:
- perf: in-place mode no longer pre-flushes current-turn messages. The flush
ran INSERTs that the immediately-following replace_messages(compressed)
DELETE+reinsert discarded -- pure wasted writes per compaction. The
current-turn tail survives via the compressor's compressed output
(protect_last_n), not the flush. Verified no data loss; rotation still
pre-flushes (its old session row is preserved, so the flush is real there).
- quality: hoist the two shared post-write steps (update_system_prompt +
_last_flushed_db_idx = 0) below the if/else -- they ran in both branches
against agent.session_id. Removes the easiest divergence bug.
- quality: compute the compaction-boundary locals (_old_sid, _is_boundary,
_boundary_parent) ONCE instead of recomputing locals().get('old_session_id')
and the "_old_sid or agent.session_id or ''" chain three times.
- quality: initialize compacted_in_place up front and assign
agent._last_compaction_in_place directly, dropping the fragile
locals().get('compacted_in_place') reflection.
- reuse: parse the in_place config flag with utils.is_truthy_value (the
project's canonical truthy coerce) instead of a hand-rolled
str().lower() in {...} (agent_init already imports from utils).
Dropped as false positives / out of scope: gateway getattr of agent internals
(established session_id pattern), dual result-dict carry (mirrors history_offset
etc.), stringly-typed "compression" (codebase-wide convention, no constant).
Behavior-preserving: 7 in-place tests (incl. 2 new flush-guard tests) + 26
rotation/boundary/persistence/command tests green; mutation check confirms the
durable-replace guard still binds (removing replace_messages fails the test);
ruff clean. Added test_in_place_skips_redundant_preflush /
test_rotation_still_preflushes to guard the perf change.
Review (Codex + 3-agent parallel) found the first cut of in-place mode was
incomplete: it only updated the system prompt, so the persisted transcript
stayed 'full history + summary' and the next turn/resume reloaded the full
history and immediately re-compacted (a loop), and every downstream layer
that keyed off session-id rotation silently no-op'd. The session_id was
doing double duty as the 'compaction happened' signal. This wires the whole
path so removing rotation is actually complete:
Agent (agent/conversation_compression.py):
- In-place now DURABLY replaces the transcript: replace_messages(session_id,
compressed) on the same row (the canonical store the gateway reloads from),
not just update_system_prompt. Resume reloads the compacted set; no loop.
- Reset flush identity/cursor (_last_flushed_db_idx=0, _flushed_db_message_ids
cleared) so next-turn appends diff against the compacted transcript.
- Expose a rotation-independent signal: agent._last_compaction_in_place, and
in_place=True on the session:compress event.
- Fire the compaction-boundary hooks (context-engine on_session_start, memory
manager on_session_switch, reason='compression') in BOTH modes — in-place
passes the same id as parent so DAG/buffer state still checkpoints. Without
this, memory/context plugins miss every in-place compaction.
Gateway auto-compress (gateway/run.py):
- Read agent._last_compaction_in_place; set history_offset=0 on rotation OR
in-place (both return the compacted set, so slicing past the pre-compaction
length would drop everything). Carry compacted_in_place in the result dict.
- No extra rewrite needed: the agent shares the gateway's SessionDB, so its
replace_messages already updated the canonical store load_transcript reads.
Manual /compress (gateway/slash_commands.py):
- The throwaway /compress agent has no _session_db, so rewrite_transcript is
the durable write. Previously gated behind 'if rotated:' which treated
'id unchanged' as the #44794 data-loss failure case and SKIPPED the rewrite
— making /compress a silent no-op in in-place mode. Now rewrites on rotated
OR in_place; the data-loss guard still fires only for the genuine
no-rotation-AND-not-in-place failure.
Hygiene auto-compress already writes _compressed to the same id
unconditionally (its agent has no _session_db, can't rotate) — correct for
in-place, no change.
Tests (tests/run_agent/test_in_place_compaction.py):
- Assert the DURABLE transcript IS the compacted set after reload
(get_messages_as_conversation == compacted), message_count==2, flush
identity reset, and the rotation-independent signal set on in-place /
unset on rotation. Rotation regression guard unchanged.
Verified: 64 tests green across in-place + rotation/persistence/boundary/
concurrent/failure-sync/command/cli suites; E2E both modes (durable replace,
gateway offset=0, rotation preserves old transcript); ruff clean. Still
default-off.
Context compression today rewrites the message list AND rotates the
session id — it ends the session, forks a parent_session_id child, and
renumbers the title (name -> name #2). That moving identity key is the
root cause of a whole bug cluster: /goal lost (#33618), pending response
lost at the split (#14238), orphan sessions (#33907), TUI sid desync
(#36777), FTS search gaps + duplicate sidebar entries (#45117), null
continuation cwd (#42228), and title-rename dead-ends (#48989). It also
forced a large defensive apparatus (compression lock, contextvar/env/
logging triple-sync, orphan finalization, gateway SessionEntry
re-propagation, tip projection) whose only job is surviving a
mid-conversation id change.
Add a compression.in_place config flag (default False during rollout).
When True, compaction rewrites the transcript and rebuilds the system
prompt but keeps the SAME session_id: no end_session, no child row, no
title renumber, no contextvar/logging re-sync, no memory/context-engine
session-switch. The conversation keeps one durable id for life, like
Claude Code / Codex. Compaction is lossy by design — the pre-compaction
transcript is summarized away, not archived.
The rotation path is unchanged when the flag is off (moved verbatim into
an else branch). Staged rollout: this PR ships the option behind a
default-off flag for live validation; a follow-up flips the default and
deletes the now-redundant rotation machinery, superseding the 14 open
band-aid PRs in this area.
- hermes_cli/config.py: add compression.in_place (default False), documented
- agent/agent_init.py: resolve the flag -> agent.compression_in_place
- agent/conversation_compression.py: branch compress_context() on the flag
- tests/run_agent/test_in_place_compaction.py: in-place invariants +
rotation regression guard + config default
The pre-flush of current-turn messages (#47202) runs in BOTH modes, so no
boundary data loss. Prompt-cache invariant preserved: the system-prompt
rebuild is the same single sanctioned invalidation that already happens
during compaction — no NEW invalidation. Message alternation preserved.
The image-too-large reactive shrink (try_shrink_image_parts_in_messages)
conflated two independent constraints: it always rejected a resize whose
re-encoded bytes were >= the original, even when the shrink was driven by a
PIXEL-DIMENSION cap (Anthropic many-image 2000px) rather than the byte budget.
Downscaled screenshot PNGs routinely re-encode LARGER in bytes, so the
dimension-correct result was discarded and the image left oversized -> the
provider re-rejected on retry and the session wedged forever.
Fix: track which constraint triggered the shrink (bytes vs dimension) and gate
the accept on the SAME axis.
* dimension path: accept the result as long as it is now within max_dimension,
regardless of byte size (verify via Pillow; fall back to the byte gate only
when the re-encode can't be decoded).
* bytes path: still require bytes to shrink, but ALSO re-check the per-side cap
when it's active — _resize_image_for_vision returns a best-effort, possibly
over-cap blob when it exhausts its halving budget on a very-high-aspect
image, so a byte-shrink alone can leave it over the dimension cap and
re-brick on retry.
Extend the unshrinkable-oversized guard to the pixel axis so a partial shrink
doesn't burn the one-shot retry.
Single shared agent path -> fixes CLI, TUI, and gateway alike.
Adds a real-Pillow runnable proof (repro_48013_image_shrink_brick.py) that
reproduces the issue's per-image table (bricks 3/5 before, passes 5/5 after)
plus unit invariants for the dimension and bytes accept/reject paths,
partial-progress accounting, and the bytes-path still-over-cap regression
surfaced by adversarial review.
Closes#48013
Locks the contract that a non-retryable failure (a Cloudflare 403
"managed challenge" page) returns a short, HTML-free `error` field —
guarding the field path where the raw page was dumped to Discord as
~31 messages.
The test drives the standard chat-completions path with a concrete
model so the turn actually reaches `client.chat.completions.create`,
where the mocked 403 is raised. It asserts the create call happened
(guarding against a vacuous pass — an empty model on the Codex
Responses path would otherwise abort on a validation ValueError before
any API call) and that the summarized error includes "403" while
excluding <html> / _cf_chl_opt. The non-retryable abort path is
provider-agnostic; a Cloudflare managed-challenge 403 can surface on
any provider behind Cloudflare.
Resolves conflicts from the OpenViking churn that merged after #32445 was
opened (#48042/#47662 session-switch + write hardening, #47311/#47973):
- plugins/memory/openviking/__init__.py: keep both __init__ field groups
(the PR's _runtime_start_* alongside main's _prefetch_threads/_shutting_down).
- tests/plugins/memory/test_openviking_provider.py: keep BOTH the PR's new
setup-validation tests and main's session-switch/concurrency tests (disjoint
additions to the same region).
Two fixes layered while reconciling (contributor work otherwise preserved):
- Restore the merged tenant-header contract (#22414/#21232). The PR had changed
_VikingClient defaults to '' and made empty account/user OMIT the tenant
headers; main's contract is that empty falls back to 'default' and the
X-OpenViking-Account/User headers are ALWAYS sent (ROOT API keys need them).
Reverted the constructor to 'account or os.environ.get(..., "default")' and
updated the two PR tests that asserted the omit-when-empty behavior.
- Close a secret-file TOCTOU in the setup writers. _write_env_vars and
_write_ovcli_config wrote the api_key/root_api_key file and chmod 0600
AFTERWARD, leaving a world-readable window on newly-created files. Added
_precreate_secret_file() to create with 0600 before any secret bytes land.
Salvage corrections on top of @XVVH's #44341:
- Make native web_search injection a 1:1 swap for an already-present client
web_search function, NOT an additive grant. The original unconditionally
appended {"type":"web_search"} on every is_xai_responses turn with any
tools, force-enabling Grok server-side search even when the user never
enabled the web toolset (bypassing Hermes web-provider config + tool-trace
plumbing). Now gated on a client web_search actually being present.
- Reconcile grok-composer context to 200000 (merged in #47908) rather than
262144; 200k is xAI's published usable context window for Composer 2.5,
262144 is the /v1/responses input+output budget.
- Update tests to match scoped behavior + add a no-web-toolset guard test.
- AUTHOR_MAP entry for #44341 salvage.
Incomplete-guard (server-side *_call items at in_progress no longer flip
has_incomplete_items) and preflight built-in-tool allowlist kept as-is.
- model_metadata: grok-composer-2.5-fast → 262144 (OAuth slug not in /v1/models)
- codex transport: inject native {"type":"web_search"} for is_xai_responses;
drop client web_search to avoid duplicate-name 400s
- codex adapter: do not treat in-progress server-side *_call items as incomplete
- tests: adapter, transport build_kwargs, model_metadata, oauth recovery
Avoid applying text-only persist_user_message overrides to multimodal current-turn user messages. Early crash-resilience persistence mutates the same messages list later used for the API call, so clobbering list content drops ACP image blocks before model dispatch.\n\nAdd regression coverage for both text override behavior and multimodal preservation.\n\nCloses #44242
Follow-up to salvaged PR #41624:
- Remove stray urllib.parse import in run_agent.py (cherry-pick cruft, unused)
- Add tests: session:compress emits with correct context, no-callback is
safe, and a callback exception does not break compression
When display.memory_notifications is set to 'verbose', skill_manage
notifications now show meaningful change details instead of just the
generic tool message.
Before (verbose mode):
💾📝 Patched SKILL.md in skill 'gogcli' (1 replacement).
After (verbose mode):
💾📝 Skill 'gogcli' patched: "old pitfall text..." → "new pitfall text..."
Changes:
- skill_manager_tool.py: _patch_skill() now includes old/new string
previews (truncated to 200 chars) in the result via '_change' key.
_create_skill() and _edit_skill() include skill description from
frontmatter for verbose create/edit notifications.
- run_agent.py: Background review notification builder now reads the
'_change' dict from skill tool results and formats descriptive
notifications per action type (patch → old→new diff, create/edit →
description preview). Falls back to generic message when _change
data is unavailable (backwards compatible).
This is especially useful when subagents patch skills, since neither
the user nor the parent agent can see what the subagent changed.
Keep request dump writes on the shared atomic JSON path, add regression coverage for request body/error/stdout redaction, and map the salvaged contributor email for release attribution.
A Claude refusal (HTTP 200, stop_reason="refusal", empty content) was
laundered into a generic retry loop and surfaced as a misleading
"rate limited / invalid response" or "no content after retries" error,
burning paid attempts reproducing a deterministic refusal.
This hit two distinct paths:
- Direct Anthropic (anthropic_messages): validate_response rejected the
empty-content refusal *before* normalize_response mapped refusal ->
content_filter, so it fell into the invalid-response retry loop.
- Nous Portal / OpenAI-compatible (chat_completions): the portal surfaces
a Claude refusal via message.refusal with empty content, which sailed
past validation and died in the empty-response retry loop.
Fix (one unified content_filter dispatch for all backends):
- AnthropicTransport.validate_response: accept empty content when
stop_reason == "refusal" so it flows to normalize_response.
- ChatCompletionsTransport.normalize_response: promote message.refusal to
content + a content_filter finish reason.
- conversation_loop: handle finish_reason == "content_filter" - fire the
api_request_error hook (content_policy_blocked), try a configured
fallback once, else return a clear terminal refusal message. Never retry
a deterministic refusal.
Supersedes #43084, which fixed only the direct-Anthropic path and could
not reach the chat_completions/portal path.
Tests: transport-level (validate_response refusal, message.refusal
promotion) + end-to-end loop (refusal surfaced, exactly one API call).
(cherry picked from commit 01f546f92c)
Parse provider-reported image pixel ceilings so many-image Anthropic requests can recover by shrinking Retina screenshots below the stricter limit instead of retrying the same rejected payload.
Follow-up to the #44837 clamp: a min() clamp only fixes cursor overshoot
past the new end of the list. When repair_message_sequence drops/merges
messages at indexes below the cursor, the clamp leaves the cursor pointing
past unflushed rows and the turn-end flush silently skips them.
Extract repair_message_sequence_with_cursor(): snapshot the flushed prefix
by object identity before repair, then recompute the cursor as the count
of surviving flushed messages. Falls back to the clamp when no snapshot is
available. Keeps the safety guard in _flush_messages_to_session_db.
Adds targeted tests for overshoot, before-cursor compaction, no-repair,
bare-agent, and the flush guard.
The subscription-cap usage gauge (50/75/90% bands) ignored purchased
(top-up) credits: a sub user with top-up funds got a sticky warn banner
at 90% of their cap — permanently at >=100%, alongside grant_spent —
despite being fully able to keep inferencing. The cap is the wrong
denominator for an account that can keep spending.
- evaluate_credits_notices: purchased_micros > 0 suppresses the usage
band (grant_spent already covers the cap-reached + top-up case with
the remaining balance). A top-up landing mid-session clears any
showing band; spending top-up down to 0 resumes the gauge.
- New display.credits_notices config (default true): false silences all
credits notices. State capture and /usage are unaffected. Read once
per agent (cached) in _emit_credits_notices, fail-open true.
- Docs: configuration.md display block.
Multimodal turns carry message content as a list of typed parts
({type: "text"|"image_url", ...}). _sync_external_memory_for_turn
passed that list straight into MemoryManager.sync_all, and providers
feed it to regexes — Honcho's sync_turn calls sanitize_context, where
re.sub raised 'expected string or bytes-like object, got list'. Every
turn with an attached image silently never synced.
Flatten to plain text at the boundary: text parts joined, images noted
as an [N image(s)] marker so the attachment isn't erased from recall.
Fixing here covers all providers instead of patching each plugin.
(cherry picked from commit 705bdb6ffe)
IAM policies scoped to bedrock:InvokeModel only (a common least-privilege
setup) reject converse_stream() with AccessDeniedException. The agent loop
hard-prefers streaming and the denial never matched the 'stream not
supported' auto-fallback, so InvokeModel-only users looped on AccessDenied
forever.
- agent/bedrock_adapter.py: new is_streaming_access_denied_error()
detector (ClientError code check + wrapped-SDK message match);
call_converse_stream() falls back to converse() on denial.
- agent/chat_completion_helpers.py: bedrock_converse streaming branch
retries inline via converse() and sets _disable_streaming so later
turns skip the doomed stream attempt; the chat-completions retry
block also recognizes the denial for the AnthropicBedrock SDK path
(message pre-check avoids importing bedrock_adapter — and its lazy
boto3 install — for unrelated providers).
Both paths print a one-line notice telling the user which IAM action
restores streaming.
The thinking-signature recovery in agent/conversation_loop.py popped
reasoning_details from messages, then continued to retry. That had two
defects.
First, the strip never reached the wire payload. api_messages is built
once at the start of the turn by shallow-copying every entry in messages
(line 919 area). Each api_messages entry has its own reference to the
same reasoning_details list. When build_api_kwargs runs on every retry
iteration of the inner while-loop, it consumes api_messages, not
messages. Popping reasoning_details from messages left api_messages
untouched, so the retry's request still carried the same thinking
blocks Anthropic had just rejected. The classifier latched
thinking_sig_retry_attempted = True after the first attempt, and the
loop terminated with max_retries_exhausted on the same 400.
Second, the pop mutated the canonical message list. messages is the
same list _persist_session writes to state.db and the session
transcript, so a single recovery permanently wiped every signed
thinking block from the stored conversation. Subsequent turns reloaded
the stripped state, hit the same 400 ('invalid signature' or 'cannot
be modified', see #24107), and the agent stopped responding entirely.
Cascading compaction-ended sessions then chained off the corrupted
parent and the affected chat could not produce a response on any
future turn.
Move the strip onto api_messages, which is the API-call-time list
rebuilt into kwargs on every retry. messages is no longer touched, so
disk I/O stays clean and the recovery actually reaches the wire.
Observed against the native Anthropic Messages API on claude-opus-4-7
and claude-opus-4-8 with the interleaved-thinking-2025-05-14 beta on
hermes-agent 0.12.0 and 0.14.0. PR #24107 narrows the trigger; this
change makes the recovery do what it always claimed to do, and
prevents the destructive aftermath.
Tests cover the api_messages strip in isolation: pop on a shallow copy
does not affect the source, the canonical messages list survives the
strip, idempotency on a duplicate firing path, and a no-op when no
reasoning_details exist on the messages.
Related: #24107, #26959, #17861.
Third-party OpenAI-compatible endpoints (self-hosted gateways, OpenRouter,
Azure proxies) fronting gpt-4o / gpt-4.1 / gpt-5+ / o1-o4 models silently
received max_tokens and 400'd with unsupported_parameter, because the three
kwarg-selection sites only checked base_url_hostname(...) == "api.openai.com"
and fell through to max_tokens on every other host. The constraint is
enforced server-side by the model family, not by the URL, so name-based
detection is required as a fallback.
Changes:
- utils.py: new shared helper model_forces_max_completion_tokens(model) that
prefix-matches gpt-4o, gpt-4.1, gpt-5, o1, o3, o4 families on normalized
(lowercased, vendor-prefix-stripped) names.
- run_agent.py: _max_tokens_param ORs the helper into the URL check.
- agent/auxiliary_client.py:
- auxiliary_max_tokens_param gains an optional keyword-only model arg.
- _build_call_kwargs inline branch applies the same check for both
provider == "custom" and non-custom paths.
Tests:
- tests/test_model_forces_max_completion_tokens.py: 31 new cases covering
positive families, negatives (classic gpt-4, claude, llama, mistral, qwen,
deepseek), vendor prefixes, case-insensitivity, whitespace, None/empty,
and substring-not-prefix guards.
- tests/run_agent/test_run_agent.py::TestMaxTokensParam: 5 new model-based
cases (custom + gpt-5.4, openrouter + gpt-4o-mini, custom + o1-preview,
classic gpt-4-turbo keeps max_tokens, llama3 keeps max_tokens).
- tests/agent/test_auxiliary_client.py::TestAuxiliaryMaxTokensParam: new
class, 7 tests covering the URL x model matrix.
The test keyed the 'which call raises' decision on a shared invocation
counter (first call → raise, second → success), then asserted the error
landed in messages[0] (c1) and success in messages[1] (c2). But
_execute_tool_calls_concurrent runs the two web_search calls on a thread
pool with no ordering guarantee — c2's handler can be invoked first, take
the 'first call raises' branch, and the error ends up in messages[1].
Results are ordered by tool_call_id, so messages[0] (c1) was then 'success'
and the assertion failed.
It passed in isolation but reliably failed under CI's full parallel slice
(8 xdist workers) where the scheduler actually interleaves the two handlers.
Fix: tie the raise to a specific tool call via its arguments (q=boom raises,
q=ok succeeds) instead of invocation order, and assert tool_call_id ↔ content
pairing explicitly. Deterministic regardless of thread scheduling — verified
10/10 in isolation and the full TestConcurrentToolExecution class (32) green.
* fix(stream): don't report dropped mid-tool-call streams as output truncation
A streaming tool call whose SSE ends with no finish_reason (the upstream
delivers the tool name + opening '{' then closes the connection cleanly,
no terminator, no [DONE]) was stamped finish_reason='length' by the mock
builder. That routed it through the output-cap truncation path: 3 useless
max_tokens-boosted retries, then the misleading 'Response truncated due to
output length limit' error — even though the model never reported hitting
any cap.
Reproduced live on nvidia/nemotron-3-ultra:free via the Nous dedicated
endpoint, which stalls/drops during large tool-arg generation (50s-4m41s).
Now: when tool args are incomplete AND the provider sent no finish_reason,
tag the response as a partial-stream stub so the loop reports an honest
mid-tool-call drop and asks the model to chunk its output (existing
continuation machinery), instead of escalating output budget and lying.
A provider-reported finish_reason='length' still takes the real-truncation
path unchanged.
* test(stream): update truncated-tool-args test for drop-vs-cap split
test_truncated_tool_call_args_upgrade_finish_reason_to_length pinned the
old behaviour where ANY incomplete tool args → finish_reason='length' with
tool_calls preserved. That single-chunk-no-finish_reason scenario is exactly
the mid-tool-call stream drop now reclassified as a partial-stream stub.
Split into two tests matching the new contract:
- no finish_reason + incomplete args → PARTIAL_STREAM_STUB_ID, tool_calls=None,
_dropped_tool_names set (the drop path)
- explicit finish_reason='length' + incomplete args → tool_calls preserved,
'length' upgrade unchanged (the genuine output-cap path)
Lift the 5 agent-construction/session-resume methods out of HermesCLI into
hermes_cli/cli_agent_setup_mixin.py:CLIAgentSetupMixin. Behavior-neutral; cli.py
14139 -> 13492 LOC.
Methods moved (~647 LOC): _ensure_runtime_credentials, _resolve_turn_agent_config,
_init_agent, _preload_resumed_session, _display_resumed_history. All self.* calls
resolve unchanged via the MRO (HermesCLI(CLIAgentSetupMixin, CLICommandsMixin)).
Import split (same recipe as #41942): 2 neutral deps (sys, _escape) imported at
the mixin module top; 12 cli.py-internal helpers/constants (AIAgent, ChatConsole,
CLI_CONFIG, _cprint, _DIM, _RST, _accent_hex, ...) imported lazily per-method
(from cli import ...) so the mixin never imports cli at module scope -> no cycle.
Repointed one source-inspection change-detector (test_callable_api_key.py) to read
the mixin file where the method now lives.
test_gateway_run_clamped read gateway/run.py asserting the /usage stats handler
clamps pct with min(100, ...). That handler moved to gateway/slash_commands.py
in this PR's extraction; repoint the guard so it still fires on clamp removal.
tests/run_agent/ + tests/gateway/ 8024 passed / 0 failed.
VolcEngine's api/plan endpoint occasionally leaks raw XML attribute
fragments into tool_use.name when its protocol-translation layer
converts the model's native XML-style tool emission to Anthropic
Messages tool_use blocks, producing names like:
terminal" parameter="command" string="true
execute_code" parameter="code" string="true
session_search" parameter="session_id" string="true
The corruption happens server-side at the provider, but it breaks
every tool call for affected users — no normalization rule in
repair_tool_call can rescue them, so each request runs through three
retries and then aborts as partial.
Add an early sanitizer in agent_runtime_helpers.repair_tool_call that
trims at the first ' " ', " ' ", '<', or '>' character (idx > 0
only) so the rest of the existing repair pipeline (lowercase /
snake_case / fuzzy match) can resolve the cleaned name normally.
Whitespace is deliberately NOT a separator — the legitimate
"write file" -> write_file repair path (covered by
test_space_to_underscore) must keep working.
Tests: 11 new regression cases in TestVolcEngineXmlPollution
covering all three observed polluted names, CamelCase + pollution
mix, single-quote variants, angle-bracket variants, clean-name
passthrough, and the whitespace-preservation guard. All 18 pre-
existing repair tests still pass (29 total in the file).
Phase 1 of the god-file decomposition plan. run_conversation's ~470-line
once-per-turn setup block (stdio guarding, retry-counter resets, user-message
sanitization, todo/nudge hydration, system-prompt restore-or-build,
crash-resilience persistence, preflight compression, the pre_llm_call hook, and
external-memory prefetch) is moved verbatim into build_turn_context(), which
returns a TurnContext dataclass the loop unpacks.
Behavior-neutral move-and-name refactor: the builder mutates `agent` exactly as
the inline code did; only the locals the loop reads back are returned.
- run_conversation: 4602 -> 4217 LOC (-385)
- agent/conversation_loop.py: 4965 -> ~4580 LOC
- new agent/turn_context.py: focused, dependency-injected, unit-tested in isolation
Tests: tests/run_agent/ 1570 passed / 0 failed under per-file process isolation.
Relocation follow-ups: 413_compression mocks now patch both module references;
nudge/on_turn_start source-inspection guards point at the extracted module.
Xiaomi MiMo (and potentially other providers) support multimodal user
messages but reject list-type tool message content with 400 'text is not
set'. Previously this was handled reactively — the API call would fail,
images would be stripped, and the request retried, losing visual info.
Fix: add supports_vision_tool_messages field to ProviderProfile (default
True). Xiaomi sets it to False. _tool_result_content_for_active_model
now checks this field proactively and returns a text summary instead of
list content, avoiding the round-trip failure entirely.
When summary_target_ratio is large (e.g. 0.45) and the context_length is
moderate (e.g. 96000), the soft_ceiling (token_budget * 1.5) can exceed
the total transcript size. _find_tail_cut_by_tokens walks the entire
transcript without breaking early, and the resulting compress window is
either empty (compress_start >= compress_end) or a single message whose
summary-of-one overhead saves ~0 tokens.
Both outcomes cause a no-op compression that does not increment
_ineffective_compression_count, so should_compress() returns True on
every subsequent turn and the loop repeats endlessly.
Fix (two layers):
1. _find_tail_cut_by_tokens: when the backward walk consumed the entire
transcript without breaking (cut_idx <= head_end and accumulated <=
soft_ceiling), re-walk with the raw (non-inflated) token budget to
find a meaningful cut that gives the summarizer a useful middle window.
2. compress(): when compress_start >= compress_end, increment
_ineffective_compression_count and log a warning so the existing
anti-thrashing guard in should_compress() can break the loop.
Fixes#40803
Persist the inbound user turn before provider/tool execution so a crash
before run_conversation() (e.g. provider/httpx client init failure) keeps
the inbound message in the transcript. Repair stale/missing SSL_CERT_FILE
state on gateway startup, and avoid duplicate gateway fallback writes.
Custom OpenAI-compatible endpoints sitting behind a gateway/WAF can reject
the OpenAI Python SDK's default identifying headers (User-Agent: OpenAI/Python,
X-Stainless-*) and return an opaque 502/4xx even though the same request body
succeeds under curl. There was no supported way to override those headers.
Add a model.default_headers config key whose values are merged onto the
OpenAI client's default_headers, taking precedence over provider- and
SDK-supplied defaults. Applied at client construction and on every credential
swap / client rebuild so the override survives reconnects. No-op for native
Anthropic / Bedrock modes and when unconfigured.
* test(kimi): align stale parity/profile tests with thinking-xor-effort contract
ce4e74b3 (fix(kimi): send thinking xor reasoning_effort, never both)
changed the Kimi profile to emit at most one of extra_body.thinking or a
top-level reasoning_effort, and added tests/plugins/model_providers/test_kimi_profile.py
to pin it — but left two older test files still asserting the removed
'send both' behavior, turning main red for every PR branched after it.
Update the stale assertions to the xor contract:
- explicit recognized effort (low|medium|high) -> reasoning_effort only,
no thinking
- enabled w/o effort, or no reasoning_config -> thinking:enabled only,
no reasoning_effort
- disabled -> thinking:disabled only
No production change.
* test(kimi): cover remaining xor stale assertions (profile_wiring, run_agent)
Two more test files asserted the pre-ce4e74b3 'thinking + reasoning_effort
together' behavior — landed in a different CI shard so they surfaced only
after the first batch went green:
- tests/providers/test_profile_wiring.py::TestKimiProfileParity (2)
- tests/run_agent/test_run_agent.py::TestBuildApiKwargs (3: kimi-coding,
moonshot, moonshot-cn)
Same realignment to the xor contract: default/enabled-without-effort emits
thinking:enabled and no reasoning_effort; explicit effort emits
reasoning_effort only. Verified by running the full provider +
TestBuildApiKwargs Kimi surface (202 passed) plus a codebase-wide grep for
any remaining paired thinking+effort assertion (none).
_hook_jsonable() referenced SimpleNamespace without importing it, so
sanitizing any hook payload that contained one raised
NameError: name 'SimpleNamespace' is not defined.
Bedrock, Codex-responses, and the auxiliary client build their
response / message / tool_call objects as SimpleNamespace and hand the
raw objects to the post_api_request hook. The hook call sites swallow
exceptions (except Exception: pass), so the crash silently dropped the
observability hook for those providers.
Add the missing `from types import SimpleNamespace` and a regression
test covering the SimpleNamespace sanitization path.
A memory provider tool whose name collides with a built-in core tool
(e.g. clarify, delegate_task) was skipped from agent.tools at init but
lingered in MemoryManager._tool_to_provider, where the has_tool dispatch
branch could route a call to a tool that was never registered (#40466).
Block the collision at registration instead of patching dispatch:
- MemoryManager.add_provider rejects any tool whose name is in
_HERMES_CORE_TOOLS (warn + skip), so it never enters the routing table.
- get_all_tool_schemas applies the same filter, so the manager never
advertises a schema it would refuse to route.
Built-ins always win, matching the invariant used by the TTS/browser/
search provider registries. Makes the dispatch-hijack structurally
impossible regardless of branch ordering.
Closes#40466.
* feat(tui): HERMES_DEV_CREDITS live-spend dev readout (L0 tracer for usage-aware credits)
L0 of the usage-aware-credits feature: a dev-only, env-gated tracer that
exercises the real header -> CreditsState -> TUI pipe end-to-end behind
HERMES_DEV_CREDITS, de-risking the L1/L5 build before the notice policy exists.
- agent/credits_tracker.py: CreditsState + parse_credits_headers (headers are
strings -> paid_access via == "true", never bool(); retain-last-known; only
subscription_micros may be negative; *_usd kept verbatim).
- run_agent.py: _capture_credits / get_credits_state / get_credits_spent_micros,
session-start baseline latch, + dev-gated "credits" capture log.
- agent/chat_completion_helpers.py: capture on the streaming response.
- agent/agent_init.py: init _credits_state + _credits_session_start_micros.
- tui_gateway/server.py: _get_usage emits dev_credits_spent_micros only when flagged.
- ui-tui appChrome.tsx / types.ts: cents delta status segment + "(dev credits)" banner.
Off by default; silent for normal users. Validated live against staging
(capture log delta matches the TUI segment). Throwaway consumer (readout/log/
banner); credits_tracker + the capture plumbing are the real feature foundation.
* test(credits): lock parser under 9-state matrix + harden validation (L2)
Add tests/agent/test_credits_tracker.py with 92 tests covering the 9-state
matrix (healthy, sub_90pct, grant_exhausted, purchased_only, tool_pool_free,
depleted, debt, missing, no_org) plus validation edge cases: version strict==1
with warn-once latch for v>1, bool-string trap (paid_access/tool_pool_gated_off
== "true"/"false", never bool()), half-pair subscription limit treated as
both-absent while parse succeeds, USD regex ^-?\d+\.\d{2}$, non-int micros
→ None, negative non-subscription micros → None, as_of_ms junk → None, zero
limit ZeroDivision guard.
Harden agent/credits_tracker.py to match the spec:
- Add tool_pool_micros/tool_pool_gated_off/from_header fields to CreditsState
- Add depleted property (== not paid_access, never remaining==0)
- Change used_fraction guard to key off subscription_limit_micros (the actual
denominator) not denominator_kind (metadata)
- Replace fail-soft _safe_int with a sentinel-returning variant; full validation
now returns None on any malformed field rather than silently defaulting
- Add module-level warn-once latch for version > 1
- Add USD regex validation; add denominator_kind allow-list check
- Parse x-nous-tool-pool-* prefix headers (not x-nous-credits-tool-pool-*)
* feat(credits): notice spine — AgentNotice + notice_callback/notice_clear_callback + TUI binding (L1)
L1 of usage-aware credits: the driver-agnostic notice delivery spine that L4's
policy will fire through and L5's TUI render will consume.
- agent/credits_tracker.py: AgentNotice dataclass (text/level/kind/ttl_ms/key/id;
kind defaults "sticky", kept TTL-expressive for a future config seam).
- run_agent.py: AIAgent gains notice_callback + notice_clear_callback slots and
_emit_notice / _emit_notice_clear emitters (swallow all callback errors — a
notice must never break the agent loop; no-op when unbound).
- agent/agent_init.py: thread both callbacks through init_agent.
- tui_gateway/server.py: bind both in _agent_cbs → notification.show / notification.clear
WS events (snake_case payload, matching the existing gateway-event convention).
- ui-tui/src/gatewayTypes.ts: notification.show / notification.clear arms on GatewayEvent.
- tests/run_agent/test_notice_spine.py: 15 tests (emitter fire + fail-open + no-op,
signature threading, TUI binding payload shape).
Messaging push is out of v1 (binds neither callback). CLI binding + the TUI render/
decode land with L4 (firing) and L5 (render) so turn-end flush is wired correctly.
* feat(credits): threshold reconciliation policy + tests (L4.1)
* feat(credits): wire threshold policy into capture + latch (L4.2)
After a fresh header parse, _capture_credits runs evaluate_credits_notices against
the agent's _credits_latch and emits the result — clears first, then shows (so a
recovered depletion clears before the "restored" success lands, and depleted wins
the latest-wins slot). Gated on a bound notice_callback: messaging (no callbacks)
still caches state for /usage but runs no policy. Parse stays fail-open (miss →
keep last-known); the eval/emit path warns on failure rather than swallowing, so a
depletion-notice bug can't vanish silently.
- run_agent.py: _capture_credits split into parse (swallow→miss) + policy (warn);
latch lazy-guarded (object.__new__ safety).
- agent/agent_init.py: init agent._credits_latch = {"active": set(), "seen_below_90": False}.
* feat(tui): render credits notices in the status bar (L5, Strategy B)
The TUI now renders the notification.show / notification.clear gateway events the
agent emits — a level-colored notice overrides the status/verb slot when not busy.
- Notice state machine on turnController (pendingNotice + dedicated noticeTimer +
show/clear/applyNotice/flushPendingNotice/clearNoticeState). createGatewayEventHandler
decodes the events and delegates.
- Render priority busy > notice > status (appChrome StatusRule); notice text rendered
verbatim (its glyph comes from the policy), shrinkable so it never clips model│ctx;
dev-credits banner + Δ segment preserved. UiState.notice is snake_case (matches wire).
- Busy-wins: a notice arriving mid-turn is held and flushed at the THREE turn-end sites
(recordMessageComplete / interruptTurn / recordError) — never idle(), which reset()
also calls (would leak across sessions); reset() clears instead.
- Dedicated noticeTimer (never statusTimer); TTL starts on visibility with an id-guard;
latest-wins cancels the prior timer; clear is key-matched (no-op on mismatch); a sticky
survives a turn (flush no-ops with no pending); session reset clears (no cross-session leak).
- 20 tests (handler/turnController logic incl. R3-C2 timer isolation + render priority).
* feat(credits): cold-start seed for new Nous sessions (L3)
A genuinely-new Nous session has no inference header yet, so seed credits state from
the authoritative GET /api/oauth/account snapshot at session start (in the new-session
branch of _restore_or_build_system_prompt — inline, since the on_session_start plugin
hook gets no agent reference). The seed runs the shared notice policy, so a session that
opens already depleted warns IMMEDIATELY rather than only after the first turn.
- Maps the nested account fields (paid_service_access → paid_access; total_usable /
subscription / purchased on paid_service_access_info; rollover on subscription), each
None-guarded; float dollars → micros via round(d*1e6), *_usd left "" (render formats
from micros — never synthesize a verbatim usd from a float).
- Magnitudes-only: no monthlyCredits on the endpoint → subscription_limit_* unset →
used_fraction None → no warn90 from the seed (% only once a header lands, per D-E).
- Provider-guarded to Nous; fail-open (any error leaves _credits_state None, never
blocks startup); paid_access unknown ⇒ True (never falsely depleted).
- run_agent.py: extracted the warm-path policy/emit block into a shared
_emit_credits_notices() so capture and the seed fire notices identically.
* feat(credits): /usage Nous credits magnitudes view + recovery trigger (L6)
Add Nous credit dollar magnitudes to /usage (subscription / top-up / total
+ rollover + renewal + portal CTA), magnitudes-only per v1 (no % until the
account endpoint exposes a denominator). Reuses the existing account-usage
render machinery via a new pure build_nous_credits_snapshot() that maps a
NousPortalAccountInfo to an AccountUsageSnapshot; no nous branch is added to
fetch_account_usage (keeps the per-provider boundary intact).
CLI /usage also doubles as a depletion-recovery trigger: a force_fresh
account fetch, kept in a SEPARATE local so it never clobbers the
header-sourced agent._credits_state (which alone carries used_fraction). If
paid access recovered while credits.depleted is latched and a notice
consumer is bound, it reuses agent._emit_credits_notices() to clear it.
Gateway /usage displays magnitudes only — messaging binds no notice
consumer, so it performs no recovery emit.
Fail-open throughout: any portal hiccup leaves /usage unaffected.
* refactor(credits): dedupe HERMES_DEV_CREDITS flag parse via shared helpers
The dev-flag truthy check was inlined in three places. Replace with the shared
utils.is_truthy_value (run_agent.py, tui_gateway/server.py — also drops a
redundant inline `import os`) and a hoisted DEV_CREDITS_MODE export in
ui-tui/src/config/env.ts (consumed by appChrome, which also stops recomputing the
env check on every render). Behaviour-preserving; identical truthy set.
* fix(credits): cut dead /usage recovery trigger + bound portal fetches (L6 review)
Adversarial review found the /usage depletion-recovery trigger dead AND broken:
the CLI binds no notice_clear_callback, the TUI runs /usage in a separate
slash-worker subprocess (its own agent/latch), and the no-clobber rule made it
evaluate stale paid_access anyway. Recovery already happens on the next inference
(warm path), so the trigger was redundant — remove it and stop the depleted
notice over-promising.
- cli.py: remove the dead recovery block; bound the /usage portal fetch with a
10s wall-clock timeout (ThreadPoolExecutor) like the per-provider fetch —
urllib's per-socket timeout is not a wall-clock guarantee.
- agent/credits_tracker.py: reword the depleted CTA to "run /usage for balance"
(no false recovery promise; /usage shows fresh magnitudes, sticky clears next turn).
- agent/conversation_loop.py: same wall-clock timeout on the cold-start seed fetch
so a stalled portal can't hang session startup; tidy its time import.
* chore(credits): dev notice-state fixtures (HERMES_DEV_CREDITS_FIXTURE)
Throwaway dev scaffolding to exercise the notice pipeline without real spend or
Redis seeding. Set HERMES_DEV_CREDITS_FIXTURE to a state name (healthy / sub_90pct
/ grant_exhausted / depleted / clear) or a file path whose contents name a state
(re-read each turn → flip states live for recovery testing). _capture_credits
injects the chosen CreditsState instead of parsing real headers and runs the
shared notice policy. Deletable with the rest of the HERMES_DEV_CREDITS scaffolding.
* feat(credits): /usage monthly-grant % gauge
The portal /api/oauth/account subscription block now carries monthly_credits
(the per-period grant allowance, the % denominator). The consumer parsed
monthly_charge but dropped monthly_credits, so /usage stayed magnitudes-only.
Capture monthly_credits into NousPortalSubscriptionInfo + _subscription_from_payload.
build_nous_credits_snapshot emits a Subscription usage window (real % used, routed
through the existing render machinery) when monthly_credits is a finite positive
denominator and credits_remaining is finite and <= cap; otherwise it degrades to
magnitudes-only (older portals, rollover-over-cap, or non-finite payloads).
Guards (adversarial-review-driven): reject non-finite operands (json.loads parses
bare NaN/Infinity by default → would render $nan + a false 100% used), reject
bools, guard div-by-zero (cap>0), and suppress the gauge when remaining > cap
(rollover spanning the period makes the cap a nonsensical denominator → the
$X-of-$Y detail would read as a contradiction). Debt (remaining<0) clamps to 100%.
Money rule preserved: the ratio + magnitudes are computed from numeric float
account fields via display formatting, never by parsing a server *_usd string
(there are none on these dataclasses).
13 gauge tests added (tests/agent/test_nous_credits_gauge.py).
* fix(credits): show /usage Nous block whenever a Nous account is present
/usage runs in a slash-worker subprocess whose resolved inference provider is
often not "nous" even when the user has a Nous account, so gating the Nous
credits block on (provider == "nous") hid it entirely — the account data was
fully available but never rendered.
Gate instead on "a Nous account is logged in": a cheap local auth-state lookup
(get_provider_auth_state('nous') has an access_token) decides whether to attempt
the portal fetch, regardless of which provider inference runs on. In the gateway
the block is also lifted out of the 'if provider:' scope so a Nous-credentialled
user with another (or no) resident inference provider still sees their balance.
Fail-open and the per-fetch wall-clock timeout are preserved.
* fix(credits): show /usage Nous block when there's no live agent (TUI slash-worker)
In the TUI, /usage runs in a slash-worker subprocess that resumes the session
WITHOUT building an agent (self.agent is None), so _show_usage early-returned
"(._.) No active agent" before ever reaching the Nous credits block — which is
agent-independent (a portal fetch gated on Nous auth-state). Extract the block
into _print_nous_credits_block() and run it at the no-agent / no-calls
early-returns too (returns True if it printed, so the fallback message only
shows when there's genuinely nothing).
Verified live against staging: the block + monthly-grant gauge now render in the
slash-worker /usage path (previously hidden). The plain CLI REPL + messaging
paths are unchanged (they have a live agent).
* feat(credits): escalating 50/75/90 usage bands (single status line)
Replace the lone 90%-used warning with three escalating bands (50 info, 75 warn,
90 warn) shown as ONE status-bar line: it displays the highest band the
subscription grant has crossed, replaces the line as usage climbs, steps back
down on recovery, and clears below 50%. No stacking, no per-turn churn.
Bands live in a tunable CREDITS_USAGE_BANDS list; the policy derives everything
from it. Single notice key (credits.usage) with a usage_band latch field so the
notice only re-emits when the band actually changes. The crossing gate
(seen_below_90) is preserved so a fresh live session that opens mid-range stays
quiet until it has been observed below the lowest band (cold-start primes it when
it wants an open-high warning). Denominator math unchanged: % = subscription
grant burn (cap - grant_remaining)/cap, clamped [0,1]; top-up never moves the %.
Migrated test_credits_policy.py to the new key + added TestUsageBands (climb,
step-down, recovery-clear, idempotent, inclusive boundaries).
* feat(credits): hydrate notices at session OPEN via shared seed (TUI + first-turn)
Notices previously only fired inside a conversation turn (first message), so a
session that opened already depleted / past a usage band showed nothing at
'ready'. Extract the cold-start seed into a shared seed_credits_at_session_start()
and call it (a) in the TUI/desktop agent build right after the notice callback is
wired (fires at 'ready', before any message) and (b) as the first-turn fallback in
conversation_loop. Idempotent (skips once _credits_state exists) and fail-open.
The seed now maps monthly_credits -> subscription_limit_micros +
denominator_kind='subscription_cap', so used_fraction is computable at seed time
and usage-band warnings (not just depletion) hydrate on open. Primes the crossing
latch so a session opening already in a band warns immediately. Degrades to
depletion-only when monthly_credits is absent (older portals).
Adds test_credits_cold_start.py covering open-at-band, depletion, debt, no-cap
degradation, and the shared seed (fires/idempotent/skips-non-nous).
* feat(credits): /usage monthly-grant % gauge + fixture support + TUI surfacing
agent/account_usage.py: build_nous_credits_snapshot emits a subscription %% gauge
when the portal supplies a positive, finite monthly_credits denominator with
remaining <= cap (guards reject NaN/Infinity and rollover-over-cap, which would
render $nan or a contradictory $X-of-$Y); degrades to magnitudes-only otherwise.
Adds shared nous_credits_lines() (auth-gated, wall-clock-bounded portal fetch) so
the CLI and TUI /usage render the same block, and _snapshot_from_credits_state()
so HERMES_DEV_CREDITS_FIXTURE drives /usage offline too.
TUI: session.usage RPC carries credits_lines (agent-independent) and the /usage
panel renders them regardless of API-call count or resume state — previously the
TUI's separate /usage implementation only showed token counts.
Money rule preserved: %% and magnitudes come from numeric float account fields via
display formatting, never by parsing a server *_usd string.
* feat(credits): CLI REPL inline notices (parity with TUI)
The plain CLI agent bound no notice callbacks, so credit notices were TUI-only.
Bind notice_callback/notice_clear_callback on the CLI AIAgent; _on_notice renders
a single level-colored line above the prompt (error red / warn yellow / success
green / info dim) via _cprint, and seed credits at session open so a depletion or
usage-band warning shows before the first message — the same hydration the TUI
got. _on_notice_clear is a no-op (the REPL prints lines, no persistent slot).
* test(credits): add sub_50pct + sub_75pct dev fixtures for the new usage bands
The fixture set jumped 10%% -> 90%%; add sub_50pct (uf 0.5 -> band 50 info) and
sub_75pct (uf 0.75 -> band 75 warn) so the new escalating bands are exercisable
via HERMES_DEV_CREDITS_FIXTURE across all three surfaces (notice, session-open
seed, /usage gauge).
* fix(credits): usage-band notice clears on next prompt (not sticky-forever)
A 50/75/90 usage heads-up was sticky and camped the status bar indefinitely. Clear
the visible credits.usage notice when a new turn starts (startMessage), so it shows
until your next prompt then yields. The server latch is unchanged, so it won't
re-nag at the same band — it only re-shows when the band actually changes (climb)
or clears when usage drops below the lowest band. Depletion stays sticky.
* refactor(credits): consolidate the /usage credits block behind nous_credits_lines()
The CLI (_print_nous_credits_block) and the messaging gateway (_handle_usage_command)
each re-implemented the auth-gate + portal fetch + render, and both bypassed the
dev-fixture short-circuit that only the TUI honored — so /usage ignored
HERMES_DEV_CREDITS_FIXTURE on the CLI and in chat. Route both through the shared
agent.account_usage.nous_credits_lines() helper: one fetch/render path, one auth
gate, and the fixture works on every surface (~60 fewer duplicated lines).
The gateway usage test recorded only the last asyncio.to_thread call; /usage now
dispatches both the account fetch and the credits fetch, so it records every call
and matches the account fetch by its provider arg.
* fix(credits): keep the /usage gauge type-safe and log its fail-open path
_is_finite_num is now a TypeGuard[float], so the type checker narrows the gauge
operands (monthly_credits / credits_remaining) and the magnitudes passed to
_fmt_usd through it — no more None-operand warnings on the arithmetic. Add a debug
breadcrumb on the nous_credits_lines portal-fetch fail-open so a dead /usage block
is diagnosable in agent.log without a dev flag.
* fix(credits): harden the header tracker — prod-leak gate, hot-path probe, fire-and-forget seed
- Prod-leak guard: dev fixtures (HERMES_DEV_CREDITS_FIXTURE) now also require
HERMES_DEV_CREDITS, so a stray fixture var can't surface fabricated balances on a
real account. Matches the documented run workflow (both vars set together).
- Hot-path probe: parse_credits_headers checks for the version sentinel header
before allocating a lowercased copy of the response headers — skips that work on
every non-Nous API call. Behaviour-identical and still case-insensitive.
- Fire-and-forget seed: the real portal fetch in seed_credits_at_session_start now
runs in a daemon thread, so a slow/unreachable portal never delays session "ready"
(previously blocked up to 10s). The dev-fixture path stays synchronous; the thread
re-checks idempotency before hydrating (a live header may land first).
- Diagnostics: debug breadcrumbs on the parse and seed fail-open paths so a crashed
parser / dead seed is distinguishable from a legitimate no-headers miss.
Cold-start tests set HERMES_DEV_CREDITS alongside the fixture to match the gate.
* test(tui): fix env-timing in the StatusRule dev-credits assertion
DEV_CREDITS_MODE is read once at module load (config/env), so mutating
process.env.HERMES_DEV_CREDITS inside the test couldn't flip it — the dev-banner
assertion only passed if the env was exported before vitest started, and failed in a
normal run. Move that assertion to a sibling file that mocks config/env with
DEV_CREDITS_MODE: true (scoped, no module-reset / React-identity hazard).
* test(credits): cover the dev-fixture /usage render and usage-band clear-on-prompt
- _snapshot_from_credits_state (the offline /usage renderer) had no direct test:
lock the gauge math, the verbatim *_usd magnitudes, the depletion line and the
fixture marker, plus the no-cap (no gauge) and None-state cases.
- turnController.startMessage had no test for clearing the credits.usage notice on
the next prompt while leaving credits.depleted sticky.
* feat(credits): deliver credit notices over messaging gateways
Bind notice_callback/notice_clear_callback on the per-turn gateway agent
so usage-band / depletion / restored notices reach Telegram/Discord/Slack/
etc. Previously the messaging gateway bound neither callback, so the agent's
_emit_credits_notices early-returned and a chat user crossing a band got
nothing unless they ran /usage manually.
- render_notice_line(): AgentNotice -> single plaintext line (level glyph +
text), plaintext-only so it renders uniformly without per-platform escaping.
Fail-soft on malformed/empty notices.
- Standalone push for every notice (messaging has no persistent status bar):
route through the shared _deliver_platform_notice rail (honors private/
public delivery + thread metadata), scheduled onto the gateway loop via
safe_schedule_threadsafe from the agent's sync worker thread — same pattern
as _status_callback_sync.
- The fired-once latch lives on the cached (reused-in-place) agent and
persists across turns, so a band crosses once -> one push, no per-turn
re-nag. Re-fires only after idle-eviction rebuilds the agent (a reminder).
- Recovery ('Credit access restored') rides the show path (emitted as a
success notice, not a clear). notice_clear_callback is a no-op: a sent
platform message can't be cleanly retracted.
Tests: render glyph/levels/fail-soft + public/private delivery seam through
_deliver_platform_notice + no-adapter no-op.
* fix(credits): don't double the glyph on messaging notices
render_notice_line prepended a per-level glyph, but the notice policy already
bakes the glyph into the text (and the TUI + CLI render it verbatim) — so every
credit notice over messaging came out doubled ("⚠ ⚠ Credits 90% used",
"⛔ ✕ Credit access paused"). Emit the text verbatim instead; drop the now-dead
level→glyph map.
The render tests fed glyph-less text (and the success case only checked
startswith), so the doubling slipped through. Rework them around the verbatim
contract and add an end-to-end regression that runs real evaluate_credits_notices
output through render_notice_line and asserts the line is returned unchanged.
A steer rides inside a tool result (the only role-alternation-safe slot
mid-turn), so a bare "User guidance:" line reads as untrusted tool content —
well-behaved models refuse it as suspected prompt injection (observed live:
"I only follow instructions from you directly, not ones injected through
command results").
- Wrap steers in a bounded, self-describing [OUT-OF-BAND USER MESSAGE] marker
(prompt_builder.format_steer_marker), shared by both drain sites.
- Add STEER_CHANNEL_NOTE to the core system prompt so the model expects this
exact marker and trusts it as a genuine user message — while still ignoring
lookalikes buried in tool/web/file output. Static text → byte-stable prompt,
no prompt-cache regression; gated on the agent having tools.
- Desktop: steer ack is now an inline transcript note (⏩ steered · …) instead
of a toast.
Marker is intentionally static (not a per-session nonce) to honor the
byte-stable system-prompt caching policy; nonce hardening noted as follow-up.
* fix: respect disabled auto-compaction on context overflow
Port from anomalyco/opencode#30749.
When compression.enabled is false, NO automatic compaction trigger may
fire. The proactive token-threshold paths (preflight + post-response
should_compress gate) already honoured the setting, but the three
provider-overflow recovery paths in the agent loop — long-context-tier
429, 413 payload-too-large, and context-overflow — called
_compress_context() unconditionally, silently compressing and rotating
the session against the user's explicit choice.
Add a single guard at the top of the overflow-recovery dispatch: when
compression is disabled and the error is one of those three overflow
classes, surface a terminal error (compaction_disabled: True) telling the
user to /compress manually, /new, switch to a larger-context model, or
reduce attachments. Manual /compress (force=True) is unaffected — it never
enters this loop.
Tests: new TestOverflowWithCompactionDisabled (413 + 400 overflow don't
compress when disabled; control case still compresses when enabled).
Existing overflow-recovery tests updated to enable compaction explicitly
(they verify the recovery fires); fixture defaults flipped to True to
match production (compression.enabled defaults to True).
* fix(gemini): default native maxOutputTokens + strip OpenAI extra_body on Gemini endpoints
Two distinct failures hit users on the gemini provider with only Google
AI Studio keys set.
1. Truncation loop: build_gemini_request() only set maxOutputTokens when
max_tokens was non-None. Hermes passes None to mean "unlimited", but
Gemini's native generateContent does NOT treat an absent maxOutputTokens
as full budget — it applies a low internal default and stops early with
finishReason=MAX_TOKENS, truncating tool calls. The agent then retries
3x and refuses the incomplete call. Now default to the published 65,535
ceiling (shared by all current Gemini text models) when max_tokens=None.
2. HTTP 400 on Gemini endpoint: the chat_completions transport assembles
profile extra_body (Nous portal 'tags', reasoning, provider prefs) and
sends it via the OpenAI client to whatever base_url is resolved. When a
profile that emits extra_body (e.g. Nous) is active but the endpoint is a
native Gemini base_url — typical when only Google creds exist and a
fallback/aux call lands on Gemini — Google rejects the unknown 'tags'
field with a non-retryable 400. Strip all non-thinking_config extra_body
keys when the resolved endpoint is native Gemini.
Verified E2E against real transport code: tags stripped on native Gemini,
preserved on Nous and the /openai compat endpoint; maxOutputTokens=65535
on None, explicit values respected.
Anthropic enforces two independent ceilings per image:
1. 5 MB encoded byte size
2. 8000 px longest side
Hermes only guarded #1. A tall screenshot (e.g. 1200x12000 at 0.06 MB)
passes every byte check but fails the pixel check, returning a
non-retryable HTTP 400 that permanently bricks the conversation thread.
Fixes:
- error_classifier: add 'image dimensions exceed' pattern to
_IMAGE_TOO_LARGE_PATTERNS so the 400 is classified as image_too_large
and triggers the shrink/retry path instead of falling through to
non-retryable error.
- conversation_compression: check pixel dimensions (via Pillow) even
when byte size is under the 4 MB target. If max(dims) > 8000, force
shrink.
- vision_tools._resize_image_for_vision: add optional max_dimension param.
When set, images exceeding the pixel cap are downscaled even if they're
under the byte budget. The resize loop now checks both byte AND pixel
limits before accepting a candidate.
Closes#37677
Fireworks/Mistral reject HTTP 400 'Extra inputs are not permitted, field:
messages[N].tool_calls[M].extra_content' on any session whose history
contains prior Gemini tool calls. Gemini 3 thinking models attach
extra_content (thought_signature) to tool_calls; it survived to the wire
because the sanitize paths only stripped call_id/response_item_id.
Strip extra_content from the outgoing wire copy in both sanitize paths
(ChatCompletionsTransport.convert_messages + _sanitize_tool_calls_for_strict_api),
but gate it on the target model: keep extra_content for Gemini-family
targets (the thought_signature MUST be replayed or Gemini 400s), strip it
for everyone else — including non-Gemini models that inherit a stale Gemini
signature earlier in a mixed-provider session. Native Gemini is unaffected
(GeminiNativeClient bypasses these paths).
Original stored history is never mutated (only the per-call copy).
Fixes#17986.
The salvaged observer contract gated the API-request hot path on has_hook()
but left the per-tool emit ungated: every tool call ran result-field
derivation + payload dict build + invoke_hook dispatch even with zero
plugins registered.
- _emit_post_tool_call_hook now short-circuits on has_hook("post_tool_call")
and derives status/error fields lazily (after the gate, only when a
listener will consume them). status defaults to None -> derived; explicit
blocked/cancelled callers still pass status through.
- transform_tool_result emit (pre-existing hook) likewise gated on
has_hook(); skips _tool_result_observer_fields when no listener.
- Removed the now-redundant _tool_result_observer_fields pre-computation at
the three ok-path call sites (model_tools, agent_runtime_helpers,
tool_executor) — the helper derives them, so the no-listener path costs
one dict lookup and the call sites shrink.
- Tests: stub has_hook=True where payload correctness is asserted; add a
no-listener regression proving post_tool_call/transform_tool_result emit
is skipped when nothing is registered.
Adds backend-neutral observer hooks for plugins: session, turn, API
request, tool, approval, and subagent lifecycle events with stable
correlation IDs (session_id, task_id, turn_id, api_request_id,
tool_call_id, parent/child subagent ids). Extends VALID_HOOKS with
api_request_error and subagent_start.
Hot path is zero-cost when no plugin subscribes: has_hook()/presence
checks gate all payload construction, request payloads are returned
by reference when no middleware rewrites, and the sanitized response
payload no longer embeds raw response objects.
Bundles the optional NeMo-Relay observability plugin
(plugins/observability/nemo_relay) as an in-repo consumer of the new
hooks, peer to the existing langfuse plugin. Fails open when the
optional nemo-relay package is not installed.
Authored-by: Bryan Bednarski <bbednarski@nvidia.com>
Salvaged from #29722 onto current main.
A stream that drops mid-response after tokens are delivered (peer-closed
connection, stale-stream reconnect) is converted into a synthetic
finish_reason="length" stub. The conversation loop treated that network
stall as a max-output-tokens truncation: when the dropped content was a
tool call it retried exactly once, then hard-failed with "Response
truncated due to output length limit" — even on large-output models that
never hit any cap (e.g. Opus).
- Tool-call truncation now retries up to 3 times (was 1) with a
progressive max_tokens boost, and is stub-aware: a PARTIAL_STREAM_STUB_ID
stall prints "Stream interrupted mid tool-call — retrying (n/3)" instead
of the false "model hit max output tokens", and the give-up message
distinguishes a network drop from a real truncation.
- Length-continuation retries preserve the original request's output cap
as a floor, so a high provider/model default isn't silently downshifted
to 8K/12K on retry.
- Added _requested_output_cap_from_api_kwargs() helper.
Tests: stub-stall mid-tool-call recovery within 3 retries; continuation
preserves a large provider-default output cap.
Fixes#26425. Salvages the substance of #26427 (cap floor) and #9525
(retry bump), adapted to the post-refactor conversation_loop.py which
handles all three api_modes uniformly.
Co-authored-by: LeonSGP43 <cine.dreamer.one@gmail.com>
Co-authored-by: ygd58 <ygd58@users.noreply.github.com>
* feat: better composer etc
* docs: add desktop and dashboard run instructions
* fix(desktop): address security scan findings
* fix(dashboard): resolve @nous-research/ui path under npm workspaces
The sync-assets prebuild step shelled out to 'cp -r
node_modules/@nous-research/ui/dist/fonts ...' with a path relative
to apps/dashboard/. That works only when the dep is installed
locally in the dashboard workspace, but 'npm install' at the repo
root (the documented setup — see apps/desktop/README.md) hoists
shared deps to the root node_modules under npm workspaces. The
relative cp then fails with 'No such file or directory', sync-assets
exits 1, the Vite build aborts, and 'hermes dashboard' surfaces a
generic 'Web UI build failed' message.
Replace the shell one-liner with scripts/sync-assets.cjs, which
walks up from the dashboard directory looking for node_modules/
@nous-research/ui — working in both the hoisted (workspaces) and
co-located (standalone) layouts. Also guards against a missing
dist/fonts or dist/assets with a clearer error pointing at a
rebuild of the UI package rather than silently copying nothing.
* feat(desktop): support connecting to a remote Hermes backend
Add HERMES_DESKTOP_REMOTE_URL and HERMES_DESKTOP_REMOTE_TOKEN env
vars that, when set, short-circuit the local-child spawn in
startHermes() and connect the Electron renderer to an already-
running 'hermes dashboard' server reachable over the network.
Motivating use case: WSL2 users who want to run the Hermes core
(agent loop, tools, filesystem access) inside their WSL
distribution while rendering the Electron GUI on native Windows.
Before this change, the desktop app always spawned a local Python
child on the same host as the renderer, which doesn't cross the
WSL/Windows boundary.
The remote path reuses waitForHermes() as a liveness probe
(/api/status is in the backend's public endpoint allowlist), so
the connection is only returned once the backend is actually
ready. WebSocket URL derivation picks ws:// or wss:// based on
the input scheme. URL validation rejects non-http(s) schemes and
requires both env vars together to avoid a half-configured
connection that would silently fall through to the spawn path.
No behaviour change when the env vars are unset — the default
local-spawn flow is untouched.
Typical usage:
# in WSL2
hermes dashboard --tui --no-open --host 0.0.0.0 --port 9119 --insecure
# on Windows
set HERMES_DESKTOP_REMOTE_URL=http://localhost:9119
set HERMES_DESKTOP_REMOTE_TOKEN=<session token>
set HERMES_DESKTOP_IGNORE_EXISTING=1
(launch Hermes desktop)
* ci(desktop): automate desktop releases
Add GitHub Actions release channels for signed desktop installers and document the stable/nightly download paths.
* feat: file tabs
* refactor(desktop): tighten right-rail tab close API
Promote closeRightRailTab/closeActiveRightRailTab as the single
public entry point. Drops the activeTabRef + handleCloseDocument
indirection in ChatPreviewRail, the unused $rightRailHasContent
atom, and the legacy dismissFilePreviewTarget alias. -70 LOC.
* feat(desktop): polish composer pill toward reference look
Solid foreground-on-background send/voice-conversation circle (black-on-white
in light, white-on-black in dark) anchors the right edge as the primary CTA
instead of the orange theme primary. Bumps the primary control to 2.125rem so
it visually outranks the ghost mic/plus controls. Opens up the surface padding
(0.625rem x / 0.5rem y) so the input row breathes around its controls, and
nudges the corner radius from 20 to 24px for a slightly pill-ier silhouette.
LiquidGlass distortion is preserved.
* feat(desktop): add startup and onboarding flow
Add phase-based desktop boot progress, fresh-install sandbox testing, and first-run provider credential onboarding so packaged installs can start cleanly without manual settings detours.
* fix(desktop): gate prompts on provider setup
Show the desktop provider onboarding flow before prompt submission when no inference provider is configured, preventing fresh installs from falling through to backend credential errors.
* fix(desktop): surface provider onboarding from session warnings
Propagate credential warnings through session runtime info and open desktop onboarding whenever a session reports no usable provider, so unconfigured installs cannot fall through to prompt errors.
* fix(desktop): route gateway provider errors to onboarding
The "No inference provider configured" auth error reaches the renderer through gateway error events, not the prompt.submit promise; the previous patch only caught the latter, so the error toast still surfaced and onboarding never opened.
Also strip credential-shaped env vars from the test:desktop:fresh sandbox so the packaged backend can't see provider keys leaking from the launching shell.
* fix(desktop): use strict runtime check to drive onboarding
setup.status returned True whenever any provider auth state was discoverable, including indirect fallbacks like a gh-CLI Copilot token. That made desktop think the user was set up while the agent's actual resolve_runtime_provider call still raised AuthError, leaving the user with a useless toast and no onboarding.
Add a setup.runtime_check gateway method that runs the same resolver the agent uses on session creation, and switch the desktop onboarding overlay and prompt precheck to use it.
* feat(desktop): OAuth-first onboarding using existing dashboard provider API
Replace the engineer-flavored API key form with a Sign-in-first onboarding overlay that uses the dashboard's existing /api/providers/oauth catalog and PKCE/device-code endpoints (Anthropic, Nous, OpenAI Codex, etc.). API key entry is now a fallback tab with friendly provider names instead of env var prefixes, and the loud raw resolver error is gone in favor of a one-line welcome message.
* fix(desktop): polish onboarding provider list
Reorder OAuth providers so Nous Portal is first, give the segmented Sign in / API key control equal column widths, and replace the engineer-flavored backend names like "Anthropic (Claude API)" / "MiniMax (OAuth)" with friendlier in-app titles. External-CLI providers now show a softer subtitle and an external-link icon instead of a chevron.
* refactor(desktop): split onboarding overlay into store + view
Move the OAuth state machine, runtime check, copy-to-clipboard, and api-key save into store/onboarding.ts (matching the boot.ts pattern), leaving the overlay as a presentation layer that subscribes via useStore. Tabs are now table-driven, child panels read flow from the store instead of prop-drilling, and the polling/PKCE/error/success branches share a small Status atom.
* fix(desktop): external CLI providers + center mode tabs
External-CLI providers (Claude Code, Qwen Code) now open an in-overlay panel with the CLI command, copy button, and an "I've signed in" recheck instead of firing an invisible toast. Center the Sign in / API key tab control so it sits under the heading instead of hugging the left edge.
* fix(desktop): drop onboarding tabs for an inline link, group device-code waiting state
Replace the Sign in / API key tab pair with an "I have an API key" footer link under the OAuth provider list, with a "Back to sign in" affordance inside the API key form. Group the device-code "Waiting for you to authorize..." status next to the Cancel button so the alignment matches the action.
* refactor(desktop): tighten onboarding store + overlay
Drop the dead isOnboardingBusy/BUSY set, factor the catch-fallback dance into safeReq, and share a single reloadAndConnect helper between PKCE submit, device-code success, external recheck, and api-key save.
In the overlay, extract Step / CodeBlock / FlowFooter / CancelBtn / DocsLink atoms so the four sign-in panels share the same chrome instead of repeating it inline. Net effect: fewer literal divs, one place to touch the spacing, and the code-block + footer rows are reusable across future flows.
* fix(desktop): mount onboarding from frame 1 to kill the FOUT
Default onboarding.configured to null (unknown until the runtime check resolves) and have the onboarding overlay render whenever it's not yet confirmed true. The boot overlay now yields to it, so the very first paint is the Welcome card with a "While we get you set up..." progress strip instead of a flash of the chat shell between boot dismiss and onboarding mount.
The picker swaps in cleanly once the gateway opens and the runtime check confirms the user is not configured. Already-configured users see the same prep card briefly while their existing runtime warms up, then the overlay dismisses without touching the chat shell.
* fix(desktop): top-align empty sessions placeholder
The "Start a chat to build your history." empty state used a min-h-35 grid place-items-center container, which floated the text in a tall dead zone. Render it as a flat paragraph that sits right under the section header like the empty pinned state does.
* refactor(desktop): drop dead boot overlay
Onboarding overlay subsumes the boot card now that it mounts from frame 1 and renders boot progress inline. The standalone DesktopBootOverlay is unreachable in every flow (yields whenever onboarding has not confirmed configured, dismisses once it has).
* fix(desktop): hide pinned/recents sections until first session
A fresh sidebar showed the Pinned and Recent chats headers with floating empty-state copy underneath. Drop both sections (and the now-orphan SidebarEmptySessionState) when there are no sessions yet — they reappear after the first chat. Skeletons during initial load are unchanged.
* feat(gui): route embedded TUI through dashboard gateway (#21979)
Inject HERMES_TUI_GATEWAY_URL into dashboard PTY sessions so embedded ui-tui instances attach to the in-process websocket gateway, with coverage for the new env wiring.
* Add desktop remote gateway settings
Make the desktop gateway connection configurable from settings so local remains the default while remote backends can be saved, tested, and applied without environment variables.
* feat(gui): first-class Messaging page + gateway menu redesign
- Add Messaging page to the desktop app with per-platform setup,
status, and inline guidance. Catalog derives from gateway.config
Platform enum + plugin registry, so every messaging adapter the CLI
supports (Telegram, Discord, Slack, Mattermost, Matrix, WhatsApp,
Signal, BlueBubbles, Home Assistant, Email, SMS, DingTalk, Feishu,
WeCom, Weixin, QQ, Yuanbao, API server, Webhooks, plugins) shows up
without per-platform code.
- New REST endpoints: GET /api/messaging/platforms, PUT and POST
/test on the same path. Secrets go through the existing .env
pipeline; enable/disable writes config.yaml.
- Replace gateway statusbar dropdown with a richer panel: status row,
icon-only restart + system-panel actions, recent activity (with
timestamps trimmed in display, full text on hover), platform list.
- Auto-poll the messaging page every 6s (paused when hidden) so
status updates without a manual check.
- Drop Settings / Command Center from the sidebar nav (still
reachable via shortcuts and the titlebar cog).
- Flatten top corners on Messaging/Skills/Artifacts/Chat panes.
- Share new StatusDot component across messaging + gateway menu.
- Fix gateway/config.py so an explicit platforms.<name>.enabled=false
in config.yaml is honored when env tokens are present.
- pb-9 on the chat content area for breathing room above the composer.
* Potential fix for pull request finding 'CodeQL / Clear-text logging of sensitive information'
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
* pin electron version
* hide application menu on non-mac systems
* interpret compactPreview for non-string vlaues as JSON or an empty string
* fix(desktop): keep composer contenteditable mounted across stacked toggle
The composer rendered {input} inside two different parent fragments
depending on `stacked`. When auto-expand flipped `stacked` (e.g. the
moment typed text wrapped past two lines), React reconciled the two
branches as different positions and unmounted/remounted the
contenteditable. The fresh mount started empty, so any in-flight
characters — most reliably reproduced by holding a key — were lost.
Replace the conditional with a single CSS Grid whose template-areas
swap on `stacked`. The three children (menu, input, controls) keep
stable identities across the toggle; only their grid placement
changes, which the browser handles without React tearing down the
editor.
* refactor(desktop): align install layout with install.ps1 / install.sh
Make the desktop app's runtime layout match what scripts/install.ps1 and
scripts/install.sh produce, so a desktop-only user and a CLI-only user end
up with the same files in the same places and can share one install.
Layout
- ACTIVE_HERMES_ROOT = HERMES_HOME/hermes-agent (was: process.resourcesPath/hermes-agent, read-only)
- VENV_ROOT = HERMES_HOME/hermes-agent/venv (was: userData/hermes-runtime)
- desktop.log = HERMES_HOME/logs/desktop.log (was: userData/desktop.log)
- HERMES_HOME default: %LOCALAPPDATA%\hermes on Windows, ~/.hermes elsewhere
The packaged .app/.exe still ships a read-only payload at
process.resourcesPath/hermes-agent (FACTORY_HERMES_ROOT). On first launch
or after an installer-driven upgrade we sync factory -> active, then
provision the venv and run pip install -e . against the active root.
Key behaviors
- Pin HERMES_HOME in the spawned Python's env so get_hermes_home() resolves
to the same path resolveHermesHome() picked. Without this, Python falls
back to ~/.hermes on every platform - fine on mac/linux, a split-state
bug on Windows where our default is %LOCALAPPDATA%\hermes.
- Detect developer installs by .git presence at ACTIVE; never overwrite
a user's checkout via factory sync.
- Marker at ACTIVE/.hermes-desktop-runtime.json (schema v4) tracks
pyproject hash + factory version + runtime schema version. depsFresh
fast-paths when nothing changed.
- Dev (npm run dev) prefers SOURCE_REPO_ROOT over ACTIVE so devs run
their local edits, not whatever's under HERMES_HOME.
- Better error messages distinguish "no payload" from "no Python".
- Preserve a legacy ~/.hermes on Windows when no %LOCALAPPDATA%\hermes
exists, so users with prior pip/manual installs aren't orphaned.
pyproject.toml
- Promote fastapi, uvicorn[standard], ptyprocess (non-Windows), and
pywinpty (Windows) to main dependencies. The dashboard backend
(hermes dashboard) needs them at runtime; the previous lazy-import
fallback was a footgun for fresh installs.
- Empty the [pty] optional-extra; kept as a no-op back-compat alias for
any existing pip install hermes-agent[pty] invocations.
Drops the hardcoded BUNDLED_RUNTIME_REQUIREMENTS list in main.cjs - the
desktop now installs whatever pyproject.toml says, single source of truth.
Files
- apps/desktop/electron/main.cjs: runtime layout, HERMES_HOME pin,
factory->active sync, marker v4
- apps/desktop/scripts/test-desktop.mjs: track new venv location
- apps/desktop/README.md: new Setup, Runtime Bootstrap, and
Debugging sections
- pyproject.toml: fastapi/uvicorn/pty backends in main
dependencies; [pty] extra emptied
Tested locally on Windows: npm run dev boots cleanly, sessions land at
the new location, type-check + lint + test:desktop:platforms all pass.
Verified end-to-end on a fresh Win11 VM via dist:win installer.
Known gaps (filed as follow-ups, not in this PR):
- Skills not seeded on packaged installs (sync_skills only runs in
cmd_chat, not cmd_dashboard). Need to move to shared pre-dispatch.
- Git Bash not bundled or detected; agent's terminal tool errors out
with a useful message but desktop bootstrapper should pre-flight it.
- install.ps1 / install.sh should be decomposed into composable phase
libraries so the desktop bootstrapper can reuse them as a single
source of truth across all install surfaces.
* feat(desktop): theme polish, prose chat typography, composer chrome
- DS tokens/midground, Backdrop, scoped scrollbars, typography plugin + prose
- Composer liquid/radius utilities, thread font parity, tool/thinking cues
- File tree label scale, preview flex, thread retry loading + streaming tests
* feat(desktop): NSIS prereq detection page + auto-install via winget
The packaged Windows installer now detects Python 3.11+ and Git for Windows
at install time and offers to install missing prereqs via winget. Mirrors
the prereq logic scripts/install.ps1 already runs for CLI installs, so
desktop installer users get the same out-of-the-box experience as
install.ps1 users.
Why
- Hermes' terminal tool calls bash.exe directly (tools/environments/
local.py); on Windows that's Git Bash from Git for Windows. Without it,
the agent fails on the first terminal() call.
- Hermes' Python runtime needs 3.11+. Without it, the desktop bootstrapper
errors out at venv creation.
- Both gaps surfaced on a fresh Windows 11 VM smoke test: VM had Python
pre-installed but no Git, so the agent's first terminal call failed
with "Git Bash isn't installed."
- install.ps1 has had Install-Git + Install-Uv functions for ages. The
desktop installer was the asymmetric outlier.
How — NSIS prereq page
- New file: apps/desktop/installer/prereq-check.nsh (plugged into
electron-builder via build.nsis.include)
- Real Wizard page using nsDialogs, inserted via customPageAfterChangeDir
hook (between the Directory page and InstFiles).
- Group boxes for Python and Git, each showing detection status.
- Pre-checked install checkboxes when winget is available.
- Auto-skips silently if both prereqs are already installed.
- Falls back to manual download URLs when winget itself is missing.
- Detection:
- Python: probes `py -3.11`/`-3.12`/`-3.13`/`-3.14` via the Python
launcher. Microsoft Store "Python stub" (no py.exe) is correctly
classified as not-installed.
- Git: `where git`.
- winget: `where winget` (Win10 1809+ / Win11 with App Installer).
- Install execution (in customInstall macro):
- Python: nsExec::ExecToLog with `--scope user --silent`. Per-user
install, no UAC prompt, output streams to install log.
- Git: ExecShellWait via Windows ShellExecute. Critical because Git
always installs per-machine and triggers UAC; ShellExecute preserves
the foreground focus chain across non-elevated → elevated process
spawns, so UAC actually comes to the foreground. nsExec::ExecToLog
breaks the chain because winget runs hidden.
- Both pass `--disable-interactivity --accept-package-agreements
--accept-source-agreements` to suppress winget's own dialogs.
- Verification: probes Git's standard install locations via FileExists
rather than `where git`. NSIS's process inherits PATH at startup, so
a freshly-installed Git won't be visible to `where` until restart.
- Silent installs (/S) skip the prompts; managed deploys handle prereqs
out-of-band via Group Policy / Intune.
How — Electron-side safety net
- New findGitBash() in main.cjs, parallel to findSystemPython(). Probes
the same locations as tools/environments/local.py:_find_bash() so a
positive result here means the agent's terminal tool will work.
- ensureRuntime now throws a clear, actionable error on Windows when Git
Bash isn't found, matching the existing "Python 3.11+ is required"
error path.
- Catches users the NSIS page doesn't: .msi installer users (NSIS prereq
page doesn't run for MSI), `npm run dev` users, manual installers,
anyone who unchecked the install boxes on the NSIS prereq page.
- All gated on `IS_WINDOWS`; macOS / Linux unaffected.
NSIS build issue (resolved)
- electron-builder defaults to `-WX` (warnings as errors). NSIS optimizer
emits "warning 6010: function not referenced" for our page functions
because Page custom directives don't count as references in its
static-analysis pass. The functions ARE called at runtime when NSIS
invokes the page; the optimizer just can't see it statically.
- Set `build.nsis.warningsAsErrors=false` in package.json so this
spurious warning doesn't fail the build. (Documented option from
electron-builder's nsisOptions.)
Out of scope (filed for future work)
- MSI prereq detection: Windows Installer custom actions are a different
mechanism. Enterprise deploys typically handle prereqs via GP/Intune.
- Bundle PortableGit + python-build-standalone in extraResources for
zero-network installs. ~80MB increase.
- Mac / Linux GUI prereq flows (different installer formats; Xcode CLT
covers most macOS prereqs already; Linux is per-distro hard).
Files
- apps/desktop/installer/prereq-check.nsh (new, ~290 lines NSIS)
- apps/desktop/package.json (build.nsis.include +
warningsAsErrors)
- apps/desktop/electron/main.cjs (findGitBash + preflight)
- apps/desktop/README.md (Runtime prerequisites
section)
Cross-platform impact
- macOS / Linux builds (dist:mac, dist:mac:dmg, dist:mac:zip): nsis
config is ignored entirely; .nsh is dormant.
- npm run dev: .nsh dormant; main.cjs preflight gated on IS_WINDOWS.
- scripts/install.ps1, scripts/install.sh: no reference to any new
files; CLI install paths untouched.
- Hermes CLI / dashboard / gateway: no reference; runtime untouched.
- All checks: node --check on main.cjs and test-desktop.mjs pass;
npm run test:desktop:platforms 4/4 passing; node --test green.
Tested
- npm run dist:win produces signed .exe and .msi without errors.
- Fresh Win11 VM (Python pre-installed, no Git): prereq page renders,
Python check shows detected, Git checkbox pre-checked. Click Next →
Git installs via winget with UAC prompt in foreground.
- After install completes, Hermes launches and the agent's terminal
tool can run bash commands. Verified Git Bash is detected at
`C:\Program Files\Git\bin\bash.exe` by ensureRuntime's preflight.
* feat: theme changes, composer tweaks, in app update ux, finesse
* fix(cli): seed bundled skills on dashboard + gateway entrypoints
`sync_skills(quiet=True)` was only being called from inside `cmd_chat`,
which meant `hermes dashboard` (the desktop GUI's backend) and `hermes
gateway` (Telegram/Discord/Slack/etc daemons) never seeded the bundled
skill library into ~/.hermes/skills/.
This surfaced as "No skills found" in the desktop GUI's skills panel on
fresh installs, despite the agent having access to the full bundled
library when invoked via `hermes chat`. scripts/install.ps1 worked
around it by running skills_sync.py as part of Copy-ConfigTemplates,
but that's not part of the desktop installer's bootstrap chain.
Fix
- Extract the skills-sync block from cmd_chat into a module-level
`_sync_bundled_skills_quietly()` helper.
- Call the helper from cmd_chat (preserving existing behavior),
cmd_dashboard (after the --status/--stop early-return paths and
fastapi import check, so we don't run skills_sync on management
commands or when deps aren't installed), and cmd_gateway.
Why these three entrypoints
- cmd_chat: the user's primary CLI entrypoint
- cmd_dashboard: the desktop GUI's backend; this is what `hermes
dashboard --tui` invokes when the desktop bootstrapper spawns Hermes
- cmd_gateway: long-running daemons where the user expects the agent
to have full skill access
Other entrypoints (cmd_config, cmd_doctor, cmd_login, cmd_status,
etc.) are management commands that don't need skill discovery and were
never running skills_sync in the first place — leaving them alone.
Idempotence
- tools/skills_sync.py is manifest-based: skipped skills cost
milliseconds. Calling it from multiple entrypoints adds no real
cost, and users running `hermes chat` then `hermes dashboard` get
two fast no-ops on the second call.
Failure handling
- Helper wraps skills_sync in try/except. Skills are an enhancement,
not a hard dependency — Hermes runs fine with an empty skills/ dir.
Files
- hermes_cli/main.py:
+ new helper `_sync_bundled_skills_quietly()` at module level
+ cmd_chat: replace inline block with helper call
+ cmd_dashboard: add helper call after fastapi import succeeds
+ cmd_gateway: add helper call before delegating to gateway_command
* feat(desktop): hoisted todo widget, JSON tool summaries, history grouping & timer fixes
- Hoist todo to first-class widget (shadcn checkboxes, brand colors, no
tool-accordion). Header derives label from active task; non-active rows fade.
- Replace raw JSON dumps with structured key/value summaries via
formatToolResultSummary; nested error extraction for clearer failures.
- Fix loaded-session grouping: stitch interleaved assistant/tool iterations
into one bubble instead of orphaned synthetic messages.
- Stable tool/thinking timers via keyed registry so unmount/scroll doesn't
reset elapsed counts; gate "running" on real live thread state.
- Reorganize chat-only assistant-ui components under components/chat/.
* fix(desktop): address CodeQL alerts on PR #20059
- settings/helpers.ts: harden setNested against prototype pollution.
POLLUTING_PATH_PARTS check is now applied at every assignment site
(loop + leaf) and uses Object.defineProperty so CodeQL can see the
guard inline rather than via a helper function call.
- lib/markdown-preprocess.ts: rebuild the dangling-fence close regex
from a fence-char + length instead of marker.replace(...). The marker
is captured by `(`{3,}|~{3,})` so it can only be backticks or tildes,
but CodeQL was tracing tainted input text into the RegExp source and
flagging hostname dots from input as part of the pattern (false
positive js/incomplete-hostname-regexp on the test fixture URLs).
Reconstructing from a literal char breaks the dataflow.
- scripts/notarize-artifact.cjs: drop args from the run() rejection
message. Args carry --key-id / --issuer / key file path; the existing
outer catch already squashes errors to a generic line, but CodeQL was
flagging the args.join(' ') as clear-text logging of APPLE_API_KEY_ID.
Composer DOM-text-as-HTML alerts (composer/index.tsx:379, :547) are
already addressed in 4dd9732a9 — innerHTML assignment was replaced with
renderComposerContents which builds DOM via replaceChildren / append
text nodes (no HTML interpretation).
* fix(desktop): inline prototype-pollution guard so CodeQL sees it
CodeQL's dataflow doesn't follow the helper-function guard inside
`safeSet`, so it kept flagging Object.defineProperty as prototype-
polluting. Inline the literal `__proto__`/`constructor`/`prototype`
check at the assignment site to break the dataflow.
Behavior unchanged — same set of disallowed keys, same throw.
* feat(ui-tui): resolve links to readable page titles
Mirror desktop pretty-link behavior in the TUI by resolving HTTP links to page titles with shared caching and safe fetch filters, plus slug-based fallbacks so chat links stay readable even when title fetch fails.
* fix(desktop): drop RegExp from dangling-fence close detection
Previous attempt tried to break the dataflow by reconstructing the
close-fence regex from a literal char + marker.length, but CodeQL still
traced marker.length back to input and kept flagging the test-fixture
URLs as hostname-regex sources (js/incomplete-hostname-regexp).
Replace `new RegExp(...)` + `closeRe.test(body)` with a string-only
hasCloseFenceLine() helper that splits on '\n' and uses ===. No regex
on this path now, so input data can no longer reach a RegExp source.
Behavior preserved: matches lines that are (whitespace + marker +
whitespace), which is what the original `\n[ \t]*${marker}[ \t]*(?=\n|$)`
matched. All 12 markdown-text tests still pass.
* fix(process-registry): suppress windows-footgun false positive on guarded killpg
Keep the existing POSIX-only process-group teardown path, but make the
signal selection explicit via getattr and add an inline windows-footgun
suppression marker on the guarded os.killpg line so the Windows footgun
check no longer blocks CI on this intentionally platform-gated code.
* feat(desktop): reconcile live tool events, polish thread chrome, harden boot
- chat-messages: match tool rows by overlapping query/context/preview values
so preview-first `tool.progress` rows reliably adopt later stable-id
`tool.start` payloads instead of spawning ghost rows or mis-merging
parallel same-name calls; preserve prior args/result across phases.
- tui_gateway: emit full args + parsed result on `tool.start` / `tool.complete`,
drop redundant `tool.started` re-emit from `tool.progress`.
- electron/main: prefer SOURCE_REPO_ROOT before PATH `hermes` in dev so
local backend edits actually run; split hardening helpers into
`electron/hardening.cjs` with tests.
- thread/tool UI: one-shot enter animation keyed by stable ids, braille
spinner for running rows, Cursor-like disclosure rows, drill-down +
duration/count formatting via new tool-fallback-model.
- composer: extract `text-utils`, drop liquid-glass overrides.
- right-rail: split preview-pane into preview-console / preview-file.
- runtime: incremental external-store runtime + runtime-readiness gate;
onboarding store + tests; route-resume hook test.
- regression tests for live tool reconciliation (parallel tools, id-less
progress, preview-first rows, structured args/results).
* feat(desktop): add ripgrep to NSIS prereq page + polish layout
Add ripgrep as a third (recommended) prereq alongside Python and Git in
the NSIS prereq detection page, and clean up the page layout based on
on-VM testing.
Why ripgrep
- Hermes' search_files tool calls `rg` directly for content + filename
search (tools/file_operations.py:1382). Falls back to grep/find from
Git Bash when missing — works but slower and noisier (no .gitignore
awareness).
- ~5MB winget install via `BurntSushi.ripgrep.MSVC --scope user` — no
UAC prompt, parallel to how Python installs.
- scripts/install.ps1 already installs ripgrep as part of
Install-SystemPackages; this brings the desktop installer to parity.
Why "recommended" not "required"
- Python and Git are hard requirements: without them the agent runtime
or terminal tool refuses to start. The bootstrapper preflight throws.
- ripgrep is a performance enhancement: missing it just means slower
searches. Page wording reflects this; failure to install is logged
but doesn't show a MessageBox or block.
Layout polish (response to on-VM screenshot review)
- Wizard header now correctly reads "System Requirements" instead of
the leftover "Choose Install Location" from the previous page. Set
via `GetDlgItem $HWNDPARENT 1037/1038` + WM_SETTEXT — the standard
NSIS pattern for overriding the page header on a custom Page.
- Removed redundant in-body title + verbose intro paragraph; the
wizard header IS the title now. Body has one short intro line.
- Group boxes tightened to 26u with content positioned just below the
groupbox title (not top-anchored status + bottom-anchored checkbox
with empty space in the middle). All three panels + footer fit
comfortably in 126u, well under the 140u page limit.
- Checkbox labels simplified: dropped "(per-user, no admin prompt)"
and "(administrator approval required)" suffixes. The footer note
still calls out UAC for Git when relevant.
- Footer text trimmed to fit cleanly without clipping.
Install order (in customInstall macro)
- Python → ripgrep → Git
- Python and ripgrep are silent and run first; Git's UAC prompt comes
last so the user's approval interaction isn't interrupted by silent
activity afterwards.
Skip behavior unchanged
- All three detected → page auto-skips via Abort
- Silent install (/S) → customInstall winget block skips
- User unchecks all → page advances without running winget
Files
- apps/desktop/installer/prereq-check.nsh: ripgrep detection block,
ripgrep page panel + checkbox, ripgrep customInstall block,
GetDlgItem header override, layout reflow
- apps/desktop/README.md: Runtime prerequisites section updated to
list ripgrep as recommended, with manual winget command
* feat(desktop): add model-confirmation step to onboarding
After OAuth/API-key login completes, onboarding now shows a confirmation
card with the curated default model and a Change button before dropping
the user into chat. Closes the gap where the desktop's `model.default`
was empty after first launch and the agent had to fall back to whatever
heuristic happened to fire — leaving users wondering "why am I getting
sonnet-4 when I logged into Nous Portal?"
Why
- Desktop onboarding only persisted credentials, never `model.default`.
The CLI's `hermes model` command pairs provider + model selection,
but the desktop's onboarding skipped the model step entirely.
- Result: users saw whichever model the agent's auto-fallback picked,
unpredictably and undocumented.
- For the BUILD demo we want users to land on the model they expect
for their provider, with a clear "this is what you're getting" UI
and a one-click path to change it before chatting.
How
- New `confirming_model` flow status carries the just-authenticated
provider slug, current default model, label, and a saving flag.
- `completeWithModelConfirm()` runs after credentials succeed: reloads
env, verifies runtime, fetches /api/model/options to find the curated
first-model for the provider, persists it via /api/model/set, then
transitions into `confirming_model`.
- If anything fails (no providers returned, network error), falls
through to the previous behaviour — onboarding completes without
the confirm step. Polish, not a hard requirement.
- All four credential paths (device_code OAuth, PKCE OAuth, external
CLI flow, API key) now use completeWithModelConfirm instead of
reloadAndConnect.
UI
- `ConfirmingModelPanel` shows: green "<provider> connected" banner,
card with "Default model: <name>" + Change button, and a "Start
chatting" CTA that finalises onboarding.
- Reuses the existing `ModelPickerDialog` (the same picker available
from the chat shell) for the change-model UX. Search, filtering,
multi-provider listing — all already built.
- Stacking: ModelPickerDialog defaults to z-130, which renders UNDER
the onboarding overlay (z-1300) and breaks pointer events. Added
optional `contentClassName` prop to ModelPickerDialog so callers
can override; onboarding passes `z-[1310]`.
Provider-slug matching
- For OAuth flows: pass `provider.id` directly as the preferred slug.
- For API-key flows: `OPENROUTER_API_KEY` → "openrouter" via env-key
prefix strip. Also includes the user-visible label as a fallback
candidate.
- fetchProviderDefaultModel falls back to the first authenticated
provider in the response if no preferred slug matches — so even a
miss still surfaces a reasonable default.
Files
- apps/desktop/src/store/onboarding.ts:
+ new `confirming_model` flow variant
+ fetchProviderDefaultModel + completeWithModelConfirm helpers
+ setOnboardingModel (optimistic update + revert on failure)
+ confirmOnboardingModel (finalises onboarding from the card)
- reloadAndConnect (replaced; the four call sites now go through
completeWithModelConfirm)
- apps/desktop/src/components/desktop-onboarding-overlay.tsx:
+ ConfirmingModelPanel component
+ new branch in FlowPanel for status `confirming_model`
+ ModelPickerDialog usage with z-[1310] content class
- apps/desktop/src/components/model-picker.tsx:
+ optional `contentClassName` prop on ModelPickerDialog so the
dialog can be stacked on top of other fixed overlays
Tested
- `npm run type-check` passes
- `npx eslint` clean on touched files
- Live test in `npm run dev`: cleared onboarding cache, walked
through Nous device-code flow, saw confirm card with curated
default, clicked Change → ModelPickerDialog rendered above the
onboarding overlay with working pointer events, picked a different
model, "Start chatting" persisted to ~/.hermes/config.yaml.
* fix(desktop): suppress generic provider warning in onboarding
Hide the red setup notice when the message is the generic missing-provider guidance, since onboarding already presents provider auth actions. Centralize provider-setup matching across desktop hooks and add coverage for the matcher.
* fix(desktop): add 2u clearance below prereq checkboxes
Group box bottom border was clipping the checkboxes by 1-2px.
Bumped each box height 26u→30u; checkboxes now sit 2u above the bottom border.
* fix(nix): refresh dashboard lockfile hash
Update the web npm deps hash in nix/web.nix to match the committed apps/dashboard/package-lock.json so bb/gui passes the nix lockfile check.
* fix(desktop): install TUI deps in release workflow
Ensure desktop release builds install the standalone ui-tui package before bundling the TUI payload.
* fix(desktop): run release builder from app package
Invoke the desktop builder through the package script so electron-builder uses apps/desktop/package.json.
* fix(desktop): expand release artifact names safely
Build desktop artifact names from workflow version/channel while preserving electron-builder platform macros.
* fix(desktop): use package artifact naming in release workflow
Let electron-builder's desktop package config provide platform-specific artifact extensions while the workflow injects the release version/channel metadata.
* fix(nix): fetch dashboard npm deps from package root
Point the dashboard npm dependency fetch at apps/dashboard so Nix can find the package lockfile after the dashboard move.
* fix(nix): build dashboard from package directory
Set the web package source root to apps/dashboard so npm patch/build phases run beside the dashboard lockfile while keeping apps/shared available as a sibling.
* feat(desktop): render LaTeX math via KaTeX after streaming completes
Add @streamdown/math plugin to the chat markdown renderer.
Inline ($x^2$) and block ($$...$$) math both supported with
singleDollarTextMath enabled. Plugin is gated to non-streaming state
to match the existing pattern for syntax highlighting — math renders
when the message completes, avoiding KaTeX re-render churn during
streaming. KaTeX CSS is imported in styles.css; ~30KB CSS + ~430KB
JS added to the bundle. Smoothness improvements during streaming
deferred to a follow-up.
* perf(desktop): memoize KaTeX renders so math streams without re-rendering
Wrap rehype-katex with a per-equation LRU cache (keyed by
displayMode + source text) and re-enable math during streaming.
Stock @streamdown/math runs rehype-katex on every markdown commit,
so each new token re-katexes every equation in the message. For
math-heavy responses (an equation derived step-by-step) that's
hundreds of ms of wasted work per token and the streaming UI
chokes. With memoization, each equation pays katex.renderToString
exactly once; subsequent tokens re-walk the tree but hit cache for
unchanged equations.
The wrapper mirrors rehype-katex's semantics exactly: same class
detection (language-math, math-inline, math-display), same
<pre>-walk-up for fenced math blocks, same parent.children.splice
replacement, same SKIP traversal, same strict-then-lenient render
strategy with VFile message reporting.
Cached children are structuredCloned on each splice so downstream
rehype plugins or toJsxRuntime can't mutate the cache.
* fix(desktop): declare katex-memo deps directly + drop per-app lockfile
katex-memo.ts (added in 112cad59b) imports hast-util-from-html-isomorphic,
hast-util-to-text, remark-math, katex, and unist-util-visit-parents but
those were never added to apps/desktop/package.json. They were silently
resolving via @streamdown/math at the workspace root, which broke the
moment `npm i --prefix apps/desktop` ran with the per-workspace lockfile
because that install only consults apps/desktop/package.json. Add them
as direct deps, plus unified/vfile/@types/hast for the type imports.
Also delete apps/desktop/package-lock.json — root package.json declares
workspaces: ["apps/*"], so npm manages all lockfile state at the root.
The stale per-app lockfile is what made `npm i --prefix apps/desktop`
diverge from the workspace install in the first place and left an empty
apps/desktop/node_modules/@assistant-ui/ stub that Vite's dep optimizer
then tried (and failed) to open at @assistant-ui/core/dist/internal.js.
* feat(desktop): disable Backdrop noise overlay by default
The noise overlay defaulted to on, which adds a busy speckle layer over
the whole window for every new user. Flip the Leva default to off; the
toggle stays in Backdrop / Noise for anyone who wants it back.
* fix(desktop): polish LaTeX rendering — currency, code blocks, brackets
Five distinct bugs surfaced from a math-heavy stress test:
1. Adjacent code fences glued together. scrubBacktickNoise's
second-pass regex /``\s*``/g matched the LAST 2 backticks of
one fence + whitespace + FIRST 2 backticks of the next, collapsing
two blocks into one. Fixed with lookbehind/lookahead so we only
match exactly 2 backticks not part of a longer run.
2. Whitespace eaten between fences and following content.
stripPreviewTargets internally calls .trim() which strips leading/
trailing whitespace from each split-segment. For segments between
two fences this collapsed \n\n to '', gluing fence close to next
block. Fixed by capturing leading/trailing whitespace at the call
site and restoring it after the transform.
3. Currency dollar signs eaten as math. With singleDollarTextMath:true
remark-math greedy-matched any pair of $, so '$5 ... $10' became
one inline math span. Added escapeCurrencyDollars to escape $<digit>
patterns to \$<digit> in prose segments (not in code). Trade-off:
math expressions starting with a digit (rare — '$5x = 10$') get
escaped too. Mirrors the convention in ChatGPT/Claude's UIs.
4. \(...\) and \[...\] LaTeX brackets unsupported. Models often
emit these instead of $...$ / $$...$$. Added
rewriteLatexBracketDelimiters preprocessor pass.
5. ```latex / ```tex blocks were being routed to KaTeX via a
rewrite to ```math. Aligns with GitHub markdown convention:
```math = render as math; ```latex / ```tex = LaTeX/TeX
source code (syntax highlighted, not rendered). Conflating them
broke teaching/showing-source use cases. MATH_FENCE_LANGUAGES
pruned to {'math'} only.
Also flipped parseIncompleteMarkdown to true (was !isStreaming) so
the math parser can't see $ inside streaming-but-not-yet-closed code
fences. Shiki was already deferred via defer={isStreaming} so this
doesn't introduce new tokenization cost.
Test: 18/18 existing tests still pass; one test updated to expect
escaped \$ in currency-prose-with-URL case.
* fix(desktop): detect Python via registry/filesystem; pin to 3.11–3.13
Two related fixes for Python detection on Windows:
1. py.exe (Python launcher) is missing from per-user installs that
didn't check the launcher option, so 'py -3.X --version' alone
misses real Python installs. User-reported case: clean Win11 +
official Python.org 3.14 install -> 'where py' returned nothing,
our installer offered to install Python again. Both NSIS prereq
page and main.cjs now probe in this order:
1. py.exe launcher (when present)
2. PEP 514 registry: HKLM/HKCU\SOFTWARE\Python\PythonCore\<v>\InstallPath
3. Filesystem: %ProgramFiles%\Python<v>, %LocalAppData%\Programs\Python\Python<v>
Crucially, we never fall back to running 'python.exe' from PATH
on Windows — the WindowsApps stub at %LOCALAPPDATA%\Microsoft\
WindowsApps\python.exe is a redirector that opens the Microsoft
Store window if no Store Python is installed. Triggering that
during boot would be terrible UX. Registry/filesystem probes
never execute the binary.
2. Drop 3.14 from the supported version set. Several Hermes deps
(notably pywinpty, which carries Rust crates like
windows_x86_64_msvc) don't yet publish 3.14 wheels. With wheels
missing, 'pip install -e .' falls back to building from sdist,
which needs a Rust toolchain — users see 'could not compile
windows_x86_64_msvc build script' on first run. install.ps1
sidesteps this by pinning to 3.11 via uv; the desktop installer
doesn't yet have the same uv-managed-Python pathway, so for now
we accept 3.11/3.12/3.13 and tell winget to install 3.11 if
none of those are present. Revisit when the wheel ecosystem
catches up to 3.14 (~early 2026).
* feat(desktop): Cron, Profiles, usage analytics, and titlebar fixes
- Add Cron and Profiles sidebar routes with full CRUD-style flows and API wiring.
- Extend Command Center with auxiliary task overrides and a Usage panel (7d/30d/90d).
- Fix titlebar geometry for WSL/Windows (native overlay width, tool spacing).
- Remove stray merge conflict markers from pyproject.toml optional deps.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(title-bar): position sidebar toggle button
* feat(desktop): composer queue — queue many, edit/delete/cancel-edit, Cursor-style
Press Enter while busy with a draft to queue it; with no draft to interrupt
and send the next queued turn. Auto-drains one queued turn each time the
session settles, same as Cursor. Queue persists across reloads so an
interrupted-and-queued turn isn't lost on refresh.
Each queued row supports edit-in-composer (with explicit Save/Cancel),
send-now (↑), and delete. Drain skips only the entry currently being
edited so the rest of the queue keeps flowing.
Queue dequeue is transactional — an entry only leaves the queue after
`prompt.submit` is accepted, so a rejected submit doesn't drop the turn.
Also shrinks the `[interrupted]` marker to a muted one-liner and drops
its assistant footer so it stops looking like a real reply.
* fix(desktop): handle empty usage analytics totals
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(desktop): address PR review titlebar and usage races
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(desktop): add MCP settings and live subagent tree
Surface configured MCP servers in Settings with JSON edit/save and a gateway-backed reload action so users can manage tool servers without falling back to slash commands.
Track live subagent gateway events in a desktop store, show active subagent counts in the Agents statusbar item, and replace the Agents overlay stub with a live spawn tree for the active session.
* fix(desktop): move power-user views out of sidebar
Keep Cron and Profiles available through lower-prominence chrome entry points so the workspace sidebar stays focused on core chat navigation.
Co-authored-by: Cursor <cursoragent@cursor.com>
* refactor(desktop): subagent overlay reads like a live transcript, not a dashboard
Strip the card chrome and rewire /agents to feel like peeking into the
child agent's stream:
- subagents store: single `stream` of typed entries (thinking/tool/progress/
summary) replaces the parallel notes/thinking/tools arrays. Drop unused
fields (toolsets, depth, apiCalls, reasoningTokens, sessionId).
- agents view: no OverlayCards, no boxed stream, no per-row borders. Goal +
status pill + indented stream lines, full row width.
- Group root spawns into "Delegation N" sections when batch shape + spawn
time match — hides task-index interleaving and makes hierarchy obvious.
- Sort tree by spawn time, then task_index. Step indicator is one colored
pill (primary while running, emerald when done) inside the row, not a
trailing pill that wrapped under the chevron.
- Tree picks up `subagent.start` (not only `spawn_requested`) and prunes
delegate-tool fallback rows once native subagent events land for the
session — fixes duplicate "Delegated task" rows alongside the real ones.
* feat(desktop): Esc closes every OverlayView-based overlay
Lift the keyboard handler into the shared OverlayView so Agents, Settings,
Command Center — and anything we build on top of it later — all dismiss on
Esc by default. Nested Radix dialogs stop propagation themselves, so a
modal opened inside an overlay (e.g. model picker inside Settings) still
closes the modal first, not the overlay underneath.
Drop the now-redundant Esc handlers in Settings (kept Cmd/Ctrl+P) and
Command Center.
* fix(desktop): drop numbered step pill on subagent rows
The pill was getting clipped at the overlay edge anyway. Just use the
status glyph (●/✓/✗/■/○) — the delegation header already conveys
"3 workers, 3 active", and order in the list implies which step you're
looking at.
* fix(desktop): drop noisy "returned N items / empty object" stub strings
When a tool returns nothing useful, the row should be silent — the title
("Search Files", etc.) already tells the user what happened. Counting the
fields in an opaque payload is engineer-noise.
`formatToolResultSummary` and `minimalValueSummary` now return '' for
empty arrays / records / unrecognized values; tool-fallback already hides
the detail section when its body is empty.
* refactor(desktop): subagent rows borrow chat tool patterns (fade-in, lucide glyphs, shimmer)
Pull the agents view closer to how chat tool blocks render:
- statusGlyph() returns the same lucide BrailleSpinner / CheckCircle2 /
AlertCircle vocabulary as tool-fallback's statusGlyph
- Stream lines fade-in via useEnterAnimation (one-shot WAAPI), keyed per
entry so streamed deltas settle in instead of popping
- Subagent rows fade in too, and pick up the existing data-slot=tool-block
spacing rules between blocks
- Active stream line trails a BrailleSpinner instead of a hand-rolled
pulsing rectangle
- Goal text drops FadeText (which forces nowrap); keep FadeText only for
the single-line meta subtitle
- Running rows shimmer the title — same affordance the chat thinking row
uses
* refactor(desktop): make /agents subagent-only, drop sidebar + dead sections
Activity rail and History stub were both noise. Strip the split layout,
sidebar, route enum, and the rail/stub helpers — the overlay is now just
the spawn tree, centered in a max-w-3xl column so it stops claiming the
whole screen for one section's worth of content.
* feat: update cron modals
* Add dedicated GUI log stream for dashboard debugging.
Capture dashboard and PTY websocket lifecycle failures in gui.log and expose it via hermes logs.
* Improve desktop runtime UX by surfacing inference readiness in gateway status and hardening WSL link opening.
This also stabilizes markdown code/table block spacing and adds root-install guards so desktop dev runs use a healthy workspace dependency tree.
* Log detailed GUI websocket failure metadata.
Capture richer reject/disconnect/send/parse context for dashboard gateway websocket flows so GUI connection failures are diagnosable from logs.
* Default dashboard startup logging to GUI mode.
Detect the dashboard subcommand during early CLI bootstrap so gui.log is attached from process start and GUI startup failures are always captured.
* Clean up gateway status conditionals and logging bootstrap mode detection.
Simplify nested dashboard gateway status branches for readability and use a concise first-subcommand check when selecting early GUI logging mode.
* add logging to nsis installer
* feat: glass ui pass
* fix(desktop): persist inline assistant errors across hydrate/resume
- Detect provider failure text arriving via message.complete
(HTTP 4xx, "API call failed after N retries", Provider/Gateway
error: ...) and persist as an inline assistant error instead of
regular completion text, blocking the hydrate that was wiping it.
- preserveLocalAssistantErrors: merge by id so same-id hydrated
messages keep their local error, and preserve the optimistic
user+error pair as a unit (with tail-user dedupe).
- Hook all hydrate/resume writers (use-session-actions resume +
fallback, hydrateFromStoredSession, syncSessionStateToView) into
the merge so stale snapshots can't clobber a failed turn.
- Add error to chatMessagesEquivalent so the resume diff actually
sees error-only changes and paints them.
- editMessage on a failed turn now submits a plain resend (no
truncate_before_user_ordinal) and retries plainly on the
"no longer in session history" race.
Style polish on touched files:
- Inline error: text-only treatment (no card).
- User stop / edit-composer send: shared Tabler IconPlayerStopFilled
glyph + shared icon-button class slot for parity.
* feat(desktop): theme xterm with active light/dark mode
The right-sidebar terminal hardcoded a light palette, which read poorly
on the dark glass surface. Subscribe to `useTheme().resolvedMode` and
hot-swap `term.options.theme` so Shift+X (and any other mode change)
updates the terminal in place without tearing down the PTY session.
Dark mode uses xterm's built-in defaults (white fg/cursor + vivid ANSI
16) with just a transparent background so the glass shows through;
light mode keeps the existing hand-tuned overrides for legibility on a
bright surface.
* feat(sidebar): right-click + drag-reorder sessions and workspaces
- Wire right-click on session rows to open the same actions menu;
suppresses the OS-native context menu so Windows stops looking awful.
- Share dropdown + context menu items via useSessionActions() driving
a single declarative ItemSpec[]; render polymorphic over MenuItem.
- New shadcn ContextMenu primitive mirroring DropdownMenu styling.
- Restore drag-and-drop reordering for Agents (lost during the cwd
cleanup) and add reordering of workspace groups via a right-side
grab handle. Pinned reorder unchanged.
- Generic orderByIds<T> replaces the duplicated session/group orderers;
useSortableBindings() hook collapses the two Sortable wrappers.
- cursor-pointer on every actionable element; cursor-grab on handles.
- KISS pass: baseName() helper, AGE_TICKS table, single WORKSPACE_PAGE
constant, flatter SidebarSessionsSection render.
* feat(desktop): solarize the xterm palette in both light & dark
xterm's default ANSI 16 is tuned for dark and reads candy-bright on the
light glass surface (vivid cyans/greens). Ship the canonical Solarized
palette (Schoonover) for both modes — same 16 accents either way, only
fg/cursor swap between `base00/01` (light) and `base0/1` (dark), so a
prompt's colors look uniform across a Shift+X toggle.
Background stays transparent in both modes — Solarized's cream/slate
backgrounds would fight the glass.
* feat(desktop): virtualize chat thread + sidebar via TanStack Virtual
Replaces `use-stick-to-bottom` and per-row session rendering with
`@tanstack/react-virtual`, matching what Cursor uses.
Chat thread (`thread-virtualizer.tsx`):
- Natural-flow virtualization (padding spacers, not absolute items) so
`position: sticky` on the human bubble still resolves cleanly against
the scroller.
- Custom at-bottom anchor: pins when armed, disarms on user-driven
upward scroll, re-arms at bottom, jumps on session switch +
`thread.runStart`.
- Loading indicator and `--thread-last-message-clearance` move to a
real `[data-slot=aui_composer-clearance]` node; drops the brittle
`:nth-last-child(1 of …)` rule that can't fire reliably under
virtualization.
Sidebar (`virtual-session-list.tsx`):
- Flat agents list virtualizes at >=25 rows; pinned and
workspace-grouped paths stay direct-render.
- `SortableContext` keeps all IDs; only the window mounts; dnd-kit's
`setNodeRef` is merged with `virtualizer.measureElement` so rows
participate in both DnD hit-testing and TanStack measurement.
Drops `use-stick-to-bottom`. Streaming test gets a global
`offsetWidth/offsetHeight` stub so the virtualizer's viewport sizing
works in jsdom; the scroll-up-doesn't-pull-back invariant still passes.
* feat: more ui qa
* fix(desktop): trim sidebar terminal startup spacer
Drop zsh's initial spacer row before writing the first terminal prompt so new sidebar terminal sessions do not open with a selectable blank line.
* chore: uptick
* feat(desktop): thin installer + first-launch install.ps1 bootstrap
Converges the Windows packaged desktop installer onto a single canonical
install topology: drop the Electron shell only (~80MB instead of ~500MB),
clone Hermes Agent at a build-time-pinned commit on first launch via
install.ps1's stage protocol, and treat the resulting git checkout at
%LOCALAPPDATA%\hermes\hermes-agent\ as the canonical install location
(same path the CLI installer uses). Future updates flow through the
existing applyUpdates() git-pull path.
Replaces the previous fat-installer architecture where the .exe bundled
a pre-staged hermes-agent source tree under resources/hermes-agent/ that
was then sync'd into ACTIVE_HERMES_ROOT at launch -- a complicated
factory-vs-active dance with several footguns (FACTORY_HERMES_ROOT
mismatch on path resolve, isGitCheckout guard regressions, pyproject
hash drift detection inside the sync loop).
Architecture overview
---------------------
Build time
apps/desktop/scripts/write-build-stamp.cjs writes
apps/desktop/build/install-stamp.json with {commit, branch, builtAt,
dirty}. Honours $GITHUB_SHA / $GITHUB_REF_NAME in CI, falls back to
`git rev-parse HEAD` locally.
apps/desktop/scripts/stage-native-deps.cjs copies the runtime subset
of @homebridge/node-pty-prebuilt-multiarch from the workspace-root
node_modules into apps/desktop/build/native-deps/. Workspace dedup
hoists this dep to the root, out of reach of electron-builder's
`files:`-restricted collector; staging gives us a deterministic
path to extraResources.
electron-builder ships both into resources/install-stamp.json and
resources/native-deps/ respectively.
Boot resolver (electron/main.cjs)
Resolver order:
1. HERMES_DESKTOP_HERMES_ROOT override
2. SOURCE_REPO_ROOT (dev mode)
3. ACTIVE_HERMES_ROOT git checkout WITH .hermes-bootstrap-complete
marker -- the post-install fast path
4. `hermes` on PATH (CLI-installed user adding the desktop)
5. pip-installed hermes_cli via system Python
6. bootstrap-needed sentinel -> hand off to runBootstrap
Deletes the entire FACTORY_HERMES_ROOT / RUNTIME_MARKER /
syncTreeExcludingVenv machinery (-200 lines). The isGitCheckout
guard that bit us in the install.ps1 PR is gone.
First-launch bootstrap (electron/bootstrap-runner.cjs)
1. Resolve install.ps1: prefer SOURCE_REPO_ROOT/scripts (dev), else
download from GitHub raw at INSTALL_STAMP.commit (cached at
HERMES_HOME\bootstrap-cache\install-<sha>.ps1).
2. Fetch the stage manifest via install.ps1 -Manifest -Commit X
-Branch Y.
3. Iterate stages: install.ps1 -Stage <name> -NonInteractive -Json
-Commit X -Branch Y per stage.
4. On all stages green: write the .hermes-bootstrap-complete
marker with {schemaVersion, pinnedCommit, pinnedBranch,
completedAt, desktopVersion}.
Per-run log to HERMES_HOME\logs\bootstrap-<ts>.log. Cancellation
via AbortSignal. Manifest cache so retries don't re-download.
Install overlay (src/components/desktop-install-overlay.tsx)
Mounted alongside the existing onboarding overlay; flexbox card
with header (static) + middle (scrollable) + footer (failure-only,
static). Subscribes to hermes:bootstrap:event IPC + resyncs from
hermes:bootstrap:get on mount/reload. Renders:
- 14-stage checklist with per-stage state icons
- Overall progress bar + current-stage spotlight
- Auto-expanded installer-output panel on failure
- "Copy output" button (full ring buffer + error to clipboard)
- "Reload and retry" wired through hermes:bootstrap:reset to
clear main.cjs's latched failure
Synthetic empty-manifest event from main.cjs flips the overlay to
'active' immediately so the slow install.ps1 download doesn't
leave the user staring at the generic Preparing splash.
Failure latching (main.cjs)
bootstrapFailure module-scope variable holds the rejection after
install.ps1 fails. startHermes() throws the latched error
immediately when set, bypassing the entire ensureRuntime +
runBootstrap chain. Without this, the renderer's ensureGatewayOpen
retries would re-run install.ps1 in a 5-10 min hot loop while the
user was still reading the failure overlay. Cleared via
hermes:bootstrap:reset on user-driven retry.
Unsupported-platform overlay (1F)
macOS / Linux packaged builds (no install.sh stage protocol yet)
emit an unsupported-platform event with a copy-pasteable install
command + docs URL. Dedicated overlay branch with "Copy command"
+ "I've run it -- retry" buttons.
install.ps1 additions (Phase 1F.3 + 1F.5)
-----------------------------------------
New -Commit and -Tag string params. Precedence Commit > Tag >
Branch. Honoured by all three code paths (update / fresh clone /
ZIP fallback), with archive URL selection that handles each
ref-type variant. Detached-HEAD checkouts intentionally -- they're
pins, not branches the user pulls into.
EAP=Continue wrap around the new pin-step git invocations. `git
fetch origin <commit>` writes the routine 'From <url>' info line to
stderr; under the script's global EAP=Stop that terminates the
script even though fetch+checkout succeed. Matches the established
pattern in Install-Uv, Test-Python, _Run-NpmInstall.
Backend fix (hermes_cli/web_server.py)
--------------------------------------
CORS allow_origin_regex now accepts Origin: 'null'. Packaged
Electron loads index.html via file://; Chromium sets the WebSocket
upgrade Origin header to the opaque origin 'null', which the old
regex rejected with HTTP 403 before gateway_ws() ever ran. This
failure mode was masked in the older FACTORY_HERMES_ROOT
architecture because the resolver often found an existing hermes
on PATH with different binding behavior.
Security maintained: localhost-only bind keeps cross-machine pages
out; per-process session token still gates every authenticated
/api/ endpoint regardless of Origin.
Desktop QoL
-----------
DevTools is now enabled in packaged builds (F12 / Cmd+Opt+I).
Field-debugging trade-off: tiny attack surface increase versus
a much better support story when CSP / WS / theme issues surface.
NSIS prereq-check page deleted (-767 lines). The standard
Welcome -> License -> Directory -> InstallFiles -> Finish wizard
now installs without custom Python/Git/ripgrep detection -- those
prereqs are install.ps1's job at first launch.
Test infrastructure (Phase 1G)
------------------------------
apps/desktop/scripts/test-desktop.mjs rewritten as a cross-platform
bundle validator (was darwin-only and asserted on dead factory-
payload paths):
NEGATIVE: hermes_cli/main.py is NOT shipped (regression guard)
POSITIVE: install-stamp.json carries a real commit + branch
POSITIVE: node-pty native deps shipped under resources/native-deps
POSITIVE: renderer dist/index.html reachable (asar or unpacked)
New nsis mode and npm run test:desktop:nsis script.
Validated end-to-end on clean Win10 VM
--------------------------------------
Confirmed: NSIS installer drops Electron shell, app launches,
install overlay shows progress, install.ps1 clones the pinned
commit, 14 stages run to completion, marker written, backend
spawns, WebSocket connects, onboarding overlay asks for API key,
main UI loads, integrated terminal works.
Failures handled: bootstrap stays failed (no hot-loop retry),
"Copy output" gives actionable transcript, "Reload and retry"
explicitly re-runs install.ps1.
What's deferred
---------------
- MSIX wrapping (Phase 2): same Electron .exe under MSIX manifest
with runFullTrust, signed and submitted to Microsoft Store.
- install.sh stage protocol parity (Phase 2): once shipped, the
unsupported-platform overlay becomes drive-it-yourself and
macOS/Linux packaged installers gain feature parity with Windows.
* feat(desktop): persistent terminal pane + fullscreen takeover
Adds a VSCode-style "focus terminal" toggle to the right sidebar's Terminal
tab that takes over the chat pane area without unmounting the shell. The
xterm host is mounted once at the layout root and CSS-overlayed onto
whichever <TerminalSlot /> is currently active, so the PTY session,
scrollback, selection, focus, and WebGL renderer survive every toggle.
Also:
- WebGL renderer (matching dashboard ChatPage) so Hermes' TUI skins paint
faithfully instead of muting through xterm's default DOM renderer
- File drag/drop from the project tree or OS into xterm — paths are
shell-quoted (zsh/bash/pwsh/cmd) and written straight into the PTY
- Solarized dark canvas with brights promoted to real accent variants
(Schoonover's UI-gray brights washed out every TUI accent)
- Strip NO_COLOR/FORCE_COLOR/COLORFGBG/TERM=dumb leaking from non-tty
parents (CI runners, Cursor's agent shell) so the embedded shell gets
truecolor regardless of how Electron was launched
- rAF-debounced ResizeObserver — running fit.fit() synchronously during
sibling pane transitions crashed the WebGL texture-atlas rebuild
* fix(install.ps1): strip UTF-8 BOM regression that broke 'irm | iex'
The canonical install flow
irm https://raw.githubusercontent.com/.../scripts/install.ps1 | iex
fails on PowerShell 5.1 with a cascade of 'The assignment expression
is not valid' errors at every param() default value:
[string]$Branch = 'main',
~~~~~~
The assignment expression is not valid. The input to an assignment
operator must be an object that is able to accept assignments...
Root cause: scripts/install.ps1 carries a UTF-8 BOM (0xEF 0xBB 0xBF)
as its first three bytes. 'irm' returns the response body as a string;
on PS 5.1 the BOM survives into that string as a leading \ufeff
character. 'iex' then evaluates the string and PS's parser chokes
on the invisible character before param() -- error recovery proceeds
into the body but every assignment is reported as broken.
This was the exact failure mode the install.ps1 hardening pass (PR
#27224) deliberately fixed by stripping the BOM and ensuring the
file body is pure ASCII. Commit 4279da4db ('fix(windows): make
PowerShell installer parse in 5.1') re-introduced the BOM later,
unintentionally undoing the irm|iex compatibility fix; the merge
that brought it into bb/gui carried it forward.
Fix: strip the three BOM bytes. File body is verified pure ASCII
(any-byte > 127 returns false), so PS 5.1 with no BOM falls back to
Windows-1252 decoding which is identical to ASCII for our content.
Both install paths now work:
- 'irm ... | iex' (canonical CLI)
- 'powershell -File install.ps1' (programmatic / desktop bootstrap)
* install.ps1: detect ARM64 Windows reliably for Node and Git stages
Add a Get-WindowsArch helper that reads Win32_Processor.Architecture
via CIM (invariant to PowerShell host bitness) with PROCESSOR_ARCHITEW6432
fallback. Use it in:
- Install-Git: previously only triggered the arm64 PortableGit asset
when invoked from a native-ARM64 PowerShell host. WoW64 / emulated
x64 hosts (the default powershell.exe on Windows-on-ARM) saw
PROCESSOR_ARCHITECTURE=AMD64 and fell through to the x64 PortableGit
build, leaving ARM64 users on emulated Git for Windows.
- Test-Node: previously hardcoded the Node download to win-x64 on any
64-bit OS, so ARM64 users always got x64 Node under Prism emulation
even though Node ships an arm64 build for Windows. The winget
fallback now also passes --architecture arm64 on ARM64.
Python remains x86_64 by design: uv intentionally prefers
windows-x86_64 cpython on ARM64 hosts for ecosystem (wheel)
compatibility (see astral-sh/uv#19015).
* install.ps1: harden Install-SystemPackages against winget msstore failures
The previous winget invocation discarded stdout/stderr and trusted no
signal at all -- not the exit code (winget exits 0 even when it bails
"please specify --source"), not output (sent to Out-Null), not the
catch handler (winget returning 0 means no exception fires). The only
trust signal was a post-install Get-Command rg / Get-Command ffmpeg
check, which would also miss the package because %LOCALAPPDATA%\
Microsoft\WinGet\Links (where winget puts command aliases) is added to
PATH by AppExecutionAlias machinery only in fresh shells. End result on
machines where the msstore source has a cert problem (0x8a15005e --
common on Windows-on-ARM and some corporate networks): silent failure,
no log, no breadcrumb, and the user is told the install succeeded.
Specifically:
- Pin --source winget on every winget install call. Defeats the broken-
msstore-source path. We ship nothing from msstore so this is safe and
forward-compatible.
- Add --exact --id for a tighter package match.
- Capture each winget invocation's combined stdout/stderr + exit code to
%TEMP%\hermes-winget-<pkg>-<n>.log instead of Out-Null. On the happy
path the log is deleted after the post-install check confirms the
binary is on PATH; on failure the log is kept and its path is named in
a Write-Warn so the user has something to grep.
- Refresh PATH to include %LOCALAPPDATA%\Microsoft\WinGet\Links in
addition to the User/Machine env-var hives, so Get-Command sees newly-
installed winget aliases in the same process.
- No behavior change on the happy path. Same Write-Info/Success/Warn
cadence, same fallback order (winget -> choco -> scoop -> manual),
same $script:HasRipgrep / $script:HasFfmpeg outputs.
Verified end-to-end on a real Snapdragon ARM64 Windows host: ripgrep
uninstalled, stage re-run, [OK] ripgrep installed in 1.4s, ok:true.
* desktop: swap node-pty fork for upstream microsoft/node-pty 1.1.0
The previous dependency, @homebridge/node-pty-prebuilt-multiarch@0.13.1,
publishes no win32-arm64 prebuilds on its v0.13.x line, and its v0.14.x
betas (which do add an arm64 Windows build) ship no electron-vXXX-win32-
arm64 prebuilds at all -- so packaged Electron 40 builds (NMV 143) would
fail at runtime even on a successful npm install. Net effect: the
desktop's integrated terminal was unbuildable on Windows-on-ARM, in
both dev (npm install fails: 404 fetching the node-vXXX-win32-arm64
prebuilt) and packaged builds (no Electron-ABI prebuilt exists).
The homebridge fork was originally created because upstream node-pty
shipped no prebuilds at all. That hasn't been true since node-pty@1.0
(April 2024), which:
- bundles prebuilts for mac (arm64+x64) and Windows (arm64+x64) directly
inside the npm tarball -- no GitHub-Releases fetch, no missing-binary
failure mode
- uses N-API (node-addon-api) for ABI stability across Node and Electron
major versions, so the same pty.node binary loads under Node 22 (dev)
and Electron 40+ (packaged) without per-ABI rebuilds
- is what VS Code, Hyper, and Theia actually ship
API surface is identical (spawn / onData / onExit / write / resize /
kill) -- no call-site changes needed.
Specifically:
- apps/desktop/package.json: replace the @homebridge fork with
node-pty@1.1.0 (exact pin). Widen `asarUnpack` from `["**/*.node"]`
to also unpack `**/prebuilds/**`, because node-pty ships runtime-
execed helpers alongside its .node files (darwin spawn-helper has no
extension and would not be matched by `**/*.node`; conpty.dll,
OpenConsole.exe, winpty.dll, winpty-agent.exe on Windows are also
exec'd at runtime and cannot live inside asar).
- apps/desktop/electron/main.cjs: update both require() strings to
match the new package name and the new staged path under
resources/native-deps/node-pty/.
- apps/desktop/scripts/stage-native-deps.cjs: point at node_modules/
node-pty. node-pty's prebuilts live under prebuilds/<plat>-<arch>/
(not build/Release/), so update the include glob to copy that dir.
Per-arch staging keeps the resource bundle small (target arch comes
from npm_config_arch when electron-builder cross-builds, else
process.arch). Explicitly enumerate file types in the prebuilds glob
so the ~25 MB of .pdb debug symbols that prebuild-install bundles
for Windows crash analysis don't bloat the installer (29 MB -> 2.6 MB
staged on win32-arm64). Re-assert +x on the darwin spawn-helper
defensively, since a stripped mode bit would manifest as a silent
ENOENT at first pty.spawn().
- apps/desktop/scripts/test-desktop.mjs: update expectedNativeDepPaths()
and its assertion site to look at prebuilds/<plat>-<arch>/ instead of
build/Release/. Add an explicit spawn-helper-exists check on darwin
so a regression in the asarUnpack glob would fail loudly in CI rather
than at first PTY spawn.
Trade-off: Linux end-users lose prebuilts and fall back to building
node-pty from source on `npm install`. Acceptable because Hermes
ships no Linux desktop builds (desktop-release.yml matrix is mac + win
only, package.json declares no `linux` target), and Linux developers
hacking on the desktop already need a C++ toolchain for the rest of
the stack.
Verified on Windows 11 ARM64 (Snapdragon):
npm install -> exit 0
node -e "require('node-pty').spawn(...)" round-trip -> OK
stage-native-deps -> 27 files, 2.6 MB
load from staged tree (simulates packaged fallback) -> ConPTY
round-trip OK
* desktop+gateway: harden Slack socket recovery and Windows restart dedupe (#28873)
* desktop+gateway: harden Slack socket recovery and Windows restart dedupe
Fix Slack Socket Mode reliability by adding a watchdog/reconnect path so silent socket task drops no longer leave the adapter stuck. Harden Windows gateway lifecycle by avoiding desktop-binary path collisions, making gateway PID scans case/extension tolerant, and reusing in-flight restart actions to prevent duplicate gateway spawns.
* test(slack): add Socket Mode watchdog/reconnect behavioural coverage
Drive the new Slack Socket Mode self-healing logic through a fake AsyncSocketModeHandler so we can simulate the P0 silent-hang failure mode (task exit, transport disconnected, intentional shutdown, concurrent reconnect attempts) without touching real Slack.
* fix(slack,desktop): address Copilot review on watchdog races and path normalization
- connect(): explicitly cancel + await the prior socket watchdog before flipping _running, so an old monitor cannot exit between teardown and respawn (Copilot #1)
- _socket_watchdog_loop: wrap the body in try/except + add a done-callback that respawns on unexpected crash, so a transient bug cannot permanently disable self-healing (Copilot #2)
- normalizeExecutablePathForCompare: use the resolved path for realpathSync so non-string inputs cannot leak through (Copilot #3)
- Add tests for crash-recovery and atomic watchdog replacement across reconnects
* fix(slack): tighten connect() error path and clarify watchdog test intent
Address Copilot review round 2.
- connect(): wrap _start_socket_mode_handler/_ensure_socket_watchdog in a focused try/except so any failure rolls back partially-started handler/task state and leaves _running=False, ensuring the platform lock is always released by the outer finally
- Defer _running=True until after the handler is actually started so the watchdog observes a live socket task immediately and never spins against a half-built adapter
- Rename test_watchdog_self_restarts_after_unexpected_crash to test_watchdog_cancellation_does_not_respawn (matches what it actually asserts) and add test_watchdog_unexpected_exit_respawns_via_done_callback that drives a real RuntimeError through _on_socket_watchdog_done and verifies a fresh task replaces the crashed one
* fix(web_server): serialize action spawn check+store under a threading lock
Address Copilot review round 3.
FastAPI runs sync handlers on its threadpool, so two near-simultaneous /api/gateway/restart (or /api/hermes/update) requests could both observe "no live process" in _spawn_hermes_action's poll-based dedupe and double-spawn. Add a module-level _ACTION_SPAWN_LOCK around the entire check + Popen + _ACTION_PROCS store sequence so the dedupe is atomic across threads.
* fix: address Copilot review round 4
- slack.disconnect(): mirror connect()'s defensive cleanup — catch the broad Exception path on watchdog await so handler shutdown and lock release still run if the watchdog raised before cancellation took effect
- web_server._spawn_hermes_action: wrap subprocess.Popen in try/except so a missing executable / permission error closes the log file handle, writes a failure marker, and re-raises instead of leaking a file descriptor
- gateway._scan_gateway_pids: drop the over-broad "hermes.exe --profile" / "hermes.exe -p" patterns that would match any Hermes CLI subcommand using a profile flag (e.g. `hermes.exe --profile foo dashboard`); rely on the "hermes.exe gateway" + "hermes-gateway.exe" tokens instead
- tests: tighten _fake_create_task to assert coroutine input and return a real asyncio.Task that stays pending until pytest teardown, and update the three callsites whose mocked AsyncSocketModeHandler.start_async returned a non-coroutine value
* fix(slack): reset multi-workspace state on reconnect
Address Copilot review round 5.
connect() is reentrant (gateway restart, in-process reconnect), but it was leaving _bot_user_id / _team_clients / _team_bot_user_ids populated from the previous session. A reconnect that rotated the primary token or dropped a workspace would silently keep the stale bot user id and stale workspace client maps, leading to dispatch against gone workspaces.
Clear these three pieces of state right after _stop_socket_mode_handler() and before the auth_test loop, then let the loop repopulate from the current tokens. Add test_reconnect_refreshes_multi_workspace_state to lock it in.
* nix: package apps/desktop as .#desktop (#28964)
Adds nix/desktop.nix building the Electron renderer with buildNpmPackage
and wrapping nixpkgs' electron binary. Reuses .#default by setting
HERMES_DESKTOP_HERMES to its hermes binary, so the desktop's resolver
picks up the fully-wired nix hermes (venv, bundled skills/plugins,
runtime PATH) without reimplementing agent resolution.
- nix/desktop.nix: renderer + electron wrapper
- nix/hermes-agent.nix: finalAttrs form, exposes hermesDesktop in passthru
- nix/packages.nix: exposes .#desktop + adds to fix-lockfiles
- apps/desktop/package-lock.json: standalone hermetic lockfile
nix build .#desktop && nix run .#desktop both clean.
* fix(desktop): probe steps 4 & 5 of resolveHermesBackend before trusting
A user-reported failure on Windows-on-ARM: a pre-installed Python 3.13
on PATH makes findSystemPython() succeed, so resolveHermesBackend
returns a backend pointing at it -- but hermes_cli isn't in that
interpreter's site-packages. The spawn dies with ModuleNotFoundError
and the user sees a dead GUI instead of the first-launch installer.
Same shape can hit step 4 (existing `hermes` on PATH) when a stale
shim survives a partial uninstall.
Add cheap exit-code probes -- `python -c "import hermes_cli"` for
step 5, `<hermes> --version` for step 4 -- and fall through to step 6
(bootstrap-needed) on failure. install.ps1 then runs as if on a clean
box and the venv gets built.
Probes live in a standalone electron/backend-probes.cjs module so they
can be unit-tested with node --test, same pattern as bootstrap-platform.cjs
and hardening.cjs. New test file wired into test:desktop:platforms.
* test(desktop): allow `node-pty` bare-require in packaged entrypoints
Pre-existing failure on bb/gui since c858484b4 swapped the node-pty
fork for upstream microsoft/node-pty 1.1.0. main.cjs intentionally
bare-requires node-pty (it's hoisted by workspace dedup in dev, and
staged to resources/native-deps via scripts/stage-native-deps.cjs +
extraResources for packaged builds, with a try/catch fallback at
line ~38). The allowlist hadn't been updated to match -- same shape
as `electron`, which was already allowed.
* chore(deps): refresh root lockfile for dashboard @nous-research/ui 0.14.0
apps/dashboard/package.json was bumped to @nous-research/ui 0.14.0 (+
flag-icons ^7.5.0, motion ^12.38.0) but the root package-lock.json was
never refreshed. Running `npm install` from the repo root now
materialises 0.14.0's transitive closure (launder, bumps for
@nanostores/react, nanostores, sanitize-html, tailwind-merge).
No code changes; purely a lockfile catch-up so fresh checkouts on bb/gui
get a working dashboard install.
* chore(desktop): bump version to 0.0.1
First non-placeholder version so electron-builder's artifactName template
produces `Hermes-0.0.1-win-x64.exe` instead of the obviously-unreleased
`Hermes-0.0.0-...`. No release process yet; this just stops the artifact
filename from telling users "you got a debug build."
Bumped in three slots that all carry the desktop app's version:
- apps/desktop/package.json (source of truth)
- apps/desktop/package-lock.json (per-app lockfile, kept for CI parity)
- root package-lock.json's apps/desktop workspace entry
Identity-of-build for first-launch bootstrap continues to come from
build/install-stamp.json (commit SHA + builtAt), unchanged.
* fix: fs icon color
* perf(desktop): cut per-keystroke layout + listener churn in chat composer
Empirical work via CDP harnesses under apps/desktop/scripts/ (see
profile-typing-lag.md):
jsListeners growth (per round of 200 chars + GC):
before: +35 (verified leak — listeners stuck after 1st trigger popover use)
after: +0
Four narrow edits in src/app/chat/composer/index.tsx:
1. Drop the per-keystroke `editorRef.current.scrollHeight` read used to
decide composer expansion. Replace with `draft.length > 60` heuristic;
the existing ResizeObserver still catches edge cases. `scrollHeight`
is a forced-layout call and was firing on every char until the first
wrap.
2. Bucket measured composer height to 8px before writing
`--composer-measured-height` / `--composer-surface-measured-height`
on `documentElement`. Without this, the editor grows ~1px per char,
setProperty fires every keystroke, computed style is invalidated tree-
wide.
3. Remove the dead `$composerDraft` two-way sync. Nothing outside the
composer subscribed to that atom (verified via grep). Two useEffects
on `[draft]` were pushing draft→atom and atom→aui per keystroke for
no consumer. Also drop the per-keystroke
`reconcileComposerTerminalSelections` call; it was pruning stale
labels for `terminalContextBlocksFromDraft`, but that helper already
ignores labels not in the current submitted text, so pruning per
keystroke was just bookkeeping.
4. `refreshTrigger` fast-bails when the draft contains neither `@` nor
`/`. Previously `textBeforeCaret(editor)` ran on every input/keyup
regardless; `range.toString()` inside is O(n) over draft length.
Synthetic typing latency p50/p90/p99 is similar before vs after on a
freshly-loaded session (Blink can already handle ~30cps typing into a
contentEditable on its own); the real win is the listener leak being
gone and the global computed-style invalidations dropping ~8× when the
composer is sitting at a fixed height row.
The `Enter → stall` follow-up (see profile-typing-lag.md §"Submit /
TTFT stall") is unmeasured here — needs a throwaway session because
the harness fires a real prompt. Not blocking this commit.
* perf(desktop): cut FadeText forced layouts during streaming
The slowest user-felt path is typing into the composer while the
assistant is streaming. Profile (scripts/profile-under-stream.mjs):
FadeText measureOverflow self time: 35.8 ms → 18.1 ms (-50%)
total active CPU during 7s window: ~150 ms → ~50 ms
Two changes in src/components/ui/fade-text.tsx:
1. Drop the `useEffect([children])` that re-ran `measureOverflow`
(reads scrollWidth + clientWidth — forced layout) on every parent
re-render. `useResizeObserver` already fires the same callback on
mount and whenever the host span's box size changes; that covers
the only case where overflow state can legitimately change. The
previous explicit useEffect was a forced-layout flush on every
parent render, which during streaming meant every token tick.
2. Wrap the component in `memo` with a custom comparator that
short-circuits the entire render when scalar string `children` and
the className/fadeWidth/style props are unchanged. The hot path
was tool-fallback's title chips being re-rendered by parent
streaming updates even though their text was stable; memo+
comparator skips that.
Also adds two harness scripts under apps/desktop/scripts/:
- latency-under-stream.mjs (key→paint latency while a turn streams)
- profile-under-stream.mjs (CPU profile while a turn streams)
Updates profile-typing-lag.md with the streaming numbers and confirms
the Enter→paint submit path is already fast (≤320ms on the populated
session; the 2s "stall after Enter" the user noticed once was a
one-time cold-start, not reproducible at the UI layer).
I'd guess the felt jank in real use is fast-burst typing during a
long-form streaming reply (code blocks + markdown lists multiply the
per-token render cost). The CPU savings here scale linearly with
token volume.
* chore(desktop): drop diag scratch scripts no longer needed
* docs(desktop): correct leak-typing numbers on a real session
Re-ran the leak harness on a populated session (Phaser thread) for both
unpatched and patched builds. The original 'listener leak' was transient
warm-up cost, not a steady-state leak — both versions show 0 listener
growth/round in steady state.
The load-bearing number is forced layouts per character:
unpatched (HEAD~2): 7.02 layouts/char
patched (HEAD): 2.35 layouts/char (3× fewer)
The patches reduce per-char forced-layout work to Blink's natural floor.
Document node count and heap are flat in both builds.
* perf(desktop): fix "Enter jumps up" on long threads
User reported: after pressing Enter on a long thread, the view jumps up
— the just-submitted message disappears below the fold. Confirmed via
apps/desktop/scripts/measure-jump.mjs:
before: distFromBottom 0 → 49.5px, sticks there permanently
after: distFromBottom 0 → ~0 (worst case 4px for one frame)
Root cause in useThreadScrollAnchor (thread-virtualizer.tsx):
1. The sticky-bottom logic disarmed on any scroll event where
`scrollTop < lastTopRef.current`. That check can't distinguish a
user scrolling up from a programmatic `pinToBottom` write that
the browser clamped short of bottom (because content also grew in
the same frame, so `scrollTop = scrollHeight` lands at
`scrollHeight - clientHeight` for the OLD scrollHeight, which is
now below the NEW scrollHeight). Result: sticky-bottom disarmed
permanently on the user's first submit.
2. There was no synchronous pin tied to React's commit phase. By the
time the ResizeObserver fired and re-pinned, the user had already
seen ~50ms of "message below the fold" — visually that reads as the
view jumping up.
Fix:
- `programmaticScrollPendingRef` counter tracks scroll events we
expect to be ours (one per `pinToBottom` write). The scroll handler
skips the disarm check when consuming a pending tick, keeps the
arm bit true, and re-pins synchronously if the browser clamped us
short of bottom. A depth cap (8) breaks runaway loops in
pathological streaming-burst layouts.
- `useLayoutEffect` on `groupCount` increase pins BEFORE the browser
paints, eliminating the visible ~50ms window between optimistic
user-message insert and the RO/scroll-event chain firing.
Verified on the long Cloud Shadows thread (7-8 turns, ~11k px tall):
all three repro runs now hold within 0–4 px of bottom across the
post-Enter transition. Submit latency unchanged (paint 77–107 ms),
streaming-typing latency unchanged.
Also adds three debug harnesses:
- measure-jump.mjs — sample thread scroll across Enter
- probe-thread.mjs — dump current thread / scroll state
- diag-jump.mjs — intercept scrollTop + RO + mutations across Enter
* perf(desktop): rate-limit thread auto-pin during streaming
Follow-up to the Enter-jump fix. The first version did a synchronous
re-pin loop inside the on-scroll handler when the browser clamped our
`scrollTop = scrollHeight` write short of the new bottom; that gave a
tight 4 px visible jump on Enter, but during streaming the
ResizeObserver fires many times per second as content grows, and each
RO callback re-entered the pin loop. CPU profile showed
`Virtualizer.getMaxScrollOffset` climbing to 22 ms self over a typing-
during-streaming window — the sync re-pin path was paying tanstack-
virtual's recompute cost ~3× per token.
Re-architect:
- RO callback coalesces to one pin per animation frame. Streaming-rate
RO bursts now cost the same as a single per-frame pin.
- The on-scroll programmatic-counter guard remains (it's what prevents
the false-disarm bug when the browser clamps a write). It no longer
does sync re-pins; the next RO/rAF will catch up.
- The useLayoutEffect on groupCount (the path that fires on user
submit / new turn arrival) ALSO schedules one rAF pin in addition to
the synchronous pin. This catches the case where React mounts the
new message in a second commit (after our layout effect ran), which
grows scrollHeight again. Two pins instead of a tight loop, paid only
once per turn change.
Net effect on the Cloud Shadows long thread:
enter-jump transient: 12–20 px for 1 frame (was 49 px permanent)
CPU during stream+type: `getMaxScrollOffset` dropped out of top-5
self-time list
typing-during-stream: p50 ~10 ms paint, p99 ~20 ms (1 frame),
occasional 40 ms+ outliers during burst
token arrivals
Also adds scripts/profile-long-stream.mjs: 20-second streaming profile
with per-500ms FPS histogram + content-length tracking, so we can see
whether streaming render cost grows with message length (it doesn't —
sustained 60 fps).
* perf(desktop): use textContent for trigger precondition
Replace composerPlainText() call inside refreshTrigger's no-trigger
fast-bail with a textContent check. textContent is a browser-native
flat traversal; composerPlainText walks recursively with chip-aware
logic. We only need to know if @ or / appears; either way the trigger
char will be in textContent because chips contain @ in their refText.
Profile shows composerPlainText was ~18ms self over a 12s typing-during-
stream window, called from refreshTrigger on every keystroke. Most of
that was the precondition check (the trigger detection path is the
slow path but only runs when a trigger char is present).
* Revert "perf(desktop): use textContent for trigger precondition"
This reverts commit a6a78ff08a.
* Revert "perf(desktop): cut FadeText forced layouts during streaming"
This reverts commit 88e7d7537c.
* Revert "perf(desktop): cut per-keystroke layout + listener churn in chat composer"
This reverts commit bff1b3261d.
* Revert "Revert "perf(desktop): cut per-keystroke layout + listener churn in chat composer""
This reverts commit b7b378e3a4.
* Revert "Revert "perf(desktop): use textContent for trigger precondition""
This reverts commit 0739588f48.
* chore(desktop): synthetic-stream perf harness + scripts
Drops the React `<Profiler>` approach (no-op because Vite is currently
serving the production React build) in favor of an externally-observable
measurement stack: rAF frame intervals, `PerformanceObserver({entryTypes:
['longtask']})`, and a `MutationObserver` on the live streaming message.
Adds a synthetic stream driver — `window.__PERF_DRIVE__.stream({...})` —
that pushes tokens through the live `$messages` atom at a controlled rate,
so the assistant-ui runtime, incremental repository, and Streamdown
markdown pipeline see the same workload they'd see during a real LLM
stream, without the LLM cost.
The driver lives in `src/app/chat/perf-probe.tsx`; `main.tsx` side-imports
it under `import.meta.env.MODE !== 'production'` so it tree-shakes out of
prod builds. (Using `MODE` rather than `DEV` because our Vite setup
currently reports `DEV=false` even under `vite dev` — see the dev-build
note in `profile-typing-lag.md`.)
Scripts:
- measure-synthetic-stream.mjs drive synthetic + record frame/longtask/mutation
- profile-synth-stream.mjs CPU profile + top self-time during synthetic
- measure-real-stream.mjs same harness, real LLM stream
- profile-real-stream.mjs CPU profile bracketing the real stream window
- eval.mjs / reload.mjs small CDP helpers
A real-LLM measurement on Cloud Shadows (gpt-4o-mini, 39 s window) showed
12 longtasks in the same 75-127 ms range the synthetic predicted, so the
synthetic is a faithful proxy.
* perf(desktop): memo FadeText so it skips re-renders when text unchanged
FadeText is used 110+ times inside `tool-fallback.tsx` on a tool-heavy
thread. During streaming each parent re-render previously triggered the
component's `useEffect([children])`, which forced a `scrollWidth` layout
read even when the title text was unchanged. The `useResizeObserver` was
already covering the genuine resize case, so that effect was strictly
redundant work.
Drops the effect and wraps the component in `React.memo` with a custom
comparator that field-compares `className`, `fadeWidth`, and `style`,
plus identity-compares `children` (scalar fast-path; correct for JSX
nodes too since a new node should force a re-render).
Verified via temporary render counter on the 34 MB
`session_20260514_215353_fe0ac8` thread (110 FadeText instances): a
2 s synthetic stream went from ~11k FadeText render calls to 122 —
roughly one render per truly-new instance instead of one per parent
commit per instance.
Doesn't move the longtask needle on its own (Streamdown's markdown
re-parse dwarfs it) but eliminates a steady CPU floor and a class of
forced layouts during streaming. Profile-typing-lag.md documents the
full investigation, including the remaining Streamdown cost as the
real source of the perceived "5 fps moment" hitches.
* perf(desktop): memoize MarkdownText plugins to stop churning Streamdown
The inline `plugins={{ math: mathPlugin, ...(isStreaming ? {} : { code }) }}`
on `<StreamdownTextPrimitive>` constructed a new object literal on every
parent render. That broke `<Streamdown>`'s outer memo and forced its
internal `rehypePlugins` / `remarkPlugins` array useMemos to rebuild,
which propagates a new identity into every `<Block>` and defeats Block's
memoization for stable historical blocks.
After memoizing on `[isStreaming]` (the only real dimension of variance),
CPU profile during a 5 s synthetic stream on the 34 MB session shows
`parser` self-time dropping out of the top 10, `compile` cut roughly in
half, and `bn$1` / `m$1` (micromark internals) leaving the top entries.
Doesn't move the visible longtask count on its own — Streamdown's
per-Block parse cost still dominates whenever the last block's content
changes — but it removes a class of unnecessary re-parses for historical
blocks during streaming. See `scripts/profile-typing-lag.md` for the
full investigation.
* perf(desktop): floor assistant-text flush gap to 33ms for predictable batching
`scheduleDeltaFlush` previously coalesced via `requestAnimationFrame`
only. The "at most one flush per frame" guarantee that gives you is fine
for fast streams (>~80 tok/sec) where multiple tokens arrive within a
single frame, but breaks down at typical LLM token rates (30-80 tok/sec)
where each token arrives slower than the rAF cadence and triggers its
own React commit + Streamdown markdown re-parse.
Track `lastFlushAt` and require at least 33 ms between two flushes.
React 18+ auto-batching probabilistically already collapsed some of
these, but the floor makes it deterministic.
A/B on the 34 MB session, 300 tokens at 50 tok/sec (markdown chunks):
| | avgFps | p99 frame | LTs / 5 s | max LT |
|---|---|---|---|---|
| no floor (current rAF) | 54.0 | 38 ms | 2.0 | 145 ms |
| 33 ms floor (this PR) | 54.3 | 41 ms | 1.7 | 110 ms |
`inter-mutation` p50 also tightens from 22-28 ms to a clean 33 ms,
which is the expected signature of a deterministic floor. Doesn't fully
solve the user's perceived hitches — Streamdown's per-Block parse cost
when the last block grows past ~2 k chars is still the elephant — but
it consistently shaves the worst-case longtask and makes the streaming
cadence visibly steadier.
Also threads a matching `flushMinMs` option through the synthetic
stream driver in `perf-probe.tsx` + `scripts/measure-synthetic-stream.mjs`
so the harness can A/B both regimes without spending LLM credits.
See `scripts/profile-typing-lag.md` for the full investigation.
* perf(desktop): useDeferredValue for streaming markdown so parses don't block input
Streamdown's per-Block parse cost grows with the live tail's length and
is unavoidable inside the block-memo pattern (industry standard, see
findings doc). The fix is to stop having that work block the main thread.
`<DeferStreamingText>` is a 12-line wrapper that reads message-part state
via `useMessagePartText`, runs it through `useDeferredValue`, and
re-publishes via assistant-ui's `<TextMessagePartProvider>`. The inner
`<StreamdownTextPrimitive>` reads the deferred value through the normal
`useMessagePartText` hook — no fork, no internal-path imports, fully on
assistant-ui's public API. React's concurrent scheduler then:
- abandons in-flight deferred renders when a newer token arrives, so
intermediate states get skipped under fast streams
- deprioritises the markdown render when the main thread has urgent
work (typing, scroll), so input stays responsive even while a
100ms parse is queued
Streamdown already uses `useTransition` for its block-array setState;
this lifts the deferral up to the consumer boundary so it covers the
whole pipeline (preprocess → split → repair → parse → render).
A/B on the 34 MB session, 300 tokens at 50 tok/sec, markdown chunks
(four trials each, with the 33ms flush throttle on for both):
| | avgFps | p99 frame | LTs/5s | max LT | typing-while-stream p95 |
|---|---|---|---|---|---|
| pre | 54.3 | 41 ms | 1.7 | 110 ms | ~17 ms |
| post | 58.5 | 31 ms | 2.0 | 117 ms | 14-18 ms |
Longtask count + max LT unchanged — useDeferredValue doesn't reduce
CPU, only its priority. The avgFps lift and p99 frame drop are the
proof that the existing CPU is no longer blocking 60 fps cadence. One
clean run logged MUTATIONS=0 — React skipped every intermediate text
state and only committed the final one (textbook deferred-value
behaviour).
The actually-reduce-CPU path is replacing the parser with a state
machine like Flowdown — left for a future PR; see
`apps/desktop/scripts/profile-typing-lag.md` for the full investigation.
* feat(desktop): add hermes gui launcher
* feat(desktop): launch packaged gui builds by default
* bump gui version to 0.0.2
* fix(dashboard): allow file:// origin on loopback WS + diagnostic logging
Upstream commit 2e66eefbc ("fix(dashboard): validate WebSocket Host
and Origin") added a WebSocket Host/Origin guard to block DNS
rebinding against the dashboard. The guard rejects any Origin whose
scheme is not http/https or whose netloc is empty — which includes
Electron's renderer Origin: file:// when the desktop app loads its
bundle from disk in production mode.
That makes the bb/gui Electron desktop unable to open the gateway
WebSocket against the embedded backend on Windows / macOS prod
builds. The renderer reports "Desktop boot failed" and the backend
logs:
WARNING hermes_cli.web_server: gateway-ws reject
peer=127.0.0.1:NNNN reason=non_loopback_or_bad_origin
bound_host=127.0.0.1 close_code=4403
DNS-rebinding requires a DNS-resolvable hostname; file:// has no
host component and therefore cannot be the attack vector this guard
exists to block. When bound to a loopback interface (127.0.0.1 /
::1 / localhost), accept file:// origins so desktop wrappers can
attach. Non-loopback binds (operator opted into network exposure)
keep rejecting file:// — the loose policy doesn't apply.
Also adds per-reason diagnostic logging in
_ws_host_origin_is_allowed, so future ws-guard rejections name the
specific clause that fired (bad_host / bad_origin_scheme /
origin_host_mismatch) instead of the opaque
"non_loopback_or_bad_origin" surfaced at the call site.
Verified against tests/hermes_cli/test_web_server_host_header.py
(all 11 upstream tests still pass) and hand-tested by opening the
bb/gui Electron desktop dev build against the patched backend.
* fix(tui_gateway): restore _content_display_text helper
Bb/gui had dropped the helper but the orchestrator code merged from main
still calls it (_inflight_text, _message_preview). Re-add the definition
verbatim from main so session.create / _start_inflight_turn don't crash
with NameError on first prompt submit.
* fix(tui-gateway): restore _content_display_text helper lost in main merge
The May 27 merge of origin/main into bb/gui re-introduced two callers of
_content_display_text (in _inflight_text and _history_to_messages) but
dropped the helper definition itself, leaving an unresolved reference.
NameError fires on every user message via _start_inflight_turn ->
_inflight_text, taking down both the TUI and the desktop (which share
this gateway backend) the moment input is dispatched.
Restores the helper verbatim from main (commit 36c99af37) -- pure
structured-content text extractor, no other dependencies.
* fix(telegram): import Set for _dm_topic_chat_ids annotation
self._dm_topic_chat_ids: Set[str] = {...} at line 460 references Set
but only Dict, List, Optional, Any are imported from typing. The file
has no 'from __future__ import annotations', so the annotation is
evaluated at runtime and raises NameError on TelegramAdapter
construction.
* fix(setup): drop shadowing inner importlib.util re-imports
_print_setup_summary and _setup_tts_provider each had 'import
importlib.util' inside a try: block nested deeper in the function
body. Python flips importlib to function-local for the whole scope,
so earlier references in the same function (the neutts branches at
lines 493 / 1109) hit UnboundLocalError before the late import can
run.
The top-of-module 'import importlib.util' at line 14 already covers
both call sites, so dropping the redundant inner imports restores
the intended behavior.
* feat(install.ps1): add -IncludeDesktop switch + Stage-Desktop
The new Hermes-Setup.exe (Tauri bootstrap installer) passes -IncludeDesktop
so users who install via the GUI end up with a launchable Hermes.exe at
apps/desktop/release/<os>-unpacked/. Existing flows are unchanged:
* The 'irm install.ps1 | iex' CLI one-liner omits the flag — terminal
users don't need a prebuilt desktop binary; 'hermes desktop' builds
on demand.
* The Electron desktop's bootstrap-runner.cjs also omits the flag —
rebuilding apps/desktop from inside a running Hermes.exe would try
to overwrite the live binary on disk and fail.
Stage-Desktop runs after Stage-NodeDeps so workspace npm is already
installed when electron-builder fires. It does:
1. 'npm install' at repo root so apps/* workspaces resolve their deps
(Electron itself arrives via npm here, ~150MB)
2. 'npm run pack' in apps/desktop (tsc + vite + electron-builder --dir)
3. Probes apps/desktop/release/{win-unpacked,win-arm64-unpacked}/Hermes.exe
The --dir mode produces an unpacked launchable binary without an NSIS/MSI
installer artifact — we don't need one because Hermes-Setup.exe spawns the
unpacked binary directly via launch_hermes_desktop.
* feat(installer): Tauri bootstrap installer for first-time onboarding
Hermes-Setup.exe is a small signed Rust+Tauri binary that drives
scripts/install.ps1 stage-by-stage with a native UI matching the
desktop's design language. Replaces the chicken-and-egg pattern of
shipping a 200MB Electron app whose first launch existed only to
run install.ps1.
The architecture:
Rust backend (src-tauri/):
bootstrap.rs orchestrator -- Tauri commands, stage iteration
install_script.rs resolve install.ps1 (dev checkout, cache, GitHub raw)
powershell.rs spawn powershell, line-stream stdout/stderr, parse JSON
events.rs BootstrapEvent types -- mirror bootstrap-runner.cjs
paths.rs HERMES_HOME resolution + tracing log setup
build.rs bakes BUILD_PIN_COMMIT / BUILD_PIN_BRANCH from
'git rev-parse HEAD' at compile time
React frontend (src/):
Tauri webview rendering 4 screens (welcome / progress / success /
failure), driven by nanostores subscribing to the Rust event stream.
Visual layer reuses the desktop's styles.css wholesale via @import
so the installer and desktop never drift visually.
Distribution:
targets = ['app', 'dmg', 'appimage'] -- no NSIS/MSI wrapper. The
raw target/release/Hermes-Setup.exe IS the artifact on Windows;
.dmg + .app on macOS; AppImage on Linux. One file, double-click,
no installer-installing-an-installer pattern.
Compile-time pinning:
build.rs reads 'git rev-parse HEAD' and emits
cargo:rustc-env=BUILD_PIN_COMMIT=<sha> + BUILD_PIN_BRANCH=<branch>.
bootstrap.rs's option_env!() picks these up so the binary fetches
install.ps1 from the exact SHA it was tested against. CI / release
builds can override via HERMES_BUILD_PIN_COMMIT env var.
Windows manifest:
hermes-setup.manifest declares level='asInvoker' so the
productName 'Hermes Setup' doesn't trip Windows's installer-
detection heuristic and refuse to launch without elevation.
Also declares PerMonitorV2 DPI + UTF-8 active code page + Common
Controls v6.
Limitations of this initial version:
* No code signing -- Windows SmartScreen will warn once on Hermes-Setup.exe
('More info -> Run anyway'). The downstream binaries it produces
(Hermes.exe in win-unpacked/, the hermes CLI) are locally-built and
therefore don't carry MOTW, so they launch without SmartScreen
intervention. Cert procurement tracked separately.
* macOS and Linux build paths defined but untested -- Windows-only V1.
* fix(installer): pass -IncludeDesktop to manifest, surface launch errors, alias hermes desktop
Three bugs found in the first VM end-to-end test:
1. install.ps1 -Manifest was called WITHOUT -IncludeDesktop, so the
manifest came back with the 14-stage list (no desktop stage), the
UI showed '14 steps' and Stage-Desktop never ran. Pass the flag to
both the manifest fetch and the per-stage runs — install.ps1 gates
the desktop stage's inclusion on the flag.
2. The Success screen's Launch button silently swallowed the Tauri
error when no Hermes.exe existed (e.g. Stage-Desktop was skipped).
Wire the error through to inline UI with an alert callout, so the
user gets actionable text ('Hermes.exe missing, run hermes desktop
from a terminal') instead of an unresponsive button.
3. The Success screen tells users to run 'hermes desktop' from a
terminal but the CLI only accepted 'hermes gui' — invalid choice
for 'desktop'. Rename the subcommand canonically to 'desktop' with
'gui' as a backwards-compatible alias. Update the _SUBCOMMANDS sets
used by session-flag arg parsing + logging-mode probe so both names
route to the same logic.
* fix(install.ps1): pre-warm electron-builder winCodeSign cache + fix Stage-Desktop $HasNode false-skip
Two bugs caught in the second VM end-to-end run:
1. electron-builder's winCodeSign extraction fails on grandma-class
Windows boxes because the .7z archive contains macOS symlinks
(darwin/10.12/lib/libcrypto.dylib and libssl.dylib pointing at
versioned siblings). Creating symlinks on Windows requires
SeCreateSymbolicLinkPrivilege, a per-user right that non-admin
accounts don't have on stock Windows. Result: every fresh install
on a non-admin user fails Stage-Desktop with a 7-Zip 'cannot create
symbolic link' error, retried four times, then bails.
Fix: Initialize-ElectronBuilderCache pre-extracts winCodeSign-2.6.0.7z
ourselves with -snl (don't preserve symlinks, store as resolved file
content) AND -x!darwin (skip the entire macOS subtree — irrelevant
on Windows). Writes to electron-builder's expected cache dir before
electron-builder gets a chance to try its own broken extraction.
Idempotent — fast-paths via signtool.exe sentinel check.
2. Install-Desktop's first guard was 'if (-not $HasNode) skip'.
$HasNode is set by Stage-Node into $script:HasNode, but in
cross-process driver mode (each -Stage NAME is a fresh powershell.exe
spawned by Hermes-Setup.exe), that script-scope variable from the
PREVIOUS process is invisible — so the guard always fired and
Install-Desktop returned in 900ms with a misleading
'Node.js not available' reason. The real npm probe below it never
got to run. Fix: re-probe npm directly via Get-Command when $HasNode
is empty/false, since by that point Stage-Node has already verified
Node is installed and the only question is whether *this* process
can see it on PATH (it can — installer-wide PATH update from Stage-Node).
* fix(install.ps1): tell electron-builder we're NOT signing instead of pre-extracting winCodeSign
The previous commit (c7e46f9f3) worked around the winCodeSign-symlinks-
on-Windows extraction crash by pre-extracting the archive ourselves with
-snl + -x!darwin. That fix was correct but addressed the wrong layer.
The deeper question: why was electron-builder fetching winCodeSign at all
when we have no signing cert configured? Answer: electron-builder
unconditionally pre-warms the toolchain assuming any build MIGHT sign.
The cert auto-discovery never finds anything (we never set CSC_LINK
or anything else), so the signing never happens — but the 100MB fetch
of winCodeSign and its broken-on-Windows symlink extraction does.
Set CSC_IDENTITY_AUTO_DISCOVERY=false (with WIN_CSC_LINK and
WIN_CSC_KEY_PASSWORD also explicitly cleared as belt-and-suspenders)
before invoking npm run pack, and electron-builder skips the entire
winCodeSign apparatus. No download, no extraction, no privilege check.
Env vars are saved/restored around the invocation so we don't leak
the override into Stage-PlatformSdks etc.
Net: removes the 100-line Initialize-ElectronBuilderCache helper that
manually downloaded + extracted winCodeSign-2.6.0.7z. Replaced with
3 env-var assignments. The produced Hermes.exe is functionally
identical — just no longer carries a code-signing-machinery dependency
we never used.
* fix(installer): bump bootstrap-installer.log to capture stage transitions + every install.ps1 line
Diagnosing the second VM failure was impossible because bootstrap-installer.log
contained only the 'starting' banner. Two causes:
1. emit_log() inside run_bootstrap() was tracing::debug! — dropped on the
floor under the default INFO env-filter.
2. The per-stage sink callbacks (on_stdout_line / on_stderr_line) only
emitted Tauri events to the frontend; they never tee'd to the log file
at all. When the failure route mounts, the Tauri event stream is the
only place the script output lived, and it gets discarded.
3. The Failed / Stage / Manifest / Complete lifecycle frames in emit_event()
were also Tauri-only — so even the 'which stage failed' frame never
reached the log.
Fixes:
* emit_log() → tracing::info!
* Sink callbacks tee stdout to info!, stderr to warn!, with stage label
as a structured field for grep'ability
* emit_event() now matches on the variant and logs each lifecycle frame
at the right level: Failed → tracing::error!, others → info!
Result: a failing install leaves a complete forensic trail in
bootstrap-installer.log — manifest stage list, every install.ps1
stdout/stderr line tagged by stage, the stage transitions, and the
final error. Same path as before so nothing the user does changes.
* fix(install.ps1): Stage-NodeDeps cross-process $HasNode + stream npm install output to bootstrap log
VM run 3 diagnosis: node-deps stage skipped on the VM (logged
'Skipping Node.js dependencies (Node not installed)') and then
desktop's npm install failed with exit 1 and zero diagnostic detail.
Two root causes:
1. $HasNode false-skip in Stage-NodeDeps — same cross-process bug
pattern we fixed for Stage-Desktop in c7e46f9f3. Stage-Node ran
in process A and set $script:HasNode = $true, then exited. Stage-
NodeDeps ran in fresh process B (Hermes-Setup.exe -Stage NAME
spawns each stage independently), where that variable doesn't
exist. Re-probe via Get-Command npm instead of trusting the
stale script-scope global. The previous stage already verified
Node so the re-probe succeeds.
2. npm install --silent + Tee to TEMP file hid the real error.
When the workspace install failed on the VM, the actual reason
was buffered in $env:TEMP\hermes-npm-desktop-install-*.log and
the user saw only 'exit 1'. Drop --silent so npm streams its
full output, drop the TEMP-file dance — the Tauri installer's
streaming sink already tees every stdout/stderr line to the
rolling bootstrap-installer.log, so a side log file is dead
weight that hides the very error we need.
After this, the bootstrap log on a failure will contain npm's full
output (deprecation warnings, ETARGET, native-module compile errors,
whatever) tagged with stage=desktop, making the actual cause
diagnosable instead of an opaque exit code.
* fix(install.ps1): restore Initialize-ElectronBuilderCache (CSC env vars alone aren't enough)
VM run 4 diagnosis: even with CSC_IDENTITY_AUTO_DISCOVERY=false set,
electron-builder still fetches winCodeSign and signs bundled binaries.
The log shows the signing happens BEFORE the cache extraction:
• signing with signtool.exe ...\winpty-agent.exe
• signing with signtool.exe ...\OpenConsole.exe
• downloading winCodeSign-2.6.0.7z
• <symlink privilege error>
Cause: node-pty's bundled prebuilds are listed in apps/desktop's
asarUnpack ['**/*.node', '**/prebuilds/**']. electron-builder
re-signs anything unpacked from asar, regardless of whether OUR
binary gets signed. The signtool invocation needs winCodeSign on
disk, which needs the .7z extracted, which hits the macOS-symlink
crash on non-admin Windows.
The CSC env vars I added in d5fe46727 only kill IDENTITY DISCOVERY
(so OUR Hermes.exe stays unsigned, which is fine — we have no cert).
They don't prevent the toolchain fetch for the bundled-prebuild
re-sign. I removed the pre-extract in d5fe46727 thinking the env
vars subsumed it; that was wrong. Both are needed.
Restoring Initialize-ElectronBuilderCache verbatim from c7e46f9f3
and keeping the CSC env vars. Wrote a clearer doc-comment at the
call site explaining the two-knob interaction so future maintainers
don't drop one half again.
* fix(desktop): disable signtool via signtoolOptions.sign=null, drop dead winCodeSign pre-extract
VM run 5 diagnosis: the pre-extract from 3b29e65c1 ran (extracted 83
files, 24MB) but produced ZERO files at the expected sentinel path
'/winCodeSign-2.6.0/windows-10/x64/signtool.exe'.
Cause: the .7z archive's root entries are 'windows-10/', 'darwin/',
'linux/', etc. — not 'winCodeSign-2.6.0/<arch>'. Extracting with
'-o$cacheRoot' put files at $cacheRoot/windows-10/..., NOT at
$cacheRoot/winCodeSign-2.6.0/windows-10/.... I had the directory
nesting wrong from the start.
And then we observed: electron-builder downloads winCodeSign-2.6.0.7z
under a random numeric filename ('384387955.7z') regardless of what's
already extracted in the parent dir. The cache key isn't the dirname;
it's content-addressed. So the pre-extract approach was doomed even
if the path nesting had been right.
Actual fix: signtoolOptions.sign=null in apps/desktop/package.json's
win build config. electron-builder honors this and skips the bundled-
prebuild signing entirely — no signtool invocation, no winCodeSign
fetch, no symlink-privilege crash. The previous failures all stemmed
from electron-builder pre-signing node-pty's bundled .exes
(winpty-agent.exe, OpenConsole.exe) which are already author-signed
upstream; re-signing with our nonexistent cert was overwriting good
sigs with nothing useful anyway.
Cost: when we DO get a real cert later, we'll add it back with the
sign function pointing at the cert chain. Until then, all-null is
the correct config and unblocks every non-admin Windows user.
Removed Initialize-ElectronBuilderCache (the dead pre-extract).
Removed the call site. Kept the CSC_IDENTITY_AUTO_DISCOVERY env
vars as belt-and-suspenders against a future electron-builder
change that might revive cert auto-discovery.
* fix(desktop): use no-op sign function instead of sign=null
VM run 6 still hit the symlink crash even with signtoolOptions.sign=null.
electron-builder 26.8.1 treats null as 'use the default signtool path'
rather than 'skip signing', so the winCodeSign fetch + extraction still
fired for the bundled prebuild re-sign.
The Electron docs (electronjs.org/docs/latest/tutorial/code-signing)
make it clear signing is OPTIONAL and unsigned apps work fine — users
just see SmartScreen on first launch. The electron-builder mechanism
for 'don't actually sign anything' is to supply a custom sign function
(via signtoolOptions.sign: '<path-to-cjs-module>') that resolves
without invoking signtool.
build-noop-sign.cjs is that module — a 5-line async function that
returns undefined. electron-builder calls it for every binary it would
have signed, gets back a resolved promise, and considers each binary
'signed.' No signtool spawn, no winCodeSign fetch, no symlink crash.
When Nous's cert arrives, replace this file with a real signing hook
(@electron/windows-sign-based or a direct signtool invocation). The
architecture's signing-ready and the cutover is a one-file edit.
* fix(desktop): signAndEditExecutable=false to skip signtool path entirely
After reading app-builder-lib/winPackager.js line 216 + 231 directly:
signAndEditExecutable is the ACTUAL hardcoded gate that short-circuits
both signApp() (which signs Hermes.exe + every shouldSignFile match
including bundled prebuilds) AND createTransformerForExtraFiles().
None of signtoolOptions.sign / sign:null / sign:<custom-fn> gate the
winCodeSign download — that happens before they're consulted.
What we lose: rcedit also runs through signAndEditResources, so
disabling this drops PE metadata (file properties showing 'Hermes' /
'Nous Research' / file description). Cost is real but bounded:
* Hermes.exe filename, icon, asar contents, app identity intact
* Task Manager shows 'Hermes.exe' (the filename) not 'Hermes' (PE
description) — minor downgrade
* Start menu, taskbar, window title all work normally
* SmartScreen will warn once (unsigned, same as before)
When the cert lands, flip signAndEditExecutable back to default true,
both signing AND rcedit return, PE metadata is restored.
Removes the no-op sign function (build-noop-sign.cjs) since
signAndEditExecutable=false prevents signtool from being invoked at
all — the custom hook never gets called either.
* feat(install.ps1): write .hermes-bootstrap-complete marker at end of install
The desktop app's main.cjs resolver ladder has a 'bootstrap-needed' rung
that fires when .hermes-bootstrap-complete is missing from
ACTIVE_HERMES_ROOT. Pre-Hermes-Setup, this marker was written by the
packaged-desktop's own bootstrap-runner.cjs at the end of its install
flow. Now that Hermes-Setup.exe runs install.ps1 directly, install.ps1
needs to own the marker — otherwise the desktop sees no marker on first
launch and triggers its legacy first-launch bootstrap (re-running
install.ps1 from inside Electron, the exact recursion Hermes-Setup.exe
was supposed to obviate).
Implementation:
* New Stage-BootstrapMarker (worker) → Write-BootstrapMarker (helper)
* Slotted in the manifest right after platform-sdks, before the
interactive configure/gateway stages, so it runs unconditionally
when the install reaches the finalize phase
* Schema mirrors apps/desktop/electron/main.cjs writeBootstrapMarker /
isBootstrapComplete EXACTLY: {schemaVersion: 1, pinnedCommit,
pinnedBranch, completedAt}. Schema version stays at 1 so old
desktops that read marker files written by future install.ps1s
can still parse them.
* pinnedCommit comes from -Commit flag (Hermes-Setup.exe passes it)
or falls back to 'git rev-parse HEAD' in InstallDir
* pinnedBranch from -Branch flag, defaults to 'main' matching
install.ps1's own param default
Two PS-5.1 gotchas baked into comments:
* The ?. null-conditional operator doesn't exist pre-PS7; use
explicit if-checks on Get-Command results
* Set-Content -Encoding UTF8 emits a BOM in 5.1 and Node's plain
JSON.parse rejects BOM — write via .NET's UTF8Encoding(false)
to produce BOM-less JSON the desktop's readJson() can parse
* feat(installer): drive in-app updates through the Tauri installer
Converge update on the same principle as bootstrap: one driver owns all
repo mutation. The desktop becomes a pure consumer that hands off to
Hermes-Setup.exe --update instead of re-implementing git/pip in Electron.
- hermes desktop --build-only: build without launching, so the installer
owns the post-update launch (CLI keeps build logic single-sourced).
- Installer AppMode {Install,Update} from argv; get_mode exposed to the UI.
- Installer self-copies to HERMES_HOME/hermes-setup.exe on install success
(no-op guard during --update re-invocation to avoid the locked-exe copy).
- Installer --update flow (update.rs): wait for the desktop to release the
venv shim, run 'hermes update --yes --gateway' (branch on exit 0/2/other),
then 'hermes desktop --build-only', then launch the rebuilt desktop. Reuses
the bootstrap event channel + progress UI via a synthetic two-stage manifest.
- Desktop applyUpdates() gutted (~105 lines of git/stash/pull/pyproject/pip
removed) -> thin handoff: spawn updater, app.quit() to free the shim.
Detection (checkUpdates, commit changelog, behind-count) kept intact.
- install.ps1 creates Start Menu + Desktop shortcuts to the packed Hermes.exe
(never bare 'hermes desktop', which would rebuild every launch).
* test update
* fix(installer): pass --branch to hermes update in the --update flow
The install is a detached-HEAD checkout of a pinned commit. Without
--branch, 'hermes update' fell back to its default (main) and switched
the checkout to main — a divergent branch that lacks the desktop CLI
command — so the update targeted the wrong branch and the rebuild stage
failed with 'invalid choice: desktop'.
Thread BUILD_PIN_BRANCH (the branch this installer was built against,
and the same branch the desktop detected the update on) into
'hermes update --branch <b>' so update + rebuild stay on-branch.
* test update
* fix(installer): stamp Hermes icon onto Hermes.exe via rcedit (no winCodeSign)
The unpacked Hermes.exe showed the stock Electron icon + name in the
taskbar because build.win.signAndEditExecutable=false disables BOTH
electron-builder's signing AND its rcedit metadata/icon stamping. That
flag is load-bearing: enabling it re-triggers signtool -> winCodeSign,
whose macOS symlinks crash 7-Zip on non-admin Windows (unfixable dead end).
Decouple identity-stamping from signing entirely: after npm run pack,
run rcedit ourselves on the produced exe.
- Add rcedit as a direct devDependency of apps/desktop (the transitive
electron-winstaller copy is fragile).
- apps/desktop/scripts/set-exe-identity.cjs: Node helper that calls
rcedit's named export to set icon + ProductName/FileDescription/
CompanyName. Node builds argv natively — avoids the PowerShell->exe
->JSON double-escaping that broke the app-builder rcedit path.
- install.ps1 Set-DesktopExeIdentity invokes the script after the build,
before shortcuts. Best-effort: failure keeps the stock icon, never
fails the install. rcedit is a pure PE editor — no signtool, no
winCodeSign, no symlinks.
Verified locally: stamping a copy of the built Hermes.exe embeds the
32x32 icon and sets ProductName=Hermes.
Also fix update-path success-screen flash: in update mode the installer
hands off + exits in ~600ms, so don't route to the 'launch Hermes'
success view (it flashed before the window closed).
* update test
* fix(desktop): show 'hermes update' guidance for CLI installs instead of dead-end error
A user who installed via the CLI (irm|iex / install.sh) then ran
`hermes desktop` has no staged hermes-setup.exe, so clicking Update
in-app hit resolveUpdaterBinary()=null and showed a misleading error
('re-run the Hermes installer') with a Try-again button that could
never succeed — a dead loop for a perfectly valid install.
Treat the no-updater case as an intentional outcome, not a failure:
- main.cjs applyUpdates returns { ok:true, manual:true, command:'hermes update' }
(no throw, no 'error' stage) when no updater binary exists.
- New 'manual' update stage + apply-state.command thread the command to the UI.
- updates-overlay ManualView: a polished terminal-native card with the
exact command and a copy button, framed as the correct path for a CLI
user rather than an error.
GUI-installer users are unaffected — hermes-setup.exe present => seamless
auto-update runs as before. Zero new process orchestration; can't fail
the update demo.
* update test
* fix(gui): pin /api/hermes/update to the current branch
The desktop command-center 'update' action hits POST /api/hermes/update,
which spawned bare `hermes update` with no --branch. cmd_update then
falls back to its default (main) and checks the working tree OUT of the
tracked branch — a bb/gui install silently jumped to main and lost the
desktop CLI.
Resolve the checkout's current branch and pass --branch <current> from
this endpoint only. The engine default (main) is DELIBERATELY unchanged:
bare `hermes update` from a terminal, the gateway /update bot command,
and the CLI/TUI relaunch path all keep their long-standing 'update against
main' contract for the existing user base. Only the GUI button is scoped
to update-the-branch-you're-on. Detached HEAD / git failure falls back to
the bare default.
* update test
* fix(desktop): branch-pin the CLI manual-update command card
The 'Update from your terminal' card (shown to CLI installs with no staged
updater) hardcoded bare `hermes update` — which defaults to main and would
switch a bb/gui (or any non-main) checkout off-branch. Same bug we fixed for
the GUI button, leaked into the card's copy text.
Resolve the checkout's current branch and show `hermes update --branch
<current>` for non-main checkouts; keep it bare for main so the card stays
clean. Best-effort: bare fallback if branch detection fails. Matches the
GUI button + installer --update contract; bare terminal/bot/TUI update
paths still default to main, unchanged.
* docs: phragg was here
* feat(desktop): lead onboarding with Nous Portal + fix fresh-install detection (#34970)
- Feature Nous Portal as the primary onboarding card (Recommended tag,
app logo, single pitch line); collapse other OAuth providers behind an
"Other providers" disclosure whose open/closed state persists.
- Surface OpenRouter as a one-click API-key option inside the disclosure;
move "I have an API key" to a quiet bottom-right link.
- Treat "no provider configured" as a normal onboarding state, not a red
error banner (provider-setup-errors copy match).
- Fix setup.runtime_check: it reported ready when the resolved runtime had
an empty credential or only implicit Bedrock/IAM, so fresh installs never
saw onboarding. Now requires a usable credential.
- Auto-wire Windows fonts for WSL2 users so the renderer renders real
Segoe UI instead of the DejaVu fallback; make WSL detection env-independent
via the /proc kernel marker.
* feat(desktop): live elapsed timer on install bootstrap steps
The first-launch install overlay showed a static "Installing" with no
motion, so long steps (notably the repo clone) looked frozen. Stamp each
stage's start time on the running transition and tick once a second so the
active step shows live elapsed (e.g. "Installing · 1:23"), plus elapsed on
the overall current-step line. Completed steps keep their final duration.
* fix(desktop): resolve PortableGit for update checks + reserve titlebar tools space
- runGit() hardcoded spawn('git'), which ENOENTs on fresh installer-driven
Windows installs (git is PortableGit under %LOCALAPPDATA%\hermes\git, never
on PATH) — so "Check for updates" failed with "Couldn't check for updates".
Add resolveGitBinary() mirroring findGitBash (PortableGit → Git-for-Windows
→ PATH) and use it in runGit.
- PageSearchShell rendered a full-width search input in the titlebar row, so
on Windows its right edge slid under the fixed top-right tools + native
window controls. Reserve that footprint via --titlebar-tools-* vars.
* fix(desktop): stop streaming caret from shifting layout on completion
The streaming caret (::after on the running message's last child) was an
in-flow inline-block adding ~0.78em of inline width, which could wrap the
last line mid-stream; when the caret is removed on completion the line
un-wraps and reflows — the visible post-response layout shift. Net-zero its
inline advance with a compensating negative margin so it paints at the text
end without consuming layout width.
* fix(desktop): stop completed-message layout shift while streaming
The assistant message action bar used `hideWhenRunning`, which unmounts it
whenever the thread is streaming. Since the bar reserves vertical space in
each completed assistant message's footer (it's invisible-until-hover via
opacity, not via mount), unmounting it collapsed every prior turn by the
bar's height — then remounting on resolve grew them back, shifting the whole
conversation (visible as "padding appears above the last user message").
Drop hideWhenRunning so the footer height is constant; the bar stays
invisible during streaming via its existing opacity/pointer-events gating.
* fix(merge): keep windows-footgun suppressions inline
* fix(merge): keep remaining gateway footgun suppressions inline
* fix(merge): restore contracts caught by main-target CI
* fix(dashboard): honor injected HERMES_DASHBOARD_SESSION_TOKEN
The desktop shell mints a session token and signs its /api + /api/ws
calls with it via HERMES_DASHBOARD_SESSION_TOKEN, but the main-merge
restored a web_server.py that ignored the env var and minted its own
random _SESSION_TOKEN -- so every desktop request 401'd and the UI
reported "gateway offline". Read the injected token (fall back to a
fresh random one) so loopback HTTP + WS auth line up.
Adds a regression test so a future merge can't silently drop the read.
* fix(desktop): align fresh-install home so upgraders don't brick
Two related first-launch bugs on machines with a legacy ~/.hermes:
- install.ps1 hardcoded $HermesHome/$InstallDir to %LOCALAPPDATA%\hermes
and ignored the HERMES_HOME the desktop passes through. The desktop
freezes HERMES_HOME at module load and prefers a legacy ~/.hermes when
%LOCALAPPDATA%\hermes is absent, so the installer wrote to a different
home than the shell read -> "Could not connect to Hermes gateway". Honor
$env:HERMES_HOME in the param defaults.
- isBootstrapComplete() trusted the marker + checkout without verifying a
runnable venv, so an interrupted/split install spawned a dead backend
instead of re-bootstrapping. Also require the venv python to exist.
* fix(dashboard): allow packaged desktop file:// origin on loopback WS
The packaged Electron desktop loads its renderer over file://, so its
/api/ws handshake carries Origin: file:// (or null). The DNS-rebinding
WebSocket Origin guard only accepted http(s) origins matching the bound
host, so it rejected the desktop's own renderer with 4403 -> "Could not
connect to Hermes gateway" on macOS.
A browser DNS-rebinding attacker can only ever present an http(s) origin
(the site hosting the malicious page); it cannot forge file://, null, or
a custom app scheme AND hold the loopback session token. So on loopback
binds we now trust non-web origins -- the token in _ws_auth_ok remains
the real authenticator. Public/gated binds still reject them, and
cross-site http(s) origins are still rejected everywhere.
* fix(desktop): resolve renderer assets relative to BASE_URL
Absolute public asset paths (/apple-touch-icon.png, /ds-assets/...) work
under the dev server but break in the packaged app, where the renderer is
loaded from file://.../index.html and a leading slash resolves to the
filesystem root -> broken onboarding provider icon and backdrop image on
macOS. Prefix these with import.meta.env.BASE_URL so they resolve next to
the bundled index.html in both dev and packaged builds.
* feat(desktop): automate first-launch bootstrap on macOS/Linux
Previously a packaged macOS/Linux app with no Hermes install hit a
dead-end ("first-launch install is not yet automated -- run install.sh
manually") because install.sh lacked the staged protocol install.ps1
exposes. Now both platforms bootstrap on first launch with the same
structured, per-step progress UI as Windows.
- install.sh: add --manifest / --stage / --json / --non-interactive plus
a stage dispatcher (prerequisites, repository, venv, python-deps,
node-deps, path, config, setup, gateway, complete). User-input stages
(setup, gateway) are skipped under --non-interactive; the in-app
onboarding overlay owns API keys/model, matching the Windows flow.
Each stage runs inside the install dir (its own process) and a new
--commit flag pins the checkout to the build-stamp SHA.
- bootstrap-runner.cjs: drive the staged manifest/stage/JSON protocol for
both install.ps1 (PowerShell) and install.sh (bash), selected by
installer kind; removed the single-blob POSIX shim.
- main.cjs: drop the macOS/Linux unsupported-platform dead-end so the
bootstrap-needed path runs the installer on every platform.
* fix(dashboard): return 404 JSON for unmatched /api paths instead of SPA HTML
The SPA catch-all (serve_spa) served index.html for any unmatched GET,
including unregistered /api/* endpoints. A missing API route therefore
came back as <!doctype html> with status 200, and JSON clients (the
desktop app's fetchJson) crashed with an opaque
'SyntaxError: Unexpected token <' instead of a clear error.
- web_server.py: unmatched /api or /api/... now returns 404 JSON
('No such API endpoint'); non-api paths still serve the SPA for
client-side routing.
- main.cjs fetchJson: detect an HTML body / text/html content-type on a
2xx response and reject with a clear message naming the URL, rather
than a raw JSON.parse SyntaxError. Empty bodies resolve to null;
malformed JSON reports the URL plus a snippet.
* say 'OS appearance' instead of 'macOS appearance'
* feat(install): add --include-desktop stage + PowerShell-style flags to install.sh
Brings install.sh to parity with install.ps1's bootstrap surface so the
shared Rust/Tauri bootstrapper (apps/bootstrap-installer) can drive a
macOS/Linux install the same way it drives Windows.
- Accept the PowerShell-style aliases the bootstrapper emits to both
installers: -Commit / -Branch (alongside existing -Manifest / -Stage /
-Json / -NonInteractive).
- Add --include-desktop / -IncludeDesktop. When set, the manifest gains a
'desktop' stage (immediately before 'complete'), and a new install_desktop
runs a root workspace `npm install` + `npm run pack` (electron-builder
--dir, signing auto-discovery disabled) to produce release/mac*/Hermes.app
-- mirroring install.ps1's Install-Desktop / Stage-Desktop.
- The flag is opt-in, exactly like Windows: the signed bootstrap installer
passes it; the Electron app's own first-launch bootstrap and the CLI
one-liner omit it (building the desktop from inside the running app would
clobber it).
* fix: tts endpoints
* macOS desktop: install + in-app self-update (#35607)
* fix(installer): align macOS HERMES_HOME with the rest of the stack
paths.rs computed the macOS Hermes home as ~/Library/Application Support/
hermes, but nothing else does: hermes_constants.get_hermes_home() (Python),
scripts/install.sh, and the Electron desktop's resolveHermesHome() all use
~/.hermes on macOS. The drift meant the Tauri installer wrote the install to
one directory and the desktop looked for it in another, so a fresh GUI
install never found its backend (the file's own comment warned this exact
drift would break things). Use ~/.hermes on macOS to match.
* fix(install.sh): always emit a stage result frame on failure
Stage helpers (clone_repo, install_deps, check_python, …) were written for
the monolithic flow and call `exit 1` on failure. Under `--stage`, that
terminated the process before the JSON result frame was printed, so the
installer's parse_stage_result saw "no frame" instead of a clean
{ok:false,...} contract response. Run the stage body in a subshell so an
`exit` only unwinds the subshell and the parent still emits the frame.
* feat(install.sh): auto-provision git on macOS/Linux (parity with install.ps1)
install.ps1 downloads PortableGit on Windows, but install.sh just printed a
"please install git" hint and exited — so a fresh Mac with no developer tools
(no Xcode CLT → no git) couldn't get past the clone step. check_git now tries
to install git before bailing:
- macOS: Homebrew if present (headless), else `xcode-select --install`
(the CLT prompt also provides the compiler some wheels need), polling for
git to appear.
- Linux: apt/dnf/pacman via sudo when available.
Falls back to the manual instructions only if auto-provision fails.
* feat(desktop): in-app GUI+backend self-update on macOS/Linux
On Windows the staged Hermes-Setup binary drives updates (quit → hermes
update → hermes desktop --build-only → relaunch). The mac drag-install has no
such binary, so "Update now" previously just printed `hermes update`.
Since there's no venv-shim file lock on POSIX, the desktop can drive the whole
update itself. applyUpdates now, when no staged updater exists on mac/linux:
1. runs `hermes update --yes [--branch <current>]` (backend git pull + deps),
2. runs `hermes desktop --build-only` (OS-aware GUI rebuild) with the
Hermes-managed Node + venv on PATH,
3. spawns a detached swapper that waits for this process to exit, dittos the
freshly built Hermes.app over the running bundle, clears quarantine, and
relaunches.
Degrades to "backend updated — restart to load the new GUI" if the rebuild
fails or there's no .app bundle to swap (dev run, Linux AppImage).
* chore: uptick
* chore: uptick
* chore: linux build
* fix(install): detect xcode-select git stub on fresh macOS
* chore: bump
* fix(desktop): repair voice dictation on Windows
Voice dictation was broken on Windows in two ways:
1. Mic access was denied. The Electron permission request handler only
granted 'media' requests whose details.mediaTypes included 'audio',
but Chromium on Windows frequently fires the mic request with an empty
mediaTypes array, so getUserMedia threw NotAllowedError. The handler
now grants audio-capture when mediaTypes includes 'audio' OR is
empty/absent, handles the 'audioCapture' permission name, and adds a
setPermissionCheckHandler (the synchronous path Chromium also consults
for getUserMedia on Windows). Video is still denied.
2. Transcripts went nowhere. The composer's insertText handler (used by
dictation and other inserts) only updated the assistant-ui composer
store via setText, never the contentEditable editor DOM. The
draft->editor sync effect only re-renders the editor when it is NOT
focused, and dictation runs while the editor has/regains focus, so the
transcript was stored but never shown and could not be sent. insertText
now renders into the editor DOM and places the caret, mirroring
appendExternalText.
Also hardens fetchJson: a 2xx response with an HTML body (or text/html
content-type) now rejects with a clear message naming the URL instead of
an opaque JSON.parse 'Unexpected token <' error.
* feat(desktop): route Nous subscribers onto the Tool Gateway from the GUI
When the GUI sets the main provider to Nous via POST /api/model/set, call
the same apply_nous_managed_defaults the CLI uses after model selection, so
GUI/onboarding users land on the Nous Tool Gateway the same way CLI users do
— no separate prompt, no duplicated logic.
Purely additive: apply_nous_managed_defaults skips any tool where the user
has a direct key (FIRECRAWL_API_KEY, FAL_KEY, etc.) or explicit config, so it
never overwrites a user's own setup. Only unconfigured tools get routed.
- web_server.py: in set_model_assignment (scope=main, provider=nous), resolve
enabled toolsets and apply managed defaults; guarded so a Portal hiccup never
blocks saving the model. Returns routed tools as gateway_tools.
- onboarding.ts: surface a 'Tool Gateway enabled' toast listing routed tools.
- types/hermes.ts: add gateway_tools to ModelAssignmentResponse.
- tests: cover nous-applies, non-nous-skips, and failure-doesnt-block-save.
* feat(desktop): mirror hermes model free/paid curation in GUI onboarding
GUI onboarding picked models[0] from /api/model/options, which ignores the
Nous free/paid tier — a free user could land on a paid default (e.g.
anthropic/claude-opus-4). Now the recommended default mirrors what `hermes
model` does.
- web_server.py: new GET /api/model/recommended-default?provider=<slug>. For
Nous it runs the same curation as the CLI (get_curated_nous_model_ids +
pricing + check_nous_free_tier + union_with_portal_{free,paid}_recommendations
+ partition_nous_models_by_tier) so free users get a free model and paid users
get the curated default. Other providers fall back to the first curated model.
Never 500s — returns empty model on error so onboarding degrades gracefully.
- hermes.ts: getRecommendedDefaultModel client + RecommendedDefaultModel type.
- onboarding.ts: fetchProviderDefaultModel prefers the recommended endpoint,
falls back to models[0] when unavailable.
- tests: free-tier picks free model, paid-tier picks curated default, failure
returns empty without 500.
* feat(desktop): show model pricing + free/paid tier gating in GUI picker
The CLI `hermes model` picker shows per-model $/Mtok pricing and gates paid
models on free Nous accounts. The GUI picker showed bare model names. Bring it
to parity across both the model-picker dialog and onboarding confirm card.
Backend:
- inventory.build_models_payload gains a pricing=True flag → _apply_pricing
enriches each provider row with formatted per-model pricing
({input,output,cache,free}) via the same _format_price_per_mtok the CLI uses,
and for Nous adds free_tier + unavailable_models (paid models a free user
can't select) via check_nous_free_tier + partition_nous_models_by_tier.
Best-effort: any pricing/tier failure is swallowed and fails open (no gating).
- /api/model/options and TUI model.options now pass pricing=True so the
global picker and in-session picker both carry pricing.
Frontend:
- ModelOptionProvider gains pricing/free_tier/unavailable_models; new
ModelPricing type.
- model-picker dialog renders In/Out $/Mtok (or a Free pill) per model, a
Free tier/Pro badge on the Nous heading, and disables + grays unavailable
paid models for free users with a 'Pro models need a paid subscription' note.
- onboarding confirm card shows the chosen model's price + tier badge.
Tests: test_inventory_pricing covers price formatting, free-tier gating,
paid no-gating, providers without pricing, and swallowed failures.
* fix(desktop): GUI model picker shows curated Nous list in curated order
Two bugs made the GUI Nous model list diverge from the `hermes model` CLI picker:
1. Backend (model_switch.py): the Nous row in list_authenticated_providers
fell through to cached_provider_model_ids("nous"), dumping the full live
/v1/models catalog (~50 vendor-prefixed models, alphabetical). Now it uses
the curated list AND applies the Portal free/paid recommendation union —
exactly like _model_flow_nous in main.py — so newly-launched models such as
stepfun/step-3.7-flash:free surface in curated order. Best-effort: falls
back to the curated list alone if the Portal fetch fails.
2. Frontend (model-picker.tsx): cmdk's Command had shouldFilter on (default),
which re-sorts items by fuzzy-match score (≈alphabetical) and ignores array
order. Set shouldFilter={false} + own the search term and do an
order-preserving substring filter, so the backend's curated order is shown
verbatim.
* feat(desktop): add/switch providers from the model picker via onboarding reuse
The model picker could only select models from already-authenticated
providers. Switching to a new provider had no in-app path. Rather than
duplicate provider UI, reuse the existing onboarding provider selector
(featured Nous + other providers + API-key form + device-code/PKCE flow +
model-confirm with pricing/tier).
- onboarding store: add a 'manual' flag with startManualOnboarding() /
closeManualOnboarding(). Manual mode forces the onboarding overlay to show
even when configured===true and refreshOnboarding no longer auto-dismisses
on runtime-ready (the app is already working — the user is just adding or
switching a provider).
- onboarding overlay: render when manual even if configured; show a Close
button (the first-run flow has none since the app can't run yet).
- model picker: 'Add provider' footer button opens the onboarding selector;
ModelResults lists only configured (model-bearing) providers.
* feat(desktop): add PUT /api/tools/toolsets/{name} enable/disable endpoint
* feat(desktop): add toggleToolset RPC binding
* feat(desktop): toolset enable/disable switch in Tools settings
* feat(desktop): tool configuration parity in GUI Tools settings
Bring the desktop GUI Tools settings to parity with the CLI `hermes tools`
for provider selection and API-key configuration.
Backend (hermes_cli/web_server.py):
- GET /api/tools/toolsets/{name}/config - provider matrix + key status
- PUT /api/tools/toolsets/{name}/provider - persist provider selection
Shared core (hermes_cli/tools_config.py):
- Extract apply_provider_selection / _write_provider_config from the
interactive _configure_provider so the CLI and GUI write identical
config keys (web.backend, tts.provider, browser.cloud_provider, plugin
image/video providers, use_gateway flags) through one code path.
Desktop UI:
- ToolsetConfigPanel: provider list with select, per-provider API-key
entry (set/replace/clear/reveal via the shared env RPCs), Ready/Needs
keys state, guidance for Nous-auth and post-setup providers.
- Wire the Configured/Needs keys pill to expand the panel inline; refresh
the toolset list after key changes so the pill updates live.
- Add getToolsetConfig / selectToolsetProvider RPC bindings + types.
Post-setup (OAuth/install) flows still defer to the CLI; see
docs spike findings for the planned /api/tools/setup/* endpoint family.
Tests: backend round-trip + 400 cases for the new endpoints and
apply_provider_selection; desktop vitest coverage for the config panel
(provider render, select, key save). No change-detector tests.
Also removes three stale completed plan docs.
* fix(desktop): show real Hermes version + sync package.json on release
The desktop app version was disconnected from the Hermes version: the
release script bumped pyproject.toml + hermes_cli/__init__.py but never
touched apps/desktop/package.json, which sat stale at 0.0.2 (lockfile at
0.0.1).
- main.cjs: hermes:version IPC now resolves __version__ from
hermes_cli/__init__.py (the canonical source release.py bumps) via a new
resolveHermesVersion() helper, falling back to app.getVersion() when the
source tree isn't readable. The About panel now always shows the live
Hermes version and can't drift.
- release.py: update_version_files() also bumps apps/desktop/package.json
in lockstep with pyproject (top-level version only; dep specs untouched).
- One-time catch-up: package.json 0.0.2 -> 0.15.1 and the lockfile root
mirrors 0.0.1 -> 0.15.1.
* fix(desktop): stamp exe identity in afterPack hook so updates stay branded
The packed Hermes.exe reverted to the stock Electron icon + "Electron" name
after an in-app update. The icon/identity stamp (rcedit) lived only in
install.ps1, but the installer's --update path rebuilds the desktop via
`hermes desktop --build-only` -> `npm run pack`, which never ran install.ps1
and so never stamped the rebuilt exe.
Move the stamp into an electron-builder afterPack hook so it runs for EVERY
packed build regardless of caller (first install, hermes desktop, the update
rebuild, or a manual npm run pack):
- set-exe-identity.cjs: refactor to export stampExeIdentity(exe, desktopRoot);
still runnable as a standalone CLI.
- after-pack.cjs (new): afterPack hook calling stampExeIdentity. Windows-only
guard; best-effort (logs + resolves on failure, never fails the build).
- package.json: register build.afterPack.
- install.ps1: remove the now-redundant Set-DesktopExeIdentity function + call;
the hook handles it during npm run pack.
electron-builder's own rcedit step stays disabled (signAndEditExecutable=false)
to avoid the signtool -> winCodeSign -> 7-Zip macOS-symlink crash on non-admin
Windows; the hook runs rcedit directly (pure PE resource edit, no signing).
* fix(desktop): export afterPack hook as exports.default so electron-builder runs it
The afterPack hook used `module.exports = fn`, which electron-builder's hook
loader doesn't pick up — it expects the function as the module's default
export (the same shape afterSign/notarize.cjs uses). The hook silently never
ran, so even first install shipped the stock "Electron" exe.
Switch to `exports.default = async function afterPack(...)`. Verified with a
real `npm run pack`: electron-builder now invokes the hook and the produced
release/win-unpacked/Hermes.exe carries ProductName/FileDescription=Hermes.
* chore(desktop): drop auto-build release CI in favor of manual build + upload
Remove desktop-release.yml (nightly-on-main + stable publish). Installers
are now built locally per platform and uploaded to a GitHub Release by hand;
the website points at them via NEXT_PUBLIC_HERMES_DL_* env. Update README +
docs and drop the dead desktop-nightly channel links.
* fix(desktop): stable shortcut icon + bust icon cache so updates repaint
Symptom on a freshly-installed laptop: Hermes.exe itself shows the correct
Hermes icon (Explorer reads the live exe's stamped PE resource), but the
desktop shortcut still draws the stock Electron icon.
Cause: New-DesktopShortcuts set IconLocation to "<exe>,0", so Windows cached
the icon it extracted from the exe at shortcut-creation time. On an update the
exe gets re-stamped, but the shortcut keeps rendering the stale cached bitmap.
- package.json: ship assets/icon.ico beside the exe via extraResources
(-> resources/icon.ico). Verified with a real npm run pack.
- install.ps1 New-DesktopShortcuts: point IconLocation at resources/icon.ico
(fallback to <exe>,0 if absent) — a dedicated .ico is cache-stable and skips
the per-exe extraction that goes stale. Then run `ie4uinit.exe -show` to bust
the shell icon cache so the shortcut repaints immediately instead of showing
the old Electron icon until reboot.
Both best-effort; never fail an otherwise-good install.
* dummy update
* feat(desktop): self-heal update branch + backend contract guard
Two fixes for the bb/gui→main transition:
- Self-update self-heals: if the tracked branch (e.g. bb/gui) no longer
exists on origin (merged + deleted), the desktop updater falls back to
main and persists it. Read-only ls-remote probe that only flips on a
definitive "ref absent" (exit 2), never on a transient network error, so
already-installed clients migrate themselves with no manual flip.
- Backend contract guard: tui_gateway reports DESKTOP_BACKEND_CONTRACT in
session runtime info; the desktop warns with a one-click "Update Hermes"
when the backend predates the GUI's required contract (e.g. a bb/gui app
pointed at a main checkout) instead of failing cryptically downstream.
* docs(desktop): rewrite README to match current install/update/build flow
The old README contradicted itself (claimed a bundled Python payload while
also saying it no longer bundles source) and predated cross-platform support.
Rewrite for accuracy: Linux is a first-class build target, install.sh/install.ps1
both drive the staged bootstrap, the real self-update handoff (Windows
Hermes-Setup vs in-app macOS/Linux), and the bb/gui→main self-heal + backend
contract guard.
* docs(desktop): rewrite README as a real product readme
Lead with what the app is and how to get it (download an installer, or
`hermes desktop` for existing CLI users) plus a plain-language feature list,
then keep contributor/build/internals as a clearly separated secondary section.
* docs(desktop): fix install framing — releases no longer auto-build installers
Lead with the install-with-Hermes path (`--include-desktop` / `hermes desktop`),
which always works, and describe prebuilt installers as manually published when
a release ships them rather than implying CI attaches them to every release.
* docs(desktop): match base repo README style
Adopt the root README's conventions: centered title + badge row, bold
one-liner intro, a feature <table> grid, --- section dividers, and a
Community / License footer.
* feat(desktop): recover from gateway boot failures + validate API keys on entry (#35864)
Fresh installs that hit a gateway boot failure had no recovery path: the
shell rendered dead ("gateway offline"), logs were undiscoverable, and a
mistyped API key was accepted because onboarding only checked credential
presence, not validity.
- Add BootFailureOverlay: a top-level recovery surface (Retry, Repair
install, Use local gateway, Open logs + inline recent logs) that mounts
on any hard boot failure, including post-install. Trims the now-redundant
recovery button from the onboarding Preparing panel.
- Add hermes:logs:reveal / :recent IPC (reveal desktop.log) and a
hermes:bootstrap:repair IPC that drops the bootstrap marker to force a
clean reinstall. Surface "Open logs" in Gateway settings too.
- Add POST /api/providers/validate: a live per-provider probe
(OpenRouter/OpenAI/xAI/Gemini key check, local endpoint connectivity)
wired into saveOnboardingApiKey so a rejected key blocks before it's
persisted, while an unreachable probe falls through (offline-safe).
* test(model-catalog): fix stale nous picker test after curated-list change
ac2e48907 made the GUI/picker Nous row use the curated list (curated["nous"]
= get_curated_nous_model_ids()) + Portal union, matching the `hermes model`
CLI — but test_picker_nous_row_uses_manifest still asserted the old 2-model
manifest snapshot, breaking the test shard.
Rewrite it as an invariant: stub the Portal union to passthrough and assert the
row equals get_curated_nous_model_ids() computed under the same conditions, so
it tracks the real contract instead of a hardcoded model list that rots on every
catalog update.
---------
Co-authored-by: emozilla <emozilla@nousresearch.com>
Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com>
Co-authored-by: Austin Pickett <pickett.austin@gmail.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: ethernet <arilotter@gmail.com>
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
PR #35718 added a per-slot "cumulative-resend" latch to the universal
streaming tool-call accumulator to fix DeepSeek / Baidu Qianfan (#35592).
The latch fires when a delta is a strict superset of the accumulated
buffer (len(_new) > len(_prev) and _new.startswith(_prev)) and then
REPLACES the buffer instead of appending.
That superset test is not an unambiguous cumulative signature. A normal
incremental stream can emit a single fragment that restates an already-
accumulated prefix — trivially common in large code-patch arguments with
repeated lines / indentation — which trips the latch and clobbers the
accumulated buffer, corrupting the tool call. Observed in the wild on
Anthropic Opus (the primary model) building a large patch: corrupted /
short arguments → finish_reason='length' dead-end → session killed.
A guessing heuristic that can silently clobber a tool-call buffer has no
place on the path every provider and model shares. Reverting restores the
known-good plain `+=` accumulator. The #35592 narrow provider bug should
be re-addressed provider-gated so it is structurally impossible to touch
Anthropic / OpenAI incremental streams, rather than via a heuristic on the
shared path.
Reverts ca03486b6.
DeepSeek / Baidu Qianfan stream tool-call arguments in cumulative mode:
each chunk resends the full arguments-so-far instead of the new fragment.
The stream accumulator blindly concatenated arg deltas with +=, turning
that into '{...}{...}{...}', which failed json.loads and got nuked to '{}'
— a silently corrupted tool call (#35592). Worse on multi-param tools
(search_files, session_search, memory replace) because longer args take
more chunks, giving more resend opportunities.
- Per-slot cumulative latch in the stream accumulator: a delta that is a
strict superset of the accumulated buffer marks the slot cumulative and
replaces (not appends); exact duplicates are dropped only after latching.
Incremental fragments are untouched (default += path).
- Backstop _collapse_repeated_json_arguments() in the repair pipeline
collapses pure identical-resend buffers (K exact repeats of a valid-JSON
unit) for providers that resend the complete object from chunk 1. Only
reached after json.loads already failed, so compliant single objects are
never touched.
Not a gateway or DeepSeek-model bug — any OpenAI-wire provider in
cumulative streaming mode is affected.
Resize vision tool-result images down to a 4 MB embed cap at load time,
not just at the 20 MB hard ceiling. A 5-20 MB image previously sailed
through the native fast path and got baked into conversation history,
where Anthropic's 5 MB per-image base64 limit rejected every subsequent
turn with a 400 — and because history is immutable, retries could never
clear it, permanently wedging the session.
Also harden the reactive shrink-recovery: it now returns False (don't
retry) when any oversized image part can't be brought under target, so
the single retry isn't burned re-sending a payload that will fail
identically. Previously it returned True after shrinking *any* part,
even when the actual oversized culprit survived.
The per-turn file-mutation verifier footer rendered failed-write paths as
bare absolute paths in the user-facing response. The gateway's
extract_local_files() scans response text for bare paths ending in a
deliverable extension (.yaml/.json/etc.), validates os.path.isfile(), and
auto-attaches matches as native uploads — so a denied write to
~/.hermes/config.yaml surfaced the path in the footer and got the
credential file silently uploaded to the messaging channel.
The gateway denylist (validate_media_delivery_path) already blocks the
config.yaml case after #35634. This is defense-in-depth at the source:
backtick-wrap every path the footer emits — both the bullet path and any
path echoed inside the tool's error preview (the protected-file denial
message embeds the path in single quotes, which do NOT block the
extractor regex). extract_local_files skips paths inside inline-code
spans, so wrapping defeats auto-attachment for ANY protected file while
keeping the path human-readable.
- run_agent.py: _format_file_mutation_failure_footer wraps bullet paths;
new _neutralize_footer_paths backticks any remaining bare path (covers
the preview echo). staticmethod -> classmethod (caller unaffected).
- tests: backtick-wrap assertion + end-to-end extract_local_files leak test.
In the concurrent tool-execution path, checkpoint preflight (write_file,
patch, destructive terminal) fired BEFORE plugin guardrail block_result
was computed. A blocked write_file could still dirty checkpoint state
(doc_modified_this_turn, _last_write_file_call_id, turn_counter).
Move checkpoint preflight to AFTER block_result computation, gated on
`if block_result is None:` — matching the invariant the sequential path
already enforces.
The max-iterations summary path (`handle_max_iterations`) hand-builds its
message list and calls `chat.completions.create()` directly, bypassing
`ChatCompletionsTransport.convert_messages()`. It only popped
("reasoning", "finish_reason", "_thinking_prefill"), so `tool_name` (SQLite
FTS bookkeeping), the `codex_*` reasoning carriers, and other internal
`_`-prefixed scaffolding leaked to the wire.
Strict OpenAI-compatible gateways (Fireworks-backed OpenCode Go, Mistral,
Moonshot/Kimi) reject these with HTTP 400 "Extra inputs are not permitted,
field: 'messages[N].tool_name'", so a long tool-using session that exhausts
the iteration budget fails to summarise instead of returning the result.
Mirror convert_messages() in this path: also drop tool_name,
codex_reasoning_items, codex_message_items, and every `_`-prefixed key.
Copy-on-write is already in place, so internal history keeps the fields for
FTS / Codex-fallback.
Adds a regression test to TestHandleMaxIterations asserting the summary
request carries none of the schema-foreign keys (fails on main, passes here).
Follow-up to the salvaged #34452 turn-completion explainer:
- Register display.turn_completion_explainer: True in DEFAULT_CONFIG so the
setting is discoverable, matching the file_mutation_verifier precedent.
- Shorten the repeated footer prefix from 'Turn ended without a usable
reply: ' to 'No reply: ' so the 10 reason variants don't all open with
the same 8-word boilerplate.
- Update the 7 assertions that referenced the old prefix.
PR #34470 adds an explainer suffix to abnormal turn endings (e.g.
max_iterations_reached) so users see why the response is short instead
of receiving a bare/blank reply. test_tool_call_validation_accepts_dict_arguments
runs the agent at max_iterations=3 which hits the explainer path; the
existing strict-equality assertion (== "done") no longer matches once
the suffix is appended.
Switch the assertion to .startswith("done") so the test continues to
verify that the models actual text survives intact while leaving the
explainer suffix wording owned by conversation_loop (where it belongs).
Test now passes (1 passed in 0.88s).
When a turn ends abnormally after substantive tool calls (empty content
after retries, a partial/truncated stream, exhausted retries, or an
iteration/budget limit), the CLI/TUI response area was left blank or
showed only a fragment (e.g. "The") with no consolidated reason. The
internal turn_exit_reason values (empty_response_exhausted,
partial_stream_recovery, etc.) were never surfaced to the user.
Add a turn-completion explainer that mirrors the existing file-mutation
verifier footer: at turn end, map an abnormal turn_exit_reason to a
short, actionable message and either replace the bare "(empty)"
sentinel or append the reason after a partial fragment. Normal
text_response exits (e.g. a terse "Done.") stay quiet.
Gated by display.turn_completion_explainer (default on) with
HERMES_TURN_COMPLETION_EXPLAINER env override, matching the
file-mutation verifier seam.
Closes#34452
* Inspired by Claude Code: /compress here [N] — boundary-aware 'summarize up to here'
Adds a user-chosen compression boundary to the existing /compress command.
/compress here [N] summarizes everything except the most recent N exchanges
(default 2), which are preserved verbatim — letting the user pick the
compression boundary instead of relying on the automatic token-budget heuristic.
Inspired by Claude Code's Rewind 'Summarize up to here' action (v2.1.139,
Week 20, May 2026): https://code.claude.com/docs/en/whats-new/2026-w20
- hermes_cli/partial_compress.py: pure split/parse helpers + seam-alternation
guard (shared by CLI and gateway).
- cli.py / gateway/run.py: route 'here [N]' / '--keep N' to partial compression;
compress only the head, re-append the verbatim tail through the seam guard.
- Preserves message-flow role alternation (seam guard merges any illegal
user->user / assistant->assistant adjacency).
- Reuses the existing _compress_context session-rotation/lock machinery — no
changes to the compression core.
- Bare /compress (full) and /compress <focus> behavior unchanged.
Tests: 12 helper unit tests + 5 CLI integration tests + E2E (interleaved
tool-call transcript, degenerate/multimodal seams, real handler path).
* fix: keep CLI context display in sync with preflight token estimate
The status bar reads compressor.last_prompt_tokens, which only updates
from a successful API response. When loaded history is oversized but
compression no-ops (e.g. the auxiliary summary model times out), no fresh
usage arrives and the bar stays frozen at the old, smaller value while the
preflight estimate reports a much larger number — looking permanently out
of sync (reported: 74.4K display vs ~144,669 preflight).
Seed last_prompt_tokens with the fresh preflight estimate (upward-only, so
a real usage figure is never clobbered and a successful compression's
downward correction still wins). Display-only; no behavioral change to
compression, caching, or the agent loop.
The rough-estimate mock supplied only 2 side_effect values but the
conversation loop calls estimate_request_tokens_rough a third time for
the post-response real-token estimate, exhausting the iterator. Use a
callable side_effect that returns 125k once (to fire preflight) then
sub-threshold values, independent of call count.
Worker threads that dispatch Hermes tools started with an empty contextvars.Context and no thread-local approval/sudo callbacks. Add tools/thread_context.propagate_context_to_thread factoring that capture/install/clear lifecycle (mirrors the GHSA-qg5c-hvr5-hjgr pattern), and refactor agent/tool_executor onto it so the security-critical logic lives in one audited place. Update the contextvar-propagation source guard for the new call shape.
Refs #33057
After the legacy session-key path was removed, two parameters became dead
surface on the Nous runtime-resolution chain:
- min_key_ttl_seconds: del'd inside refresh_nous_oauth_pure and pass-through /
telemetry-only in refresh_nous_oauth_from_state, _try_import_shared_nous_state,
_nous_device_code_login, and resolve_nous_runtime_credentials. It controlled the
now-deleted agent-key mint TTL and drives no behavior.
- inference_auth_mode: with the legacy mode gone, AUTO and FRESH are behaviorally
identical; the value only fed _normalize_nous_inference_auth_mode validation and
oauth trace output, never a branch.
Removing inference_auth_mode orphaned its whole supporting cluster
(NOUS_INFERENCE_AUTH_MODE_AUTO/FRESH, NOUS_INFERENCE_AUTH_MODES,
_normalize_nous_inference_auth_mode), and dropping min_key_ttl_seconds orphaned
DEFAULT_AGENT_KEY_MIN_TTL_SECONDS — all deleted here.
Updated every caller (run_agent, auxiliary_client, credential_pool, proxy adapter,
runtime_provider, web_server, main, auth_commands, setup) and pruned the matching
test kwargs. Deleted two tests that exercised the removed surface
(test_legacy_auth_mode_is_rejected, test_try_refresh_..._accepts_explicit_auth_mode).
No behavior change: net -134 LOC of dead code.
The scoping fix added enabled_toolsets/disabled_toolsets to the
agent_runtime_helpers sequential dispatch into handle_function_call, so
test_invoke_tool_dispatches_to_handle_function_call's assert_called_once_with
(exact match) needs the two new kwargs. Both are None for the default agent
fixture.
The two tests in TestRunConversation now verify the new behavior:
- test_kanban_block_called_on_iteration_exhaustion → verifies
_record_task_failure(outcome='timed_out') is called instead of
kanban_block
- test_no_kanban_block_when_not_in_kanban_mode → verifies the bridge
is a no-op when HERMES_KANBAN_TASK is unset
The function names are kept for diff stability; both assert against
_record_task_failure now, which is the correct contract per the gap-2
fix in this PR.
The xAI tool-schema sanitizers (strip_slash_enum, strip_pattern_and_format)
mutate their input in place — that's their documented contract. The two
call sites (chat_completion_helpers.build_api_kwargs and the auxiliary
client) were passing agent.tools straight through, so the first xAI
request would permanently strip slash-containing enum constraints and
pattern/format keywords from the per-agent tool registry.
Effect: any subsequent non-xAI call from the same agent (auxiliary task
routed to Anthropic, OpenRouter fallback, mid-session model switch) saw
the already-stripped schema with no way for the user to notice from
their config.
Fix: deepcopy tools_for_api before sanitizing at both call sites.
The slash-enum bug itself (xAI 400ing on enums with '/') was fixed
earlier by #32443 (Nami4D) — that PR landed the strip but used the
sanitizers directly without copying. This salvages #27907's correctness
contribution (the deepcopy) while skipping its redundant parallel
sanitizer (strip_xai_incompatible_enum_values is functionally
equivalent to the existing strip_slash_enum) and its preflight-
neutrality argument (we chose model-gated preflight in #32443).
3 new tests in tests/run_agent/test_run_agent_codex_responses.py:
- strips_slash_enum_from_outgoing_request — outgoing kwargs has no
slash-containing enum values (functional contract preserved).
- does_not_mutate_agent_tools — headline #27907 regression. Snapshot
agent.tools before build_api_kwargs, assert it survives intact
after. Pre-fix this assertion would have caught the mutation.
- is_idempotent_across_repeated_calls — three xAI requests in a row
each strip cleanly AND don't progressively erode the source schema.
344/344 across tests/agent/test_auxiliary_client.py,
tests/agent/transports/test_codex_transport.py,
tests/run_agent/test_run_agent_codex_responses.py, and
tests/tools/test_schema_sanitizer.py.
Co-authored-by: Gabor Barany <barany.gabor@gmail.com>
Remove unused imports (F401) and duplicate/shadowed import
redefinitions (F811) across the codebase using ruff's safe
autofixes. No behavioral changes -- imports only.
- ~1400 safe autofixes applied across 644 files (net -1072 lines)
- __init__.py re-exports preserved (excluded from F401 removal so
public re-export surfaces stay intact)
- Re-exports that are imported or monkeypatched by tests but look
unused in their defining module are kept with explicit # noqa:
F401 (gateway/run.py load_dotenv; run_agent re-exports from
agent.message_sanitization, agent.context_compressor,
agent.retry_utils, agent.prompt_builder, agent.process_bootstrap,
agent.codex_responses_adapter)
- Unsafe F841 (unused-variable) fixes deliberately skipped -- those
can change behavior when the RHS has side effects
- ruff lints remain disabled in pyproject.toml (only PLW1514 is
selected); this is a one-time cleanup, not a config change
Verification:
- python -m compileall: clean
- pytest --collect-only: all 27161 tests collect (zero import errors)
- core entry points import clean (run_agent, model_tools, cli,
toolsets, hermes_state, batch_runner, gateway)
- static scan: every name any test imports directly from an edited
module still resolves
* fix(codex): surface error code in Responses 'failed' status errors
When a Codex Responses turn ends with status=failed, the response carries
the failure details under `response.error` as
`{code, message, param, ...}`. The previous extractor pulled only
`message`, so users seeing a rate-limit failure got a bare "Slow down"
string indistinguishable from a generic stream truncation; an
internal_error with empty message degraded to a dict dump
("{'code': 'internal_error', 'message': ''}").
Extract a `_format_responses_error()` helper that:
- prefixes `code` when both code and message are present
(e.g. 'rate_limit_exceeded: Slow down')
- falls back to the bare `code` when message is empty
- accepts both dict and attribute-style payloads (SDK and JSON-RPC paths)
- preserves the prior status-only fallback when no error payload exists
Apply the same helper at the sibling site in
`codex_app_server_session.run_turn()` so codex-CLI subprocess turn
failures get the same treatment.
Tests:
- 8 new unit tests for `_format_responses_error` covering both shapes,
empty/missing fields, non-string fields, and the status-only fallback.
- 2 regression tests on `_normalize_codex_response` for failed status
with and without a code, asserting the exact RuntimeError message.
- All 3603 tests in tests/agent/ pass.
Adapted from anomalyco/opencode#28757.
* feat(prompt): universal task-completion guidance + local Python toolchain probe
Two cross-model failure modes get a single-line answer in the cached
system prompt. Both gated by config (default on), both add zero overhead
when not needed, both verified via real AIAgent prompt builds.
## What changed
`TASK_COMPLETION_GUIDANCE` — short prompt block applied to ALL models.
Targets two failure modes observed on a real Sarasota real-estate build
task: (1) Opus stopped after writing an 85-byte stub and gave a prose
response with finish_reason=stop on call #3 of 90; (2) DeepSeek pushed
through a PEP-668 wall, then returned fabricated listings instead of
admitting the blocker. Both behaviors are model-family-agnostic, so the
guidance lives outside the existing tool_use_enforcement gate (~192
tokens, paid once per session via prefix cache).
`tools/env_probe.py` — local Python toolchain probe. Detects
python3/pip/uv/PEP-668 state and emits ONE short line in the system
prompt when something is non-default. Emits NOTHING when the env is
clean (zero token cost for normal users). Skipped entirely for remote
terminal backends (docker/modal/ssh) — they have their own probe.
Example output on a broken environment (the actual case):
Python toolchain: python3=3.11.15 (no pip module),
python=missing (use python3), pip→python3.12 (mismatch),
PEP 668=yes (use venv or uv).
## Config
Both flags live under `agent.` in config.yaml, default True:
agent:
task_completion_guidance: true # universal "finish the job" block
environment_probe: true # local Python toolchain hints
Neither addition required a `_config_version` bump — deep-merge fills
defaults in for existing user configs.
## Validation
| Test surface | Result |
|---|---|
| tests/tools/test_env_probe.py | 10/10 pass (probe unit) |
| tests/run_agent/test_run_agent.py — new classes | 8/8 pass (integration) |
| TestToolUseEnforcementConfig | 17/17 pass (no regression) |
| TestBuildSystemPrompt | 9/9 pass (no regression) |
| TestInvalidateSystemPrompt | 2/2 pass (no regression) |
| tests/agent/test_prompt_builder.py | 124/124 pass (no regression) |
| tests/hermes_cli/ | 5662/5662 pass (config defaults) |
| E2E AIAgent build (broken env) | Both blocks present, 2,178 chars |
| E2E AIAgent build (clean env) | 771-char net overhead, env probe silent |
Adds an optional `messages` keyword to the `MemoryProvider.sync_turn`
contract so external/community memory plugins can receive the OpenAI-style
conversation message list for the completed turn — including assistant tool
calls and tool result content — not just the final assistant text.
Dispatch uses signature inspection (`_provider_sync_accepts_messages`): only
providers that declare a `messages` parameter (or `**kwargs`) receive it; all
existing in-tree providers keep their legacy text-only signature and are
called unchanged. No structured-trace envelope is added to core — providers
reconstruct whatever they need from the standard message list.
Also documents Memori as a standalone community memory provider.
Salvaged from #28065 — rebased onto current main.
Co-authored-by: Dave Heritage <david@memorilabs.ai>
The old test asserted that a non-MiniMax provider returning a generic
overflow (no provider-reported max) would step down to the 128K probe
tier. The salvaged fix from #33673 deliberately removes that step-down
because guessed tiers cause configured 1M sessions to silently shrink.
Update the test to assert the new contract: keep the configured 200K
window and rely on compression instead.
* fix(agent): fallback immediately on provider content-policy blocks
Provider safety-filter refusals (e.g. OpenAI Codex 'flagged for possible
cybersecurity risk', OpenAI moderation 'violates our usage policies',
Anthropic safety-system rejections, Azure content_filter) are
deterministic decisions about a specific prompt. Retrying the same
prompt up to api_max_retries times just reproduces the same refusal and
burns paid attempts before surfacing the generic 'API failed after 3
retries — <provider message>' to Telegram / cron with no indication that
the failure came from the model provider rather than Hermes itself.
Classify these as a new FailoverReason.content_policy_blocked
(non-retryable, should_fallback=True) and route them through the
existing is_client_error path so the loop:
- skips the 3x retry backoff
- activates a configured fallback model immediately
- emits a clear provider-safety message to the user (not the generic
'Non-retryable error (HTTP None)') and surfaces actionable guidance
when no fallback is configured (rephrase, narrow context, or set
fallback_model in hermes config)
- returns a final_response that explicitly tells the user this came
from the model provider, so gateway delivery is unambiguous and
cron last_status reflects the safety block rather than a vague
'agent reported failure'
Patterns are intentionally narrow — verbatim refusal phrasings keyed to
specific provider safety pipelines, not generic words like 'policy' or
'violation' that would collide with billing / format / auth errors.
Regression guards in test_18028_content_policy_blocked.py verify
billing 402s, generic 400s, and OpenRouter account-level
provider_policy_blocked remain distinct classifications.
Salvaged from #18164 onto current main (file restructure: loop logic
moved from run_agent.py to agent/conversation_loop.py, _emit_status →
_buffer_status), broadened patterns beyond the original OpenAI Codex
cybersecurity case to cover OpenAI moderation, Anthropic safety system,
and Azure content_filter; added user-actionable guidance and a clear
final_response so cron/gateway surfaces the policy block instead of a
generic non-retryable error, and added a regression-guard test module
mirroring the is_client_error predicate.
Addresses #18028.
Co-authored-by: Kuan-Chieh Huang <kchuang1015@users.noreply.github.com>
* chore: add kchuang1015 to AUTHOR_MAP
---------
Co-authored-by: Kuan-Chieh Huang <kchuang1015@users.noreply.github.com>
Users report that the CLI/gateway floods them with confusing retry chatter
during transient failures: a single 429 can produce 10+ "Provider/Endpoint/
Retrying in 5s..." lines before the request eventually succeeds. The same
firehose hits Telegram, Discord, Slack, etc. via _emit_status.
This patch defers all retry/fallback/compression status messages until we
know the outcome:
- if the turn ultimately succeeds (any path: primary recovers, fallback
activates, compression unsticks the request), the buffer is silently
dropped — the user sees nothing.
- if every retry and fallback exhausts and the turn fails, the buffer
is flushed at the terminal-failure return so the user sees the full
retry trace alongside the final error.
Backend logging (agent.log) is unchanged — every emission site still
writes to logger.warning/info, so post-mortem diagnosis is intact.
## What changed
run_agent.py: four new methods on AIAgent:
_buffer_status(msg) — defer an _emit_status call
_buffer_vprint(msg) — defer a _vprint(force=True) line
_clear_status_buffer() — drop pending messages on success
_flush_status_buffer() — replay pending messages on terminal failure
agent/conversation_loop.py:
- converted ~30 mid-process emit/vprint sites in the retry, fallback,
compression, empty-response, and stream-watchdog paths to the buffered
helpers
- added _flush_status_buffer() at every terminal-failure return so users
still see the trace when it actually matters
- added _clear_status_buffer() at the "non-empty assistant content"
point (NOT at "API call returned bytes" — empty responses still loop
through the empty-retry path and would otherwise lose their trace
between iterations)
- silenced the two "(´;ω;`) oops, retrying..." / "(╥_╥) error,
retrying..." spinner final-frame messages — the spinner now stops
cleanly so retries leave no visible residue
agent/chat_completion_helpers.py: same conversion for codex TTFB / stale-
stream / fallback-activation status messages.
agent/stream_diag.py: _emit_stream_drop now buffers instead of emitting
directly.
## Tests
tests/run_agent/test_retry_status_buffer.py: 7 unit tests covering
accumulate→flush, clear-on-success, mixed kinds, empty-buffer no-op,
re-buffer after flush, exception swallowing.
Updated 3 existing tests that mocked _emit_status to also mock (or use)
_buffer_status:
- tests/run_agent/test_run_agent.py::test_empty_response_emits_status_for_gateway
- tests/run_agent/test_stream_drop_logging.py (2 tests)
- tests/agent/test_codex_ttfb_watchdog.py (TTFB hint test)
## Validation
Live test: hermes chat -q against an unreachable endpoint with no fallback
exhausts retries and prints the full trace at the end. Same flow against
a working endpoint prints zero retry chatter.
api_messages is built once before the retry loop while the primary provider
is active. When a mid-conversation fallback switches to a require-side thinking
provider (DeepSeek/Kimi/MiMo), assistant turns built under a non-require primary
(e.g. Codex) go out without reasoning_content and the new provider rejects the
request with HTTP 400 ("reasoning_content must be passed back").
Re-apply the echo-back pad against the current provider immediately before
building the request kwargs. Idempotent and a no-op unless the active provider
enforces echo-back, so it covers all fallback paths without affecting normal or
reject-side operation.
Drafted by Claude (Opus 4.7) under human review while fixing a personal deployment.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Snapshot review_agent._session_messages before teardown so close() can
clean per-session state without dropping the user-visible
self-improvement summary. Adds two regressions:
- bg-review summarizer receives captured review-agent tool messages
after review_agent.close() runs
- context-compressor protected-head handoff rehydration populates
_previous_summary and keeps the old handoff out of newly summarized
turns
Salvaged from PR #26039 onto current main after agent/background_review.py
extraction. Original commit 63eaf6055; bg-review test updated to patch
the module-level summarize_background_review_actions in
agent.background_review instead of the now-forwarder
AIAgent._summarize_background_review_actions.
Salvages the transport-side fix from #32911 (@xxxigm). Closes#32892.
The openai SDK's responses.stream() / responses.parse() eagerly call
_make_tools(tools), which iterates tools without a None guard. Passing
tools=None raises TypeError: 'NoneType' object is not iterable before
any HTTP request is issued (openai==2.24.0).
PR #33042 already removed responses.stream() from our own Codex call
paths, so the specific iteration crash inside _make_tools is no longer
on the hot path. But the right API contract is to omit tools entirely
when there are no functions to expose — passing tools=None to the
backend is semantically wrong regardless of the SDK's iteration
behavior, and we'd hit it again on any future code path that hasn't
migrated off responses.stream().
This applies the transport-level part of @xxxigm's fix: move
'tools': response_tools into the if response_tools: branch so the
key is omitted when there are no tools, just like tool_choice and
parallel_tool_calls already are. Skips the run_agent.py-side
_strip_sdk_none_iterables helper from their PR — that path is now
obsolete because the SDK helper that needed defending is gone.
Tests
- tests/run_agent/test_codex_no_tools_nonetype.py: 6 tests trimmed
from @xxxigm's original 13-test file. Drops the obsolete tests for
_strip_sdk_none_iterables and _RecordingResponsesStream (helpers
that don't exist on main anymore), keeps the transport behavior
tests + the SDK contract sanity check that ensures we notice if
upstream ever fixes _make_tools(None).
- 6/6 passing locally.
Co-authored-by: xxxigm <tuancanhnguyen706@gmail.com>
Alex Fournier
kshitijk4poor
Enzo Adami
Brooklyn Nicholson
davidb73-hub
nanami7777777
Burke Autrey
teknium1
HexLab98
Eva
David Robertson
cresslank
Bryan Nathan
Fli
sprmn24
Marcelo
Jai Suphavadeeprasit
liuhao1024
Sk
Ignacio Pastor
webtecnica
Simplelife
Michael Huang
ScotterMonk
neo
Ziliang Peng
dmabry
Alex Jestin Taylor
giggling-ginger
Gille
joaomarcos
pedrommaiaa
Umi4Life
Kshitij Kapoor
Tosko4
gauravsaxena1997
doncazper
dsad
hmirin
isheng
AIalliAI
Tranquil-Flow
xxxigm
devatnull
David Metcalfe
eliteworkstation94-ai
nankingjing
Jaaneek
Jneeee
openhands
snav
HODLCLONE
kyssta-exe
Dutch Dim
rrevenanttt
Gustavo Mendes
Tim Roth
Omar Baradei
Jeff Watts
petrichor-op
峯岸 亮
Jan Renz
Chufeng Fan
LeonSGP43
ygd58
HiddenPuppy
Erosika
NiuNiu Xia
nightq
sgaofen
595650661
sweetcornna
infinitycrew39
EloquentBrush0x
konsisumer
LeonSGP43
herbalizer404
YuShu
diamondeyesfox
rebel
Tranquil-Flow
Hao Zhe
Shannon Sands
JP Lew
alt-glitch
tt-a1i
XVVH
Reiji Kisaragi
Wolfram Ravenwolf
SHL0MS
briandevans
0xyg3n
Xiangji
Martín Alcalá Rubí
islam666
bmoore210
Sanidhya Singh
Frowtek
Siddharth Balyan
Nate George
Bryan Bednarski
beardthelion
Max Hsu
Bartok9
firefly
Robin Fernandes
Gabor Barany
Dave Heritage
Biser Perchinkov
Stephen Schoettler