Move the two big tool-dispatch methods out of run_agent.py:
* execute_tool_calls_concurrent — 408-line concurrent path (interrupt
pre-flight, guardrail+plugin block, callback fan-out, ContextVar-
preserving ThreadPoolExecutor, periodic heartbeats for the gateway
inactivity monitor, per-tool result handling with subdir hints +
guardrail observations + checkpoint, /steer drain)
* execute_tool_calls_sequential — 441-line sequential path (the
original behavior used for single-tool batches and interactive
tools)
Both take the parent AIAgent as their first argument; AIAgent keeps
thin forwarders so call sites unchanged. handle_function_call is
routed through _ra() so tests that patch run_agent.handle_function_call
keep working. _set_interrupt likewise.
The AST guard in test_tool_executor_contextvar_propagation.py is
updated to scan both run_agent.py AND agent/tool_executor.py so it
still catches the executor.submit(_run_tool, ...) regression
regardless of which file the body lives in.
tests/run_agent/ + tests/agent/: 4313 passed (same pre-existing
test_auxiliary_client failure as before).
run_agent.py: 14309 -> 13461 lines (-848).
Move the background-review subsystem (the self-improvement loop — see the
README) out of run_agent.py into a dedicated module.
* summarize_background_review_actions — was the @staticmethod that builds
the user-facing action summary
* spawn_background_review_thread — builds the thread target + prompt;
the actual review loop body (forked AIAgent, runtime inheritance,
tool whitelist, suppression, teardown) lives in _run_review_in_thread
* build_memory_write_metadata — provenance for external memory mirrors
AIAgent keeps thin wrappers for backward compatibility AND because tests
patch run_agent.threading.Thread to assert lifecycle behavior — the
threading.Thread construction stays in AIAgent._spawn_background_review,
the inner work moves out.
tests/run_agent/ + tests/agent/: 4313 passed, 1 pre-existing failure
(test_auxiliary_client.py::test_custom_endpoint... — confirmed failing
on main before this change). 3 skipped.
run_agent.py: 15272 -> 14972 lines (-300).
Three small extractions into focused modules:
* agent/process_bootstrap.py — \_OpenAIProxy (lazy openai.OpenAI import),
\_SafeWriter (broken-pipe-resistant stdio wrapper), \_install_safe_stdio,
\_get_proxy_from_env, \_get_proxy_for_base_url. All process / IO bootstrap.
* agent/iteration_budget.py — IterationBudget class (thread-safe consume/
refund counter shared by parent agent and subagents).
run_agent re-exports every name so existing test patches like
patch('run_agent.OpenAI', ...) and 'from run_agent import IterationBudget'
keep working unchanged. Verified the patch-rebinding contract for OpenAI
explicitly.
tests/run_agent/ + tests/agent/test_gemini_fast_fallback.py:
1347 passed, 3 skipped.
run_agent.py: 15427 -> 15261 lines (-166).
Pull the 10 pure sanitization/repair helpers (\_sanitize_surrogates,
\_sanitize_structure_surrogates, \_sanitize_messages_surrogates,
\_escape_invalid_chars_in_json_strings, \_repair_tool_call_arguments,
\_strip_non_ascii, \_sanitize_messages_non_ascii, \_sanitize_tools_non_ascii,
\_strip_images_from_messages, \_sanitize_structure_non_ascii) and the
\_SURROGATE_RE constant out of run_agent.py into a new module.
These are stateless byte-walking helpers with no AIAgent dependency.
Backward compatibility: run_agent re-exports every name via a single
import block, so existing 'from run_agent import _sanitize_surrogates'
imports in tests and cli.py keep working unchanged. Same pattern the
file already uses for _summarize_user_message_for_log (codex_responses_adapter).
run_agent.py: 16077 -> 15682 lines (-395).
Mirrors openclaw beta.8's app-server resilience fixes so a stuck codex
subprocess can't burn the full turn deadline and so users get a
`codex login` pointer instead of raw RPC errors when their token expires.
- TurnResult.should_retire signals the caller to drop+respawn codex.
- Deadline-hit path and dead-subprocess detection set should_retire so
the next turn doesn't ride a CPU-spinning or auth-broken process.
- Post-tool watchdog (post_tool_quiet_timeout=90s): if a tool item
completes and codex goes silent past the threshold without further
output or turn/completed, fast-fail instead of waiting the full 600s.
Resets on any non-tool activity so normal think-after-tool flows are
not affected.
- <turn_aborted> and <turn_aborted/> in agent text are treated as
terminal — some codex builds tear down a turn that way without
emitting turn/completed.
- _classify_oauth_failure() inspects RPC error message + stderr tail
for invalid_grant / token refresh / 401 / etc. and rewrites
user-facing errors to 'run codex login'. Conservative: generic
failures still surface verbatim. Fires at turn/start failure,
turn/completed failure, and dead-subprocess paths.
- thread/start cross-fill: tolerate thread.id, thread.sessionId,
top-level sessionId/threadId so future codex schema drift doesn't
KeyError us at handshake.
- run_agent.py: when run_turn returns should_retire=True OR raises,
close + null self._codex_session so the next turn respawns.
Tests: +30 cases across session + integration suites.
tests/agent/transports/test_codex_app_server_session.py 50/50 pass
tests/run_agent/test_codex_app_server_integration.py 27/27 pass
Broader codex scope (transports + cli runtime/migration) 376/376 pass
Background review fork redirected stdout/stderr around run_conversation()
so its iteration messages stay silent. But the memory-provider teardown
(shutdown_memory_provider() and review_agent.close()) fired in the outer
finally block AFTER the redirect_stdout context exited — so provider
teardown prints (Honcho disconnect, Hindsight sync, etc.) leaked into
the parent terminal at end of every turn.
Moves the teardown inside the redirect_stdout scope on the success path
(and nulls review_agent so the finally safety-net skips double-shutdown).
The finally block is rewritten as an exception-path safety net that
re-opens a devnull redirect, since the original 'with' context has
already exited by the time finally runs.
Salvage of #25342 by @ayushere (manually re-applied + merged conflict
with current main's set_thread_tool_whitelist wiring).
When auxiliary.compression.provider is "auto", the compression model
reuses the main model's provider and base_url. The main model's
context_length was correctly picking up custom_providers per-model
overrides (via _custom_providers stored during __init__), but the
auxiliary compression model's context-length detection path in
_check_compression_model_feasibility was not passing custom_providers,
causing it to skip step 0b and fall through to models.dev.
This meant that for providers like NVIDIA NIM where the user has a
per-model context_length in custom_providers (e.g. 196608 for
minimax-m2.7), the auxiliary model would use the models.dev value
(204800) instead of the user-configured one — a subtle discrepancy
that could lead to silent compression issues when the auxiliary model
doesn't actually support the detected context length.
Fix: pass self._custom_providers (already stored as an instance attr
during __init__) to the get_model_context_length() call for the
auxiliary compression model.
Xiaomi MiMo emits reasoning via OpenAI's reasoning_content field and
requires reasoning_content on every assistant tool-call message when
replaying history. Without echo-back, subsequent API calls fail with
HTTP 400 — same shape as DeepSeek and Kimi/Moonshot thinking modes.
Adds _needs_mimo_tool_reasoning() detection (provider == 'xiaomi',
'mimo' in model, or xiaomimimo.com base url) and wires it into the
_needs_thinking_reasoning_pad() check.
Salvage of #25358 by @ephron-ren (manually re-applied — original branch
was severely stale against current main).
Follow-up on the salvaged feat commit:
- Keep the constructor / config / yaml-example default at 3 so existing
gateway and CLI users see no behavioural change. PR #13754 (which this
builds on) had lowered the default to 2 to chase pre-feature parity in
the system-prompt-present case, at the cost of quietly halving the
protected head for the gateway path (which strips the system prompt
before calling compress()). With the new "system prompt is implicit"
semantics, default 3 gives every caller a stable head shape.
- agent/context_engine.py: bring the ABC's protect_first_n docstring in
line with the new semantics so plugin context engines interpret the
config key the same way the built-in compressor does.
- tests: adjust the default-value test (3, not 2) and a stale comment;
per-test protect_first_n=2/3/1 values added in PR #13754 stay as-is
since those tests fix concrete head shapes.
The number of head messages preserved verbatim across context compactions
was previously hardcoded to 3 in AIAgent.__init__. Expose it as
`compression.protect_first_n` in config, matching the existing
`protect_last_n` pattern.
Motivation: users who rely on rolling compaction for long-running sessions
had the opening user/assistant exchange pinned as head forever, which
doesn't always match how they want the session framed after many
compactions. Lowering to 1 preserves the system prompt + first non-system
message; lowering to 0 preserves only the system prompt and lets the
entire first exchange age out naturally through the summary.
Semantics: `protect_first_n` counts non-system head messages protected
**in addition to** the system prompt, which is always implicitly protected
when present. Same meaning across both code paths:
protect_first_n=0 → system prompt only (or nothing if no system message)
protect_first_n=2 → system prompt + first 2 non-system messages (default)
This unifies the CLI path (which reads messages with the system prompt at
position 0) and the gateway path (where the gateway /compress handler
strips the system prompt before calling compress() — see
gateway/run.py L9150-9154 on the parent fork). Previously these two paths
disagreed:
CLI path: protect_first_n=1 → protect system prompt only
Gateway path: protect_first_n=1 → protect first USER turn forever
In practice on long-running gateway sessions the old semantics pinned
whatever stale aside happened to be the first user message, reinserting
it into every compaction summary indefinitely.
Default chosen as 2 (not 3) so that the effective protected head count
remains 3 messages in the common case — assuming a system prompt is
present, default protection becomes system + 2 non-system = 3 total,
matching the pre-feature behaviour where `protect_first_n` was hardcoded
to protect 3 messages total. Sessions without a system prompt will see a
small behaviour change (2 protected head messages instead of 3), but this
is the rare path and the new semantics make the system-prompt-present
case the well-defined one.
Changes:
- agent/context_compressor.py: redefine protect_first_n as the count of
non-system head messages protected beyond the implicit system-prompt
guarantee; both paths converge. Constructor default updated to 2.
- hermes_cli/config.py: add `compression.protect_first_n` default (2),
matching the new semantics. `show_config` label tweaked to
'Protect first: N non-system head messages' for clarity.
- run_agent.py: read protect_first_n from config; 0 is now valid (system
prompt is always implicitly protected).
- cli-config.yaml.example: document the new key and rationale.
- tests/agent/test_context_compressor.py: cover default, override, the
end-to-end `protect_first_n=0` and `protect_first_n=1` behaviour,
the no-system-prompt (gateway) path, and the new shared-semantics
regression test.
Fixes#13751
Tested on Ubuntu 24.04.
Belt-and-suspenders complement to the cached-system-prompt inheritance:
pin session_start and session_id to the parent's so any code path that
re-renders parts of the system prompt (compression, plugin hooks)
still produces byte-identical output. The cached-prompt assignment
already short-circuits the normal rebuild path, but these pins
guarantee parity even if a future code path bypasses the cache.
Idea from simpolism's reference PR #25427 for #25322.
Co-Authored-By: simpolism <32201324+simpolism@users.noreply.github.com>
Background review fork is supposed to hit Anthropic's prefix cache on the
parent's messages_snapshot, but currently doesn't (cache_read=0 on every
fork). Two root causes, fixed in this commit:
1. System prompt is rebuilt at fork time. _cached_system_prompt starts as
None, so run_conversation calls _build_system_prompt, which embeds a
minute-precision "Conversation started: ..." timestamp. Reviews fire
10+ turns after session start, so the minute differs from main's,
producing a 1-character diff that invalidates the byte-exact cache key.
Fix: inherit the parent's _cached_system_prompt directly (same idea as
#17089, which was self-closed for only fixing this half).
2. Tools schema was narrowed via enabled_toolsets=["memory","skills"] for
safety. Anthropic's cache key includes `tools`, which sits before
`system` in the cache hierarchy, so even byte-identical `system` won't
hit when `tools` differs from main's full set.
Fix: drop the schema-level restriction so `tools` matches main, and
deny non-whitelisted tools at runtime via the existing
get_pre_tool_call_block_message gate (hermes_cli/plugins.py:1085,
already called at all three dispatch sites). Install/clear a thread-
local whitelist (added in the previous commit) on the daemon thread.
Append a soft constraint to the review prompt so the model knows.
Real E2E on Sonnet 4.5 (12-tool task + auto-triggered review):
- Per review-call cost: $0.331 → $0.035 (~89% reduction)
- End-to-end per run: $0.848 → $0.629 (~26% reduction)
- Review fork cache_create / cache_read: 88,385 / 0 → 1,234 / 94,404
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* feat(codex-runtime): scaffold optional codex app-server runtime
Foundational commit for an opt-in alternate runtime that hands OpenAI/Codex
turns to a 'codex app-server' subprocess instead of Hermes' tool dispatch.
Default behavior is unchanged.
Lands in three pieces:
1. agent/transports/codex_app_server.py — JSON-RPC 2.0 over stdio speaker
for codex's app-server protocol (codex-rs/app-server). Spawn, init
handshake, request/response, notification queue, server-initiated
request queue (for approval round-trips), interrupt-friendly blocking
reads. Tested against real codex 0.130.0 binary end-to-end during
development.
2. hermes_cli/runtime_provider.py:
- Adds 'codex_app_server' to _VALID_API_MODES.
- Adds _maybe_apply_codex_app_server_runtime() helper, called at the
end of _resolve_runtime_from_pool_entry(). Inert unless
'model.openai_runtime: codex_app_server' is set in config.yaml AND
provider in {openai, openai-codex}. Other providers cannot be
rerouted (anthropic, openrouter, etc. preserved).
3. tests/agent/transports/test_codex_app_server_runtime.py — 24 tests
covering api_mode registration, the rewriter helper (default-off,
case-insensitive, opt-in, non-eligible providers preserved), version
parser, missing-binary handling, error class. Does NOT require codex
CLI installed.
This commit is wire-only: the api_mode is recognized but AIAgent does
not yet branch on it. Followup commits add the session adapter, event
projector, approval bridge, transcript projection (so memory/skill
review still works), plugin migration, and slash command.
Existing tests remain green:
- tests/cli/test_cli_provider_resolution.py (29 passed)
- tests/agent/test_credential_pool_routing.py (included above)
* feat(codex-runtime): add codex item projector for memory/skill review
The translator that lets Hermes' self-improvement loop keep working under the
Codex runtime: converts codex 'item/*' notifications into Hermes' standard
{role, content, tool_calls, tool_call_id} message shape that
agent/curator.py already knows how to read.
Item taxonomy (matches codex-rs/app-server-protocol/src/protocol/v2/item.rs):
- userMessage → {role: user, content}
- agentMessage → {role: assistant, content: text}
- reasoning → stashed in next assistant's 'reasoning' field
- commandExecution → assistant tool_call(name='exec_command') + tool result
- fileChange → assistant tool_call(name='apply_patch') + tool result
- mcpToolCall → assistant tool_call(name='mcp.<server>.<tool>') + tool result
- dynamicToolCall → assistant tool_call(name=<tool>) + tool result
- plan/hookPrompt/etc → opaque assistant note, no fabricated tool_calls
Invariants preserved:
- Message role alternation never violated: each tool item produces at most
one assistant + one tool message in that order, correlated by call_id.
- Streaming deltas (item/<type>/outputDelta, item/agentMessage/delta)
don't materialize messages — only item/completed does. Mirrors how
Hermes already only writes the assistant message after streaming ends.
- Tool call ids are deterministic (codex item id-based) so replays produce
identical messages and prefix caches stay valid (AGENTS.md pitfall #16).
- JSON args use sorted_keys for the same reason.
Real wire formats verified against codex 0.130.0 by capturing live
notifications from thread/shellCommand and including one as a fixture
(COMMAND_EXEC_COMPLETED).
23 new tests, all green:
- Streaming deltas don't materialize (3 paths)
- Turn/thread frame events are silent
- commandExecution: 5 tests including non-zero exit annotation +
deterministic id stability across replays
- agentMessage + reasoning attachment + reasoning consumption
- fileChange: summary without inlined content
- mcpToolCall: namespaced naming + error surfacing
- userMessage: text fragments only (drops images/etc)
- opaque items: no fabricated tool_calls
- Helpers: deterministic id stability + sorted JSON args
- Role alternation invariant across all four tool-shaped item types
This commit is a pure addition. AIAgent integration (the wire that uses the
projector) is the next commit.
* feat(codex-runtime): add session adapter + approval bridge
The third self-contained module: CodexAppServerSession owns one Codex
thread per Hermes session, drives turn/start, consumes streaming
notifications via CodexEventProjector, handles server-initiated approval
requests, and translates cancellation into turn/interrupt.
The adapter has a single public per-turn method:
result = session.run_turn(user_input='...', turn_timeout=600)
# result.final_text → assistant text for the caller
# result.projected_messages → list ready to splice into AIAgent.messages
# result.tool_iterations → tick count for _iters_since_skill nudge
# result.interrupted → True on Ctrl+C / deadline / interrupt
# result.error → error string when the turn cannot complete
# result.turn_id, thread_id → for sessions DB / resume
Behavior:
- ensure_started() spawns codex, does the initialize handshake, and
issues thread/start with cwd + permissions profile. Idempotent.
- run_turn() blocks until turn/completed, drains server-initiated
requests (approvals) before reading notifications so codex never
deadlocks waiting for us, projects every item/completed via the
projector, and increments tool_iterations for the skill nudge gate.
- request_interrupt() is thread-safe (threading.Event); the next loop
iteration issues turn/interrupt and unwinds.
- turn_timeout deadlock guard issues turn/interrupt and records an
error if the turn never completes.
- close() escalates terminate → kill via the underlying client.
Approval bridge:
Codex emits server-initiated requests for execCommandApproval and
applyPatchApproval. The adapter translates Hermes' approval choice
vocabulary onto codex's decision vocabulary:
Hermes 'once' → codex 'approved'
Hermes 'session' or 'always' → codex 'approvedForSession'
Hermes 'deny' / anything else → codex 'denied'
Routing precedence:
1. _ServerRequestRouting.auto_approve_* flags (cron / non-interactive)
2. approval_callback wired by the CLI (defers to
tools.approval.prompt_dangerous_approval())
3. Fail-closed denial when neither is wired
Unknown server-request methods are answered with JSON-RPC error -32601
so codex doesn't hang waiting for us.
Permission profile mapping mirrors AGENTS.md:
Hermes 'auto' → codex 'workspace-write'
Hermes 'approval-required' → codex 'read-only-with-approval'
Hermes 'unrestricted/yolo' → codex 'full-access'
20 new tests, all green. Combined with prior commits this PR now has
67 tests across three modules:
- test_codex_app_server_runtime.py: 24 (api_mode + transport surface)
- test_codex_event_projector.py: 23 (item taxonomy projections)
- test_codex_app_server_session.py: 20 (turn loop + approvals + interrupts)
Full tests/agent/transports/ directory: 249/249 pass — no regressions
to existing transport tests.
Still no wire into AIAgent.run_conversation(); that integration commit
is small and goes next.
* feat(codex-runtime): wire codex_app_server runtime into AIAgent
The integration commit. AIAgent.run_conversation() now early-returns to a
new helper _run_codex_app_server_turn() when self.api_mode ==
'codex_app_server', bypassing the chat_completions tool loop entirely.
Three small surgical edits to run_agent.py (~105 LOC total):
1. Line ~1204 (constructor api_mode validation set):
Add 'codex_app_server' so an explicit api_mode='codex_app_server'
passed to AIAgent() isn't silently rewritten to 'chat_completions'.
2. Line ~12048 (run_conversation, just before the while loop):
Early-return to _run_codex_app_server_turn() when self.api_mode is
'codex_app_server'. Placed AFTER all standard pre-loop setup —
logging context, session DB, surrogate sanitization, _user_turn_count
and _turns_since_memory increments, _ext_prefetch_cache, memory
manager on_turn_start — so behavior outside the model-call loop is
identical between paths. Default Hermes flow is unchanged when the
flag is off.
3. End-of-class (line ~15497):
New method _run_codex_app_server_turn(). Lazy-instantiates one
CodexAppServerSession per AIAgent (reused across turns), runs the
turn, splices projected_messages into messages, increments
_iters_since_skill by tool_iterations (since the chat_completions
loop normally does that per iteration), fires
_spawn_background_review on the same cadence as the default path.
Counter accounting:
_turns_since_memory ← already incremented at run_conversation:11817
(gated on memory store configured) — codex
helper does NOT touch it (would double-count).
_user_turn_count ← already incremented at run_conversation:11793
— codex helper does NOT touch it.
_iters_since_skill ← incremented in the chat_completions loop per
tool iteration. Codex helper increments by
turn.tool_iterations since the loop is bypassed.
User message:
ALREADY appended to messages by run_conversation pre-loop (line 11823)
before the early-return reaches us. Helper does NOT append again.
Regression test test_user_message_not_duplicated guards this.
Approval callback wiring:
Lazy-fetches tools.terminal_tool._get_approval_callback at session
spawn time, passes to CodexAppServerSession. CLI threads with
prompt_toolkit get interactive approvals; gateway/cron contexts get
the codex-side fail-closed deny.
Error path:
Codex session exceptions become a 'partial' result with completed=False
and a final_response that explicitly tells the user how to switch back:
'Codex app-server turn failed: ... Fall back to default runtime with
/codex-runtime auto.' Same return-dict shape as the chat_completions
path so all callers (gateway, CLI, batch_runner, ACP) work unchanged.
9 new integration tests in tests/run_agent/test_codex_app_server_integration.py:
- api_mode='codex_app_server' is accepted on AIAgent construction
- run_conversation returns the expected codex shape
(final_response, codex_thread_id, codex_turn_id, completed, partial)
- Projected messages are spliced into messages list
- _iters_since_skill ticks per tool iteration
- _user_turn_count delegated to standard flow (not double-counted)
- User message appears exactly once (regression guard)
- _spawn_background_review IS invoked (memory/skill review keeps working)
- chat.completions.create is NEVER called (loop fully bypassed)
- Session exception → partial result with /codex-runtime auto hint
- Interrupted turn → partial result with error preserved
Adjacent test runs confirm no regressions:
- tests/run_agent/test_memory_nudge_counter_hydration.py: green
- tests/run_agent/test_background_review.py: green
- tests/run_agent/test_fallback_model.py: green
- tests/agent/transports/: 249/249 green
Still missing for full feature: /codex-runtime slash command, plugin
migration helper, docs page, live e2e test gated on codex binary. Those
are the remaining followup commits.
* feat(codex-runtime): add /codex-runtime slash command (CLI + gateway)
User-facing toggle for the optional codex app-server runtime. Follows the
'Adding a Slash Command (All Platforms)' pattern from AGENTS.md exactly:
single CommandDef in the central registry → CLI handler → gateway handler
→ running-agent guard → all surfaces (autocomplete, /help, Telegram menu,
Slack subcommands) update automatically.
Surface:
/codex-runtime — show current state + codex CLI status
/codex-runtime auto — Hermes default runtime
/codex-runtime codex_app_server — codex subprocess runtime
/codex-runtime on / off — synonyms
Files changed:
hermes_cli/codex_runtime_switch.py (new):
Pure-Python state machine shared by CLI and gateway. Parse args,
read/write model.openai_runtime in the config dict, gate enabling
behind a codex --version check (don't let users opt in to a runtime
they have no binary for; print npm install hint instead).
Returns a CodexRuntimeStatus dataclass that callers render however
suits their surface.
hermes_cli/commands.py:
Single CommandDef entry, no aliases (codex-runtime is its own thing).
cli.py:
Dispatch in process_command() + _handle_codex_runtime() handler that
delegates to the shared module and renders results via _cprint.
gateway/run.py:
Dispatch in _handle_message() + _handle_codex_runtime_command() that
returns a string (gateway sends as message). On a successful change
that requires a new session, _evict_cached_agent() forces the next
inbound message to construct a fresh AIAgent with the new api_mode —
avoids prompt-cache invalidation mid-session.
gateway/run.py running-agent guard:
/codex-runtime joins /model in the early-intercept block so a runtime
flip mid-turn can't split a turn across two transports.
Tests:
tests/hermes_cli/test_codex_runtime_switch.py — 25 tests covering the
state machine: arg parsing (10 cases incl. case-insensitive and
synonyms), reading current runtime (5 cases incl. malformed configs),
writing runtime (3 cases), apply() entry point covering read-only,
no-op, codex-missing-blocked, codex-present-success, disable-no-binary-check,
and persist-failure paths (8 cases). All green.
Adjacent test suites confirm no regressions:
- tests/hermes_cli/test_commands.py + test_codex_runtime_switch.py:
167/167 green
- tests/agent/transports/: 283/283 green when combined with prior commits
Still missing: plugin migration helper, docs page, live e2e test gated on
codex binary. Followup commits.
* feat(codex-runtime): auto-migrate Hermes MCP servers to ~/.codex/config.toml
Translates the user's mcp_servers config from ~/.hermes/config.yaml into
the TOML format codex's MCP client expects. Wired into the
/codex-runtime codex_app_server enable path so users get their MCP tool
surface in the spawned subprocess automatically.
The migration runs on every enable. Failures are non-fatal — the runtime
change still proceeds and the user gets a warning so they can fix the
codex config manually.
What translates (mapping verified against codex-rs/core/src/config/edit.rs):
Hermes mcp_servers.<n>.command/args/env → codex stdio transport
Hermes mcp_servers.<n>.url/headers → codex streamable_http transport
Hermes mcp_servers.<n>.timeout → codex tool_timeout_sec
Hermes mcp_servers.<n>.connect_timeout → codex startup_timeout_sec
Hermes mcp_servers.<n>.cwd → codex stdio cwd
Hermes mcp_servers.<n>.enabled: false → codex enabled = false
What does NOT translate (warned + skipped per server):
Hermes-specific keys (sampling, etc.) — codex's MCP client has no
equivalent. Listed in the per-server skipped[] field of the report.
What's NOT migrated (intentional):
AGENTS.md — codex respects this file natively in its cwd. Hermes' own
AGENTS.md (project-level) is already in the worktree, so codex picks
it up without translation. No code needed.
Idempotency design:
All managed content lives between a 'managed by hermes-agent' marker
and the next non-mcp_servers section header. _strip_existing_managed_block
removes the prior managed region cleanly, preserving any user-added
codex config (model, providers.openai, sandbox profiles, etc.) above
or below.
Files added:
hermes_cli/codex_runtime_plugin_migration.py — pure-Python migration
helper. Public API: migrate(hermes_config, codex_home=None,
dry_run=False) returns MigrationReport with .migrated/.errors/
.skipped_keys_per_server. No external TOML dependency — minimal
formatter handles strings/numbers/booleans/lists/inline-tables.
tests/hermes_cli/test_codex_runtime_plugin_migration.py — 39 tests
covering:
- per-server translation (12): stdio/http/sse, cwd, timeouts,
enabled flag, command+url precedence, sampling drop, unknown keys
- TOML formatter (8): types, escaping, inline tables, error case
- existing-block stripping (4): no marker, alone, with user content
above, with user content below
- end-to-end migrate() (8): empty, dry-run, round-trip, idempotent
re-run, preserves user config, error reporting, invalid input,
summary formatting
Files changed:
hermes_cli/codex_runtime_switch.py — apply() now calls migrate() in
the codex_app_server enable branch. Migration failure logs a warning
in the result message but does NOT fail the runtime change. Disable
path (auto) explicitly skips migration.
tests/hermes_cli/test_codex_runtime_switch.py — 3 new tests:
test_enable_triggers_mcp_migration, test_disable_does_not_trigger_migration,
test_migration_failure_does_not_block_enable.
All 325 feature tests green:
- tests/agent/transports/: 249 (incl. 67 new)
- tests/run_agent/test_codex_app_server_integration.py: 9
- tests/hermes_cli/test_codex_runtime_switch.py: 28 (3 new)
- tests/hermes_cli/test_codex_runtime_plugin_migration.py: 39 (new)
* perf(codex-runtime): cache codex --version check within apply()
Single /codex-runtime invocation could spawn 'codex --version' up to 3
times (state report, enable gate, success message). Each spawn is ~50ms,
so the cumulative cost wasn't a crisis, but it was wasteful and turned a
trivial slash command into something noticeably laggy on slower systems.
Refactored to lazy-once via a closure over a nonlocal cache. First call
spawns; subsequent calls in the same apply() reuse the result.
Behavior unchanged — same return shape, same error handling, same install
hint when codex is missing. Just one subprocess per call instead of three.
Two regression-guard tests added:
- test_binary_check_cached_within_apply: enable path → call_count == 1
- test_binary_check_cached_on_read_only_call: state-report path → call_count == 1
Total tests for /codex-runtime now 30 (was 28); all 143 codex-runtime
tests still green.
* fix(codex-runtime): correct protocol field names found via live e2e test
Three real bugs caught only by running a turn end-to-end against codex
0.130.0 with a real ChatGPT subscription. Unit tests passed because they
asserted on our own (incorrect) wire shapes; the wire format from
codex-rs/app-server-protocol/src/protocol/v2/* is the source of truth and
my initial reading of the README was incomplete.
Bug 1: thread/start.permissions wire format
Was sending {"profileId": "workspace-write"}.
Real format per PermissionProfileSelectionParams enum (tagged union):
{"type": "profile", "id": "workspace-write"}
AND requires the experimentalApi capability declared during initialize.
AND requires a matching [permissions] table in ~/.codex/config.toml or
codex fails the request with 'default_permissions requires a [permissions]
table'.
Fix: stop overriding permissions on thread/start. Codex picks its default
profile (read-only unless user configures otherwise), which matches what
codex CLI users expect — they configure their default permission profile
in ~/.codex/config.toml the standard way. Trying to be clever about
profile selection broke every turn we tested.
Live error before fix: 'Invalid request: missing field type' on every
turn/start, even though our turn/start payload was correct — the field
codex was complaining about was inside the permissions sub-object we
shouldn't have been sending.
Bug 2: server-request method names
Was matching 'execCommandApproval' and 'applyPatchApproval'.
Real names per common.rs ServerRequest enum:
item/commandExecution/requestApproval
item/fileChange/requestApproval
item/permissions/requestApproval (new third method)
Fix: match the documented names. Added handler for
item/permissions/requestApproval that always declines — codex sometimes
asks to escalate permissions mid-turn and silent acceptance would surprise
users.
Live symptom before fix: agent.log showed
'Unknown codex server request: item/commandExecution/requestApproval'
and codex stalled because we replied with -32601 (unsupported method)
instead of an approval decision. The agent reported back 'The write
command was rejected' even though Hermes never showed the user an
approval prompt.
Bug 3: approval decision values
Was sending decision strings 'approved'/'approvedForSession'/'denied'.
Real values per CommandExecutionApprovalDecision enum (camelCase):
accept, acceptForSession, decline, cancel
(also AcceptWithExecpolicyAmendment and ApplyNetworkPolicyAmendment
variants we don't currently use).
Fix: rename _approval_choice_to_codex_decision return values; update
auto_approve_* fallbacks; update fail-closed default from 'denied' to
'decline'. Test mapping table updated to match.
Live test verified after fixes:
$ hermes (with model.openai_runtime: codex_app_server)
> Run the shell command: echo hermes-codex-livetest > .../proof.txt
then read it back
Approval prompt fired with 'Codex requests exec in <cwd>'.
User chose 'Allow once'. Codex executed the command, wrote the file,
read it back. Final response: 'Read back from proof.txt:
hermes-codex-livetest'. File contents on disk match.
agent.log confirms:
codex app-server thread started: id=019e200e profile=workspace-write
cwd=/tmp/hermes-codex-livetest/workspace
All 20 session tests still green after wire-format updates.
* fix(codex-runtime): correct apply_patch approval params + ship docs
Live e2e revealed FileChangeRequestApprovalParams doesn't carry the
changeset (just itemId, threadId, turnId, reason, grantRoot) — Codex's
'reason' field describes what the patch wants to do. Test config and
display logic updated to use it. The first 'apply_patch (0 change(s))'
display from the live test is now 'apply_patch: <reason>'.
Adds website/docs/user-guide/features/codex-app-server-runtime.md
covering enable/disable, prerequisites, approval UX, MCP migration
behavior, permission profile delegation to ~/.codex/config.toml, known
limitations, and the architecture diagram. Wired into the Automation
category in sidebars.ts.
Live e2e validation across the path matrix:
✓ thread/start handshake
✓ turn/start with text input
✓ commandExecution items + projection
✓ item/commandExecution/requestApproval → Hermes UI → response
✓ Approve once → command runs
✓ Deny → command rejected, codex falls back to read-only message
✓ Multi-turn (codex remembers prior turn's results)
✓ apply_patch via Codex's fileChange path
✓ item/fileChange/requestApproval → Hermes UI
✓ MCP server migration loads inside spawned codex (verified via
'use the filesystem MCP tool' prompt)
✓ /codex-runtime auto → codex_app_server toggle cycle
✓ Disable doesn't trigger migration
✓ Enable with codex CLI present succeeds + migrates
✓ Hermes-side interrupt path (turn/interrupt request issued cleanly
even if codex finishes before the interrupt lands)
Known live-validated limitations now documented in the docs page:
- delegate_task subagents unavailable on this runtime
- permission profile selection delegated to ~/.codex/config.toml
- apply_patch approval prompt has no inline changeset (codex protocol
doesn't expose it)
145/145 codex-runtime tests still green.
* feat(codex-runtime): native plugin migration + UX polish (quirks 2/4/5/10/11)
Major: migrate native Codex plugins (#7 in OpenClaw's PR list)
Discovers installed curated plugins via codex's plugin/list RPC and
writes [plugins."<name>@<marketplace>"] entries to ~/.codex/config.toml
so they're enabled in the spawned Codex sessions. This is the
'YouTube-video-worthy' bit Pash highlighted: when a user has
google-calendar, github, etc. installed in their Codex CLI, those
plugins activate automatically when they enable Hermes' codex runtime.
Implementation:
- hermes_cli/codex_runtime_plugin_migration.py: new _query_codex_plugins()
helper spawns 'codex app-server' briefly and walks plugin/list. Returns
(plugins, error) — failures are non-fatal so MCP migration still works.
- render_codex_toml_section() now takes plugins + permissions args.
- migrate() defaults: discover_plugins=True, default_permission_profile=
'workspace-write'. Explicit None on either disables that side.
- _strip_existing_managed_block() now also strips [plugins.*] and
[permissions]/[permissions.*] sections inside the managed block, so
re-runs replace plugins cleanly without touching codex's own config.
Quirk fixes:
#2 Default permissions profile written on enable.
Without this, Codex's read-only default kicks in and EVERY write
triggers an approval prompt. Now writes [permissions] default =
'workspace-write' so the runtime feels normal out of the box. Set
default_permission_profile=None to opt out.
#4 apply_patch approval prompt now shows what's changing.
Codex's FileChangeRequestApprovalParams doesn't carry the changeset.
Session adapter now caches the fileChange item from item/started
notifications and looks it up by itemId when codex requests approval.
Prompt shows '1 add, 1 update: /tmp/new.py, /tmp/old.py' instead of
'apply_patch (0 change(s))'.
Side benefit: also drains pending notifications BEFORE handling a
server request, so the projector and per-turn caches are up to date
when the approval decision fires. Bounded to 8 notifications per
loop iter to avoid starving codex's response.
#5/#10 Exec approval prompt never shows empty cwd.
When codex omits cwd in CommandExecutionRequestApprovalParams, fall
back to the session's cwd. If somehow neither is available, show
'<unknown>' explicitly instead of an empty string.
Also surfaces 'reason' from the approval params when codex provides
it — gives users more context on why codex wants to run something.
#11 Banner indicates the codex_app_server runtime when active.
New 'Runtime: codex app-server (terminal/file ops/MCP run inside
codex)' line appears in the welcome banner only when the runtime is
on. Default banner is unchanged.
Tests:
- 7 new tests in test_codex_runtime_plugin_migration.py covering
plugin discovery (mocked), failure handling, dry-run skip, opt-out
flag, idempotent re-runs, and permissions writing.
- 3 new tests in test_codex_app_server_session.py covering the
enriched approval prompts: cwd fallback, change summary on
apply_patch, fallback when no item/started cache exists.
- All 26 session tests + 46 migration tests green; 153 total in PR.
* feat(codex-runtime): hermes-tools MCP callback + native plugin migration
The big architectural addition: when codex_app_server runtime is on,
Hermes registers its own tool surface as an MCP server in
~/.codex/config.toml so the codex subprocess can call back into Hermes
for tools codex doesn't ship with — web_search, browser_*, vision,
image_generate, skills, TTS.
Also: 'migrate native codex plugins' (Pash's YouTube-video-worthy bit) —
when the user has plugins like Linear, GitHub, Gmail, Calendar, Canva
installed via 'codex plugin', Hermes discovers them via plugin/list and
writes [plugins.<name>@openai-curated] entries so they activate
automatically.
New module: agent/transports/hermes_tools_mcp_server.py
FastMCP stdio server exposing 17 Hermes tools. Each call dispatches
through model_tools.handle_function_call() — same code path as the
Hermes default runtime. Run with:
python -m agent.transports.hermes_tools_mcp_server [--verbose]
Exposed: web_search, web_extract, browser_navigate / _click / _type /
_press / _snapshot / _scroll / _back / _get_images / _console /
_vision, vision_analyze, image_generate, skill_view, skills_list,
text_to_speech.
NOT exposed (deliberately):
- terminal/shell/read_file/write_file/patch — codex has built-ins
- delegate_task/memory/session_search/todo — _AGENT_LOOP_TOOLS in
model_tools.py:493, require running AIAgent context. Documented
as a limitation and surfaced in the slash command output.
Migration changes (hermes_cli/codex_runtime_plugin_migration.py):
- _query_codex_plugins() spawns 'codex app-server' briefly to walk
plugin/list and pull installed openai-curated plugins. Failures are
non-fatal — MCP migration still completes.
- render_codex_toml_section() now takes plugins + permissions args
AND wraps the managed block with a MIGRATION_END_MARKER comment so
the stripper can reliably find both ends, even when the block
contains top-level keys (default_permissions = ...).
- migrate() defaults: discover_plugins=True, expose_hermes_tools=True,
default_permission_profile=':workspace' (built-in codex profile name
— must be prefixed with ':'). All three opt-out via explicit args.
- _build_hermes_tools_mcp_entry() builds the codex stdio entry with
HERMES_HOME and PYTHONPATH passthrough so a worktree-launched
Hermes points the MCP subprocess at the same module layout.
Live-caught wire bugs fixed during this turn:
1. Permission profile config key is top-level , NOT a [permissions] table. The [permissions] table is
for *user-defined* profiles with structured fields. Built-in
profile names start with ':' (':workspace', ':read-only',
':danger-no-sandbox'). Was emitting
which codex rejected with 'invalid type: string "X", expected
struct PermissionProfileToml'.
2. Built-in profile is , NOT . Codex
rejected with 'unknown built-in profile'.
3. Codex's MCP layer sends for
tool-call confirmation. We weren't handling it, so codex stalled
and returned 'MCP tool call was rejected'. Now: auto-accept for
our own hermes-tools server (user already opted in by enabling
the runtime), decline for third-party servers.
Quirk fixes shipped (from the limitations list):
#2 default permissions: workspace profile written on enable. No more
approval prompt on every write.
#4 apply_patch approval shows what's changing: cache fileChange
items from item/started, look up by itemId when codex sends
item/fileChange/requestApproval. Prompt: '1 add, 1 update:
/tmp/new.py, /tmp/old.py' instead of '0 change(s)'.
#5/#10 exec approval cwd never empty: fall back to session cwd, then
'<unknown>'. Also surfaces 'reason' from codex when present.
#11 banner shows 'Runtime: codex app-server' line when active so
users understand why tool counts may not match what's reachable.
Tests:
- 5 new tests in test_codex_runtime_plugin_migration.py covering
plugin discovery, expose_hermes_tools entry generation, idempotent
re-runs, opt-out flag, permissions profile.
- 3 new tests in test_codex_app_server_session.py covering enriched
approval prompts (cwd fallback, fileChange summary).
- 2 new tests for mcpServer/elicitation/request handling (accept
hermes-tools, decline others).
- New test file test_hermes_tools_mcp_server.py covering module
surface, EXPOSED_TOOLS safety invariants (no shell/file_ops,
no agent-loop tools), and main() error paths.
- 166 codex-runtime tests total, all green.
Live e2e validated against codex 0.130.0 + ChatGPT subscription:
✓ /codex-runtime codex_app_server enables, migrates filesystem MCP,
registers hermes-tools, writes default_permissions = ':workspace'
✓ Banner shows 'Runtime: codex app-server' line in subsequent sessions
✓ Shell command runs without approval prompt (workspace profile works)
✓ Multi-turn — codex remembers prior turn's results
✓ apply_patch path via fileChange request approval
✓ web_search via hermes-tools MCP callback returns real Firecrawl
results: 'OpenAI Codex CLI – Getting Started' end-to-end in 13s
✓ Disable cycle clean
Docs updated: website/docs/user-guide/features/codex-app-server-runtime.md
Full re-write covering native plugin migration, the hermes-tools
callback architecture, the prerequisites change ('codex login is
separate from hermes auth login codex'), the trade-off table now
reflecting which Hermes tools work via callback, and the limitations
list updated with what's actually unavailable on this runtime.
* feat(codex-runtime): pin user-config preservation invariant for quirk #6
Quirk #6 from the limitations list — user MCP servers / overrides /
codex-only sections in ~/.codex/config.toml that live OUTSIDE the
hermes-managed block must survive re-migration verbatim.
This already worked thanks to the MIGRATION_MARKER + MIGRATION_END_MARKER
pair I added when fixing the default_permissions wire format (so the
strip can find both ends of the managed region even with top-level
keys like default_permissions). But it was an emergent property
without a test pinning it.
Now explicitly tested:
- User MCP server above the managed block survives migration
- User MCP server below the managed block survives migration
- Both above + below survive a second re-migration
- User content (model, providers, sandbox, otel, etc.) outside our
region is left untouched
Docs added a section "Editing ~/.codex/config.toml safely" explaining
the marker contract — so users know they can add their own MCP
servers, override permissions, configure codex-only options, etc.
without fear of Hermes overwriting their work.
167 codex-runtime tests, all green.
* docs(codex-runtime): clarify the actual tool surface — shell covers terminal/read/write/find
Previous docs and PR description undersold what codex's built-in
toolset actually provides. apply_patch alone made it sound like the
runtime could only edit files in patch format — implying you'd lose
terminal use, read_file, write_file, search/find. That was wrong.
Codex's 'shell' tool runs arbitrary shell commands inside the sandbox,
which covers everything you'd do in bash: cat/head/tail (read), echo>
or heredocs (write), find/rg/grep (search), ls/cd (navigate), build/
test/git/etc. apply_patch is for structured multi-file edits on top
of that. update_plan is its in-runtime todo. view_image loads images.
And codex has its own web_search built in (in addition to the
Firecrawl-backed one Hermes exposes via MCP callback).
Docs now have a 'What tools the model actually has' section right
after Why, breaking the surface into three clearly-labeled buckets:
1. Codex's built-in toolset (always on) — shell, apply_patch,
update_plan, view_image, web_search; covers everything terminal-
adjacent.
2. Native Codex plugins (auto-migrated from your codex plugin
install) — Linear, GitHub, Gmail, Calendar, Outlook, Canva, etc.
3. Hermes tool callback (MCP server in ~/.codex/config.toml) —
web_search/web_extract via Firecrawl, browser_*, vision_analyze,
image_generate, skill_view/skills_list, text_to_speech.
Plus a 'What's NOT available' callout listing the four agent-loop tools
(delegate_task, memory, session_search, todo) that need running
AIAgent context and can't reach the codex runtime.
Trade-offs table broken out: shell, apply_patch, update_plan,
view_image, sandbox each get their own row with a one-line description
so users can see at a glance what's available natively.
Architecture diagram updated to list the codex built-ins by name
instead of 'apply_patch + shell + sandbox'.
No code changes — purely docs clarification. 167 codex-runtime tests
still green.
* fix(codex-runtime): _spawn_background_review signature + review fork api_mode downgrade
Two real bugs in the self-improvement loop integration that the previous
test mocked away.
Bug 1: wrong call signature
The codex helper was calling self._spawn_background_review() with no
args after every turn. That function actually requires:
messages_snapshot=list (positional or keyword)
review_memory=bool (at least one trigger must be True)
review_skills=bool
So the call would have raised TypeError at runtime — except the only
test that exercised this path mocked _spawn_background_review entirely
and just asserted spawn.called, so the wrong-arg shape never surfaced.
Bug 2: review fork inherits codex_app_server api_mode
The review fork is constructed with:
api_mode = _parent_runtime.get('api_mode')
So when the parent is codex_app_server, the review fork ALSO runs as
codex_app_server. But the review fork's whole job is to call agent-loop
tools (memory, skill_manage) which require Hermes' own dispatch — they
short-circuit with 'must be handled by the agent loop' on the codex
runtime. So the review fork would have run, decided to save something,
called memory or skill_manage, and silently no-op'd.
Fixed in run_agent.py:_spawn_background_review() — when the parent
api_mode is 'codex_app_server', the review fork is downgraded to
'codex_responses' (same OAuth credentials, same openai-codex provider,
but talks to OpenAI's Responses API directly so Hermes owns the loop).
Also rewrote the codex helper's review wiring to match the
chat_completions path:
- Computes _should_review_memory in the pre-loop block (was already
being computed; now passed through to the helper as an arg).
- Computes _should_review_skills AFTER the codex turn returns +
counters tick (line ~15432 pattern in chat_completions).
- Calls _spawn_background_review(messages_snapshot=, review_memory=,
review_skills=) only when at least one trigger fires.
- Adds the external memory provider sync (_sync_external_memory_for_turn)
that the chat_completions path runs after every turn.
Tests:
Replaced the broken test_background_review_invoked (which only
asserted spawn.called) with three sharper tests:
- test_background_review_NOT_invoked_below_threshold:
single turn at default thresholds → no review fires (would have
caught the original 'every turn calls spawn with no args' bug)
- test_background_review_skill_trigger_fires_above_threshold:
10 tool_iterations at threshold=10 → review fires with
messages_snapshot=list, review_skills=True, counter resets
- test_background_review_signature_never_breaks: regression guard
asserting positional args are always empty and kwargs include
messages_snapshot
New TestReviewForkApiModeDowngrade class:
- test_codex_app_server_parent_downgrades_review_fork: drives the
real _spawn_background_review function (no mock at that level),
asserts the review_agent gets api_mode='codex_responses' when
the parent was codex_app_server.
Live-validated against real run_conversation:
- Counter ticked from 0 to 5 after a 5-tool-iteration turn
- _spawn_background_review fired exactly once with kwargs-only signature
- review_skills=True, review_memory=False
- messages_snapshot was 12 entries (5 assistant tool_calls + 5 tool
results + 1 final assistant + initial system/user)
- Counter reset to 0 after fire
170 codex-runtime tests, all green.
Docs: added a Self-improvement loop section to the codex runtime page
explaining both how the trigger logic stays equivalent and that the
review fork is auto-downgraded to codex_responses for the agent-loop
tools. Also clarified that apply_patch and update_plan ARE codex's
built-in tools (the previous version made it sound like they were
separate from 'codex's stuff' — they're not, all five tools listed
in 'What tools the model actually has' section 1 are codex built-ins).
* feat(codex-runtime): expose kanban tools through Hermes MCP callback
Kanban workers spawn as separate hermes chat -q subprocesses that read
the user's config.yaml. If model.openai_runtime: codex_app_server is set
globally (which is the whole point of opt-in), every dispatched worker
ALSO comes up on the codex runtime.
That mostly works — codex's built-in shell + apply_patch + update_plan
do the actual task work fine — but it had one critical break: the
worker handoff tools (kanban_complete, kanban_block, kanban_comment,
kanban_heartbeat) are Hermes-registered tools, not codex built-ins.
On the codex runtime, codex builds its own tool list and these never
reach the model, so the worker would do the work but not be able to
report back, hanging until the dispatcher's timeout escalates it as
zombie.
Fix: add all 9 kanban tools to the EXPOSED_TOOLS list in the Hermes
MCP callback. They dispatch statelessly through handle_function_call()
just like web_search and the others — they read HERMES_KANBAN_TASK
from env (set by the dispatcher), gate correctly (worker tools require
the env var, orchestrator tools require it unset), and write to
~/.hermes/kanban.db.
Why kanban tools work via stateless dispatch when delegate_task/memory/
session_search/todo don't: those four are listed in _AGENT_LOOP_TOOLS
(model_tools.py:493) and short-circuit in handle_function_call() with
'must be handled by the agent loop' — they need to mutate AIAgent's
mid-loop state. Kanban tools have no such requirement; they're pure
side-effect functions against the kanban.db plus state_meta.
Tools exposed:
Worker handoff (require HERMES_KANBAN_TASK):
kanban_complete, kanban_block, kanban_comment, kanban_heartbeat
Read-only board queries:
kanban_show, kanban_list
Orchestrator (require HERMES_KANBAN_TASK unset):
kanban_create, kanban_unblock, kanban_link
Tests:
- test_kanban_worker_tools_exposed: complete/block/comment/heartbeat
in EXPOSED_TOOLS (regression guard for the would-hang-worker bug)
- test_kanban_orchestrator_tools_exposed: create/show/list/unblock/link
Docs:
- New 'Workflow features' section in the docs page covering /goal,
kanban, and cron behavior on this runtime
- /goal: works fully via run_conversation feedback; only caveat is
approval-prompt noise on long writes-heavy goals (mitigated by
the default :workspace permission profile)
- Kanban: enumerated which tools are reachable via the callback and
why the env var propagates correctly through the codex subprocess
to the MCP server subprocess
- Cron: documented as 'not specifically tested' — same rules as the
CLI apply since cron runs through AIAgent.run_conversation
- Trade-offs table gained rows for /goal, kanban worker, kanban
orchestrator
172/172 codex-runtime tests green (+2 from kanban tests).
* docs(codex-runtime): wire /codex-runtime into slash-commands ref + flag aux token cost
Three docs gaps caught during a final audit:
1. /codex-runtime was only in the feature docs page, not in the
slash-commands reference. Added rows to both the CLI section and
the Messaging section so users discover it where they'd look for
slash command syntax.
2. CODEX_HOME and HERMES_KANBAN_TASK weren't in environment-variables.md.
CODEX_HOME lets users redirect Codex CLI's config dir (the migration
honors it). HERMES_KANBAN_TASK is set by the kanban dispatcher and
propagates to the codex subprocess + the hermes-tools MCP subprocess
so kanban worker tools gate correctly — documented as 'don't set
manually' since it's an internal handoff.
3. Aux client behavior on this runtime. When openai_runtime=
codex_app_server is on with the openai-codex provider, every aux
task (title generation, context compression, vision auto-detect,
session search summarization, the background self-improvement review
fork) flows through the user's ChatGPT subscription by default.
This is true for the existing codex_responses path too, but it's
more visible / important here because users explicitly opted in for
subscription billing. Added a 'Auxiliary tasks and ChatGPT
subscription token cost' section to the docs page with a YAML
example showing how to override specific aux tasks to a cheaper
model (typically google/gemini-3-flash-preview via OpenRouter).
Also documents how the self-improvement review fork gets
auto-downgraded from codex_app_server to codex_responses by the
fix earlier in this PR.
No code changes — pure docs. 172 codex-runtime tests still green.
* docs+test(codex-runtime): pin HOME passthrough, document multi-profile + CODEX_HOME
OpenClaw hit a real footgun in openclaw/openclaw#81562: when spawning
codex app-server they were synthesizing a per-agent HOME alongside
CODEX_HOME. That made every subprocess codex's shell tool launches
(gh, git, aws, npm, gcloud, ...) see a fake $HOME and miss the user's
real config files. They had to back it out in PR #81562 — keep
CODEX_HOME isolation, leave HOME alone.
Audit confirms Hermes' codex spawn doesn't have this problem. We do
os.environ.copy() and only overlay CODEX_HOME (when provided) and
RUST_LOG. HOME passes through unchanged. But it was an emergent
property without a test pinning it, so adding a regression guard:
test_spawn_env_preserves_HOME — confirms parent HOME survives intact
in the subprocess env
test_spawn_env_sets_CODEX_HOME_when_provided — confirms codex_home
arg still isolates
codex state correctly
Docs additions:
'HOME environment variable passthrough' section — calls out the
contract explicitly: CODEX_HOME isolates codex's own state, HOME
stays user-real so gh/git/aws/npm/etc. find their normal config.
Cites openclaw#81562 as the cautionary tale.
'Multi-profile / multi-tenant setups' section — addresses the
related concern: profiles share ~/.codex/ by default. For users who
want per-profile codex isolation (separate auth, separate plugins),
documents the manual CODEX_HOME=<profile-scoped-dir> approach.
Explains why we DON'T auto-scope CODEX_HOME per profile: doing so
would silently invalidate existing codex login state for anyone
upgrading to this PR with tokens already at ~/.codex/auth.json.
Opt-in is safer than surprising users.
174 codex-runtime tests (+2 from HOME guards), all green.
* fix(codex-runtime): TOML control-char escapes + atomic config.toml write
Two footguns caught in a final audit pass before merge.
Bug 1: TOML control characters not escaped
The _format_toml_value() helper escaped backslashes and double quotes
but passed literal control characters (\n, \t, \r, \f, \b) through
unchanged. TOML basic strings don't allow literal control characters
— a path or env var containing a newline would produce invalid TOML
that codex refuses to load.
Realistic exposure: pathological cases like a HERMES_HOME with a
trailing newline (env var concatenation accident), or a PYTHONPATH
with a tab from a multi-line shell heredoc.
Fix: escape all five TOML basic-string control sequences (\b \t \n
\f \r) in addition to \\ and \" that we already did. Order
matters — backslash must come first or the other escapes get
re-escaped.
Bug 2: config.toml write wasn't atomic
If the python process crashed between target.mkdir() and the
write_text() finishing, a half-written config.toml could be left
behind. On NFS / Windows / some FUSE mounts this is a real concern;
on ext4/APFS small writes are usually atomic in practice but not
guaranteed.
Fix: write to a tempfile.mkstemp() temp file in the same directory,
then Path.replace() (atomic same-dir rename on POSIX, ReplaceFile on
Windows). On rename failure, clean up the temp file so repeated
failed migrations don't pile up .config.toml.* files.
Tests:
- test_string_with_newline_escaped — \n in value → \n in output
- test_string_with_tab_escaped — \t in value → \t in output
- test_string_with_other_controls_escaped — \r, \f, \b
- test_windows_path_escaped_correctly — backslash doubling
- test_atomic_write_no_temp_leak_on_success — no .config.toml.*
left over after a successful write
- test_atomic_write_cleanup_on_rename_failure — temp file removed
when Path.replace raises (simulated disk full)
180 codex-runtime tests, all green (+6 from this commit).
Footguns audited but NOT fixed (with rationale):
- Concurrent migrations race. Two Hermes processes hitting
/codex-runtime codex_app_server within seconds of each other could
cause one writer to lose entries. Low probability (you'd have to
enable from two surfaces simultaneously) and low impact (just re-run
migration). Adding fcntl/msvcrt locking is more code than it's
worth here. The atomic rename above means each individual write is
consistent — only the merge step is racy.
- Codex protocol version drift. We pin MIN_CODEX_VERSION=0.125 and
check at runtime but don't reject too-new versions. Right call —
the protocol has been stable through 0.125 → 0.130. If OpenAI
breaks it later we'd see the error in test_codex_app_server_runtime
on CI before users hit it.
Drops the duplicate _FILE_MUTATING_TOOLS frozenset in run_agent.py and
imports the canonical FILE_MUTATING_TOOL_NAMES from
agent/tool_result_classification.py (aliased as _FILE_MUTATING_TOOLS to
avoid renaming the existing call sites). Prevents future drift if
another file-mutating tool is added — only one set needs updating.
No behavior change: same frozenset({'write_file', 'patch'}), and the
117 PR-scoped tests still pass.
* feat(nous): unified client=hermes-client-v<version> tag on every Portal request
Every Hermes request to Nous Portal now carries the same
client=hermes-client-v<__version__> tag (e.g. client=hermes-client-v0.13.0
on this release), sourced live from hermes_cli.__version__. The release
script's regex bump auto-aligns it on every release.
Centralized in agent/portal_tags.py and wired into all four call sites:
- NousProfile.build_extra_body (main agent loop, every chat completion)
- auxiliary_client.NOUS_EXTRA_BODY + _build_call_kwargs (aux client)
- run_agent.py compression-summary fallback path
- tools/web_tools.py web_extract fallback
Replaces the client=aux marker added in #24194 with the unified version
tag. Tests assert against the helper output (invariant) rather than the
literal string, so they don't need updating on every release.
* feat(nous): cover /goal judge and kanban specify aux paths
Two aux-using surfaces bypassed call_llm by invoking
client.chat.completions.create() directly without extra_body, so they
were missing the unified Portal client tag:
- hermes_cli/goals.py — /goal standing-goal judge
- hermes_cli/kanban_specify.py — kanban triage specifier
Both now pass extra_body=get_auxiliary_extra_body() or None so they
inherit the version tag when the aux client points at Nous Portal, and
emit nothing otherwise (no tag leak to OpenRouter/Anthropic auxes).
The long-lived prefix-cache layout split the system prompt into stable/
context/volatile blocks and re-derived them on every API call. The
volatile tier (timestamp + memory snapshot + USER profile) ticks per
turn, so the system message bytes mutated mid-conversation and broke
upstream prompt caches (OpenRouter, Nous Portal, Anthropic).
Diagnosed via live wire-format diffing: an 8-turn conversation showed
OLD layout flipping system block[1] sha mid-session at the minute
boundary, dropping cached_tokens to 0 on that turn (cumulative
66.6% vs 83.3% for the single-block layout). Hermes invariant:
history (system + all but the last 1-2 messages) must be static.
Fix: drop the long-lived layout entirely. Single layout everywhere —
system_and_3 with one cached system string built once on first turn,
replayed verbatim on every subsequent turn. Loses cross-session 1h
prefix caching for Claude (the feature that motivated the split), but
within-session caching now actually works on every provider.
Removed:
- run_agent.py: _use_long_lived_prefix_cache flag, _long_lived_cache_ttl,
_supports_long_lived_anthropic_cache method, the long-lived branch in
run_conversation, mark_tools_for_long_lived_cache call site
- agent/prompt_caching.py: apply_anthropic_cache_control_long_lived,
mark_tools_for_long_lived_cache, _mark_system_stable_block helper
- hermes_cli/config.py: prompt_caching.long_lived_prefix and
prompt_caching.long_lived_ttl config keys
- tests/agent/test_prompt_caching_live.py (entire file)
- tests/agent/test_prompt_caching.py: TestMarkToolsForLongLivedCache,
TestApplyAnthropicCacheControlLongLived
- tests/run_agent/test_anthropic_prompt_cache_policy.py:
TestSupportsLongLivedAnthropicCache
Targeted tests: 62/62 pass.
When switching models via /model, AIAgent._config_context_length was
never cleared, so the new model inherited the previous model's context
window instead of auto-detecting the correct one via
get_model_context_length().
Clear _config_context_length to None before the runtime field swap so
the full resolution chain (custom_providers per-model, endpoint probe,
models.dev, etc.) is re-evaluated for the newly selected model.
Closes#21509
PR #24151 routed Portal Qwen (qwen3.6-plus) through the prefix_and_2
long-lived cache layout, attaching {"type":"ephemeral","ttl":"1h"}
markers to the tools[-1] entry and the stable system-prefix block.
That layout works for Portal Claude because Anthropic / OpenRouter on
Anthropic routes honour 1h TTL — but Portal Qwen ultimately proxies to
Alibaba DashScope, which documents a single "ephemeral" TTL of 5
minutes on its Context Cache. The ttl="1h" qualifier is silently
dropped upstream, so the two highest-value breakpoints (tools array +
system prefix) never land. Only the rolling-window 5m markers on the
last 2 messages cache, which matches the observed ~25% read rate.
Fix: keep Portal Qwen on cache_control via _anthropic_prompt_cache_policy
returning (True, False), but drop it from _supports_long_lived_anthropic_cache
so it rides the standard system_and_3 5m layout (system + last 3 messages,
all at 5m). Same 4 breakpoints, all in a TTL the upstream actually honours.
Refs: https://www.alibabacloud.com/help/en/model-studio/context-cachehttps://openrouter.ai/docs/features/prompt-caching (Alibaba Qwen
section: "TTL: 5 minutes")
- _supports_long_lived_anthropic_cache: Portal scope narrowed back to Claude
- tests: flip the two qwen long-lived expectations to False, retitle
non_claude_non_qwen_rejected -> non_claude_rejected
Detect when write_file / patch calls fail during a turn and are never
superseded by a successful write to the same path. When the final
text response is delivered, append an advisory footer listing the
files that did NOT change — so models that over-claim 'patched 5 files'
after 4 silent failures can't hide the lie.
Catches the failure mode reported in Ben Eng's llm-wiki session:
grok-4.1-fast issued batches of parallel patches, half failed with
'Could not find old_string', and the agent summarised the turn
claiming every file was edited. The user had to manually run
'git status' each turn to catch it.
The verifier is a pure post-hoc check on tool results — no new LLM
calls, no synthetic messages injected into history (prompt cache
preserved), no changes to tool argument dispatch. Per-turn state is
keyed by path; a later successful write to the same path clears the
failure entry so single-file retry recovery is not flagged.
Wired into both _execute_tool_calls_concurrent and
_execute_tool_calls_sequential, so batched parallel patches and one-at-
a-time edits are both covered. Footer emission happens after the
agent loop exits, before transform_llm_output / post_llm_call plugin
hooks run, so plugins still see (and can modify) the augmented text.
Config: display.file_mutation_verifier (bool, default true) +
HERMES_FILE_MUTATION_VERIFIER env override.
31 unit tests in tests/run_agent/test_file_mutation_verifier.py cover
target extraction (write_file, patch-replace, patch-v4a single and
multi-file), error-preview extraction (JSON .error field and plain
string), per-turn state transitions (first-error-wins on repeated
failure, success supersedes failure), footer rendering (truncation
at 10 entries, user-actionable hint), and env/config precedence.
Companion docs updated: user-guide/configuration.md +
reference/environment-variables.md.
Qwen models on Nous Portal (e.g. qwen3.6-plus) now get the same envelope-layout
cache_control markers and long-lived (1h cross-session) cache treatment as
Portal Claude. Portal proxies to OpenRouter with identical wire-format and
cache_control semantics, but the prior policy left Portal Qwen falling through
to the alibaba-family branch (which only matches provider=opencode/alibaba),
serving 0% cache hits and re-billing the full prompt every turn.
Scope is narrow: Portal Claude OR Portal Qwen. Other models on Portal keep
their existing behavior.
- _anthropic_prompt_cache_policy: add (is_nous_portal and qwen) -> (True, False)
- _supports_long_lived_anthropic_cache: drop Claude-only gate for Portal so
Qwen also gets the validated 1h cross-session layout
- tests cover both functions, both bare and vendored qwen slug forms, and
the rejection of non-Claude non-Qwen Portal traffic
Set HERMES_SESSION_ID using the existing session_context.py ContextVar
system for concurrency safety (multiple gateway sessions in one process
won't cross-talk). Also writes os.environ as fallback for CLI mode.
Touchpoints:
- gateway/session_context.py: Add _SESSION_ID ContextVar + _VAR_MAP entry
- run_agent.py: Set both ContextVar and os.environ at init and on
context-compression rotation
- tools/environments/local.py: Bridge ContextVars into subprocess env
in _make_run_env() (ContextVars don't propagate to child processes)
- tests/run_agent/test_session_id_env.py: 3 tests covering env, provided
ID, and ContextVar paths
execute_code subprocess already passes HERMES_* prefixed vars through
_scrub_child_env (line 82: _SAFE_ENV_PREFIXES includes 'HERMES_').
Primary use case: webhook-triggered agents that need to include a
`--resume <session_id>` takeover command in their output.
Cuts input cost for first-turn Claude requests by ~85-90% on subsequent
sessions within an hour. Tools array (~13k tokens for default toolset) +
stable system prefix (~5-8k tokens) get a 1h cache_control marker; the
volatile suffix (memory, USER profile, timestamp, session id) sits in a
separate non-cached block at the end so it doesn't poison the cross-session
prefix when it changes.
Provider gate: Claude on native Anthropic (incl. OAuth subscription),
OpenRouter, and Nous Portal (which proxies to OpenRouter). All other
providers keep today's system_and_3 layout unchanged.
Layout (4 cache_control breakpoints, Anthropic max):
1. tools[-1] -> 1h (cross-session)
2. system content[0] -> 1h (cross-session, stable prefix)
3. messages[-2] -> 5m (within-session rolling)
4. messages[-1] -> 5m (within-session rolling)
Within-session rolling shrinks from 3 messages to 2 to free the breakpoint
budget. On Claude with realistic tool loadouts the long-lived tier carries
the bulk of cross-session value anyway.
System prompt is now always assembled cache-friendly: stable identity /
guidance / skills / platform hints first, then session-stable context
files (AGENTS.md, .cursorrules), then per-call volatile content. Old
single-string callers see the same logical content (same join order),
just reordered so volatile lives at the end.
Config knobs (defaults shown):
prompt_caching:
cache_ttl: "5m" # rolling-window TTL (unchanged)
long_lived_prefix: true # opt-out switch
long_lived_ttl: "1h" # cross-session prefix TTL
Live E2E (tests/agent/test_prompt_caching_live.py, gated on
OPENROUTER_API_KEY) on anthropic/claude-haiku-4.5 with default toolset:
Call 1 (cold): cache_write=13,415 cache_read=0
Call 2 (NEW agent + msg): cache_write=391 cache_read=13,025
Cross-session reuse: 97.09%
Implementation:
* agent/prompt_caching.py: new apply_anthropic_cache_control_long_lived()
+ mark_tools_for_long_lived_cache(); existing apply_anthropic_cache_control()
preserved verbatim for the fallback path.
* agent/anthropic_adapter.py: convert_tools_to_anthropic() now forwards
cache_control onto each Anthropic-format tool dict.
* run_agent.py: _build_system_prompt_parts() returns the 3-tier dict;
_build_system_prompt() joins them (backward compatible).
_supports_long_lived_anthropic_cache() policy added next to the existing
_anthropic_prompt_cache_policy() (which now also recognises Nous Portal
Claude — pre-existing gap fixed in passing).
_build_api_kwargs() resolves tools_for_api once and propagates the
marker through all four build paths (anthropic_messages, bedrock,
codex_responses, profile/legacy chat completions).
Long-lived flag plumbed into the runtime snapshot/restore + model-switch
+ fallback-promotion paths.
Tests:
* tests/agent/test_prompt_caching.py: +8 tests (TestMarkToolsForLongLivedCache,
TestApplyAnthropicCacheControlLongLived).
* tests/run_agent/test_anthropic_prompt_cache_policy.py: +9 tests
(TestSupportsLongLivedAnthropicCache matrix across 8 endpoint classes
+ a fallback-target case).
* tests/agent/test_prompt_caching_live.py: new live E2E (skipif when
OPENROUTER_API_KEY is unset; runs outside the hermetic suite).
* Targeted suites: 327/327 pass (caching/adapter/policy/builder).
* tests/agent/ + tests/run_agent/: 3992 pass, 17 skip, 1 pre-existing
flake (test_async_httpx_del_neuter::test_same_key_replaces_stale_loop_entry,
verified failing on pristine origin/main).
Replace with for all literal-tuple
membership tests. Set lookup is O(1) vs O(n) for tuple — consistent
micro-optimization across the codebase.
608 instances fixed via `ruff --fix --unsafe-fixes`, 0 remaining.
133 files, +626/-626 (net zero).
When the user's main provider is openai-codex on the ChatGPT-account
backend (https://chatgpt.com/backend-api/codex), sending a native image
attachment encodes it as data:image/...base64,... in the input_image
field. The OpenAI Responses API on the public endpoint accepts that, but
the ChatGPT-account variant rejects it with HTTP 400:
Invalid 'input[N].content[K].image_url'. Expected a valid URL, but got
a value with an invalid format.
Hermes' image-rejection phrase list didn't include this wording, so the
error escaped the strip-and-retry branch and fell through to the generic
recovery path: model fallback → context-too-large → compression cascade
→ auxiliary OpenRouter 402 spam (issue #23570).
Add a NARROW phrase keyed on the field-path apostrophe used by the Codex
Responses error format: "image_url'. expected". This matches the actual
error format without false-tripping on generic 'Expected a valid URL'
errors from unrelated tools (webhooks, redirect_uri, etc.). Once matched,
the existing branch strips images from history, sets _vision_supported=
False for the session, and retries text-only.
Refs #23570 (1 of 3 image-replay improvements; persistence rewrite to
store image PATHS instead of inlined base64 is a separate follow-up)
* Revert "fix(goals): force judge to use tool calls instead of JSON-text replies (#23547)"
This reverts commit a63a2b7c78.
* Revert "fix(goals): forward standing /goal state on auto-compression session rotation (#23530)"
This reverts commit 4a080b1d5a.
* Revert "feat(goals): /goal checklist + /subgoal user controls (#23456)"
This reverts commit 404640a2b7.
When a kanban worker subprocess hits the iteration budget, the agent
loop strips tools and asks the model for a summary. The model cannot
call kanban_block itself at that point, so the process exits rc=0
without calling kanban_complete or kanban_block — a protocol violation
that the dispatcher detects as a fatal error, giving up after 1 failure
and stranding downstream tasks.
Fix: after _handle_max_iterations() returns, check HERMES_KANBAN_TASK
and call kanban_block with a reason describing the exhaustion. The
dispatcher then sees a clean block transition instead of a protocol
violation, and the task can be retried or escalated by a human.
Fixes [Bug] kanban-worker exits cleanly (rc=0) on iteration-budget
exhaustion without calling kanban_complete or kanban_block #23216
Salvages the three substantive low-severity fixes from Gutslabs' #1974
"misc bug fixes" bundle. The other 8 claims in that PR were either
already fixed on main with superior implementations (state lock,
firecrawl lazy import, fcntl/msvcrt guard, path normalization, schema
migrations) or did not survive review.
- run_agent: `_materialize_data_url_for_vision` uses
`NamedTemporaryFile(delete=False)`; if `base64.b64decode` raises on a
corrupt data URL the temp file would persist forever. Wrap the
write in try/except and `os.unlink` the temp on failure.
- gateway/session: `append_to_transcript` JSONL write had no error
handling, so disk-full / read-only-fs / permission errors crashed the
message handler. The SQLite write above is the primary store, so
swallow OSError on the JSONL fallback with a debug log.
- gateway/status: `_read_pid_record` reads `pid_path.read_text()` after
an `exists()` check; if the PID file is deleted between the two
calls (concurrent gateway restart) we hit an unhandled OSError.
Catch it and return None.
Adds a regression test for the tempfile cleanup; the other two paths
are defensive try/excepts on infrequent OSError that don't warrant
dedicated tests.
Co-authored-by: Teknium <127238744+teknium1@users.noreply.github.com>
When run_agent's _compress_context fires mid-turn it ends the parent
session in SessionDB and creates a new continuation session with a
fresh session_id. The /goal state is keyed on session_id in
state_meta ("goal:<sid>"), so without forwarding the goal silently
disappears: _get_goal_manager() rebinds for the new session_id,
load_goal() returns None, mgr.is_active() is False, and the
continuation loop dies with no user-visible signal.
Fix: in the same SessionDB transaction block that creates the
continuation session, copy state_meta[goal:<old>] →
state_meta[goal:<new>] when present. No-op when the user has no
active goal. Logged at INFO so a stuck loop is debuggable.
Tests cover the round-trip via SessionDB and the no-op path.
Affects all three run-conversation surfaces (CLI, gateway, TUI
gateway) because _compress_context is the single rotation site.
Closes#6051.
Reported failure mode: agent migrated to WSL2, browser launch failed
because Playwright wasn't installed yet. Background reviewer captured
the failure as a durable skill (`browser-tool-launch-issue`) and the
agent kept refusing the browser tool for weeks after Playwright was
installed and verified working. Negative claims also propagated into
unrelated skills ("browser tools do not work", "cannot use Y from
execute_code").
Root cause: `_SKILL_REVIEW_PROMPT` and `_COMBINED_REVIEW_PROMPT` both
lean hard on "be active, save things, a pass that does nothing is a
missed learning opportunity." Neither distinguished durable knowledge
from transient environment state. The reviewer was doing what it was
told.
Fix at the write site — both prompts now carry a "Do NOT capture"
section calling out:
• Environment-dependent failures (missing binaries, fresh-install
errors, post-migration path mismatches, 'command not found',
unconfigured credentials, uninstalled packages)
• Negative claims about tools or features ("X does not work")
that harden into self-cited refusals
• Session-specific transient errors that resolved before the
conversation ended
• One-off task narratives ("summarize today's market", "analyze
this PR") — also addresses the #12812 / #4538 family
Plus a positive-reframing line: when a tool fails because of setup
state, capture the FIX (install command, config step, env var)
under an existing setup/troubleshooting skill — never "this tool
doesn't work" as a standalone constraint.
Targeted tests: 24/24 passing in tests/run_agent/test_review_prompt_class_first.py
(2 new + all existing review-prompt assertions). Substring-based
checks so future prompt edits don't false-fail.
The previous PR (#22993) gave us a structured WARNING per stream drop
but the only diagnostic was 'error_type=APIError error=Network
connection lost.' — same nothing the user started with. To actually
diagnose why subagents drop streams disproportionately we need to know
WHERE the drop happened.
Adds three breadcrumbs to the agent.log WARNING:
1. Inner exception chain. openai SDK wraps httpx errors as
APIConnectionError / APIError so the catch site only sees the
wrapper. _flatten_exception_chain walks __cause__/__context__ up to
4 levels deep and renders 'Outer(msg) <- Inner(msg)' so we can
tell ConnectError vs RemoteProtocolError vs ReadError vs
ProxyError without enabling verbose mode.
2. Upstream HTTP headers. Snapshots cf-ray, x-openrouter-provider,
x-openrouter-model, x-openrouter-id, x-request-id, server, via,
etc. from stream.response immediately after open (so they survive
even when the stream dies before the first chunk). These answer
'is one CF edge / one downstream provider responsible, or random?'
3. Per-attempt counters. bytes streamed, chunk count, elapsed time on
the dying attempt, and time-to-first-byte. Distinguishes 'couldn't
connect at all' (0s, 0 bytes) from 'died after 30s mid-stream'
(very different root causes — first is auth/routing, second is
upstream idle-kill or proxy timeout).
Plumbing:
- _stream_diag_init / _stream_diag_capture_response live on AIAgent
and produce a per-attempt dict held on request_client_holder['diag']
for closure access from the retry block.
- _call_chat_completions and _call_anthropic both initialize the diag
and increment counters per chunk/event (best-effort, never raises in
the streaming hot path).
- _log_stream_retry / _emit_stream_drop accept an optional diag and
render the new fields. Final-exhaustion log goes through the same
helper so it gets the same diagnostic dump.
- UI status line gains a brief 'after Xs' suffix when timing is
available — distinguishes 'connect failed' from 'died mid-stream'
at a glance without grepping logs.
Sample WARNING after this change:
Stream drop mid tool-call on attempt 2/3 — retrying.
subagent_id=sa-2-cafef00d depth=1 provider=openrouter
base_url=https://openrouter.ai/api/v1
error_type=APIError error=Connection error.
chain=APIError(Connection error.) <- RemoteProtocolError(peer
closed connection without sending complete message body)
http_status=200 bytes=12400 chunks=47 elapsed=12.00s ttfb=0.83s
upstream=[cf-ray=8f1a2b3c4d5e6f7g-LAX
x-openrouter-provider=Anthropic
x-openrouter-id=gen-abc123 server=cloudflare]
Tests: 10 covering diag init, header capture (whitelist enforced for
PII), exception-chain walking + depth cap, log content with full diag,
log content without diag (placeholders), UI elapsed-suffix on/off.
Subagent stream drops were spamming the parent terminal with two lines
per blip ('Connection dropped...' + 'Reconnected...') while leaving zero
breadcrumb in agent.log to debug them.
Two underlying bugs, fixed together:
1. quiet_mode raised the run_agent/tools/etc. loggers to ERROR, which
filters records before root-logger file handlers see them. The comment
claimed 'File handlers still capture everything' — that was wrong.
Removed in both run_agent.py and cli.py; console quietness already
comes from hermes_logging not installing a console StreamHandler in
non-verbose mode.
2. The stream-retry blocks emitted two _emit_status calls per drop
('⚠️ Connection dropped... Reconnecting...' + '🔄 Reconnected —
resuming…') with no provider name, so multi-provider sessions had to
dig through agent.log to attribute a drop. Replaced both call sites
with a single _emit_stream_drop helper that emits ONE line naming the
provider and error class, and always writes a structured WARNING to
agent.log with subagent_id, depth, provider, base_url, error_type.
Net UX change: 6 lines per triple-subagent drop → 3 lines, each
naming the provider. agent.log now has a structured breadcrumb per
retry that didn't exist before.
Tests: 6 new tests in tests/run_agent/test_stream_drop_logging.py
covering the logger-level guard, structured WARNING content, single
status line per drop (no Reconnected follow-up), and provider naming.
When the active main model has native vision and the provider supports
multimodal tool results (Anthropic, OpenAI Chat, Codex Responses, Gemini
3, OpenRouter, Nous), vision_analyze loads the image bytes and returns
them to the model as a multimodal tool-result envelope. The model then
sees the pixels directly on its next turn instead of receiving a lossy
text description from an auxiliary LLM.
Falls back to the legacy aux-LLM text path for non-vision models and
unverified providers.
Mirrors the architecture used in OpenCode, Claude Code, Codex CLI, and
Cline. All four converge on the same pattern: tool results carry image
content blocks for vision-capable provider/model combinations.
Changes
- tools/vision_tools.py: _vision_analyze_native fast path + provider
capability table (_supports_media_in_tool_results). Schema description
updated to reflect new behaviour.
- agent/codex_responses_adapter.py: function_call_output.output now
accepts the array form for multimodal tool results (was string-only).
Preflight validates input_text/input_image parts.
- agent/auxiliary_client.py: _RUNTIME_MAIN_PROVIDER/_MODEL globals so
tools see the live CLI/gateway override, not the stale config.yaml
default. set_runtime_main()/clear_runtime_main() helpers.
- run_agent.py: AIAgent.run_conversation calls set_runtime_main at turn
start so vision_analyze's fast-path check sees the actual runtime.
- tests/conftest.py: clear runtime-main override between tests.
Tests
- tests/tools/test_vision_native_fast_path.py: provider capability
table, envelope shape, fast-path gating (vision-capable model uses
fast path; non-vision model falls through to aux).
- tests/run_agent/test_codex_multimodal_tool_result.py: list tool
content becomes function_call_output.output array; preflight
preserves arrays and drops unknown part types.
Live verified
- Opus 4.6 + Sonnet 4.6 on OpenRouter: model calls vision_analyze on a
typed filepath, gets pixels back, reads exact text from images that
no aux description could capture (font color irony, multi-line
fruit-count list, etc.).
PR replaces the closed prior efforts (#16506 shipped the inbound user-
attached path; this PR closes the gap for tool-discovered images).
Fallback chain entries with 'api_key_env: ENV_VAR_NAME' weren't being
resolved by either the init-time fallback path (line ~1660) or the
runtime _try_activate_fallback path (line ~8045). Only literal
'api_key' was honored; the snake_case 'api_key_env' alias documented
elsewhere in the config was silently dropped, so a 'provider: custom'
fallback with base_url + api_key_env worked as primary but failed as
fallback with 'no endpoint credentials found' / 401.
Adds 'or fb.get("api_key_env")' to the existing 'key_env' lookup in
both call sites, with empty-string-to-None coercion so unset env vars
don't poison the resolver.
Salvage of #22665's fallback portion. The original PR also bundled
gateway-degrade-on-no-adapters changes (those land via the carve-out
in #22853 which is the same code) and run_agent.py memory-nudge
counter hydration (issue #22357 territory, not mentioned in the
title). Drops both bundled pieces; keeps just the api_key_env fix.
Closes#5392.
Pick openrouter/pareto-code as your model and OpenRouter auto-routes each
request to the cheapest model meeting your coding-quality bar (ranked by
Artificial Analysis). The new openrouter.min_coding_score config key (0.0-1.0,
default 0.65) tunes the floor.
- hermes_cli/models.py: add openrouter/pareto-code to OPENROUTER_MODELS so
it shows up in the picker with a description
- hermes_cli/config.py: add openrouter.min_coding_score (default 0.65 — lands
on a mid-tier coder on the current Pareto frontier)
- plugins/model-providers/openrouter: emit extra_body.plugins =
[{id: pareto-router, min_coding_score: X}] when model is openrouter/pareto-code
AND the score is a valid float in [0.0, 1.0]
- agent/transports/chat_completions.py: same emission on the legacy flag
path (when no provider profile is loaded)
- run_agent.py: openrouter_min_coding_score kwarg + storage; plumbed into
both build_kwargs() invocations and the context-summary extra_body path
- cli.py: read openrouter.min_coding_score once at init, validate float in
[0,1], pass to AIAgent constructions (CLI + background-task paths)
- cron/scheduler.py, batch_runner.py, tools/delegate_tool.py,
tui_gateway/server.py: propagate the kwarg (mirrors providers_order
plumbing — subagents inherit, cron/batch read from config)
- tests: profile-level + transport-level coverage of the model gating,
unset/empty/out-of-range handling, and the legacy flag path
- docs: new 'OpenRouter Pareto Code Router' section in providers.md
Verified end-to-end against api.openrouter.ai: at score=0.65 we land on a
mid-tier coder, at omission we get the strongest. Score is silently dropped
on any model other than openrouter/pareto-code, so it's safe to leave set.
DeepSeek V4 Pro returns thinking content as typed blocks inside the
content array rather than as a top-level reasoning_content field:
[{"type": "thinking", "thinking": "..."}, {"type": "output", ...}]
_extract_reasoning only handled content as a plain string, so the
thinking text was silently dropped. On the next turn the session was
replayed without the thinking block, causing:
HTTP 400: The content[].thinking in the thinking mode must be
passed back to the API.
Fix: when content is a list and no structured reasoning field was
found, scan for items with type=='thinking' and accumulate their
'thinking' (or 'text') value into reasoning_parts. Structured fields
(reasoning, reasoning_content, reasoning_details) still take priority
so existing provider behaviour is unchanged.
Closes#21944
_try_activate_fallback() walked the chain by index without comparing
the candidate entry against the currently-failing backend. So a
misconfigured chain that listed the same provider+model as the primary,
or two custom_providers entries pointing at the same shim URL, would
loop the same failure 3x for the same backend.
After the fix, advance() skips:
- entries where (provider, model) match the current agent's
- entries with a base_url + model matching the current backend
(catches two custom_providers names pointing at the same shim)
Recursing through self._try_activate_fallback() continues to the next
chain entry; if everything matches, returns False and the caller
moves on without retrying the same broken path.
3 regression tests covering same-provider-same-model skip, same-base_url-
same-model skip, and the all-self-matching-returns-False exhaustion path.
Closes#22548 (the Hermes-side portion). The 120s timeout itself in
the downstream claude-cli shim is a deployment concern documented in
that issue's wherewolf87 comment.
Gateway creates a fresh AIAgent per inbound message in several common
scenarios: cache miss, idle eviction (1h TTL), config-signature
mismatch, process restart. A freshly-built AIAgent has
_turns_since_memory=0 and _user_turn_count=0, so the
memory.nudge_interval trigger ('_turns_since_memory >=
_memory_nudge_interval') can never be reached when these reconstructions
happen on roughly the cadence of the interval. A user can chat for hours
on Telegram without ever seeing a self-improvement review fire.
Reconstruct the counters from conversation_history at the top of
run_conversation(), right after the existing _hydrate_todo_store call.
Idempotent guard ('if self._user_turn_count == 0') means a cached agent
that already accumulated counters keeps them; only freshly-built agents
hydrate. Modulo arithmetic preserves the original 1-in-N cadence rather
than firing a review immediately on resume.
7 regression tests pinning the contract (mid-cycle history, modulo wrap,
idempotency, zero-interval skip, role==user filtering, production-code
anchor).
Closes#22357.
When session_id rotates (e.g. /new), commit_memory_session was firing
MemoryManager.on_session_end but skipping ContextEngine.on_session_end.
Engines that accumulate per-session state (LCM-style DAGs, summary
stores) leaked that state from the rotated-out session into whatever
continued under the same compressor instance.
Mirror the call shutdown_memory_provider already makes — same
lifecycle moment, same hook contract ("real session boundaries (CLI
exit, /reset, gateway expiry)"). /new is a real boundary for the old
session_id; providers keep their state but the rotated-out session_id
is done.
6 regression tests covering both-hooks-fire, no-memory-manager,
no-context-engine, both failure-tolerant paths.
Closes#22394.
SQLite's WAL mode requires shared-memory (mmap) coordination and fcntl
byte-range locks that don't reliably work on network filesystems. Upstream
documents this explicitly:
https://www.sqlite.org/wal.html#sometimes_queries_return_sqlite_busy_in_wal_mode
On NFS / SMB / some FUSE mounts / WSL1, 'PRAGMA journal_mode=WAL' raises
'sqlite3.OperationalError: locking protocol' (SQLITE_PROTOCOL). Before
this change, every feature backed by state.db or kanban.db broke silently:
- /resume, /title, /history, /branch returned 'Session database not
available.' with no cause
- gateway logged the init failure at DEBUG (invisible in errors.log)
- kanban dispatcher crashed every 60s, driving the known migration race
(duplicate column name: consecutive_failures, #21708 / #21374)
Changes:
- hermes_state.apply_wal_with_fallback(): shared helper that tries WAL
and falls back to DELETE on SQLITE_PROTOCOL-style errors with one
WARNING explaining why
- hermes_state.get_last_init_error() + format_session_db_unavailable():
capture the init failure cause and surface it in user-facing strings
(with an NFS/SMB pointer for 'locking protocol')
- hermes_cli/kanban_db.connect(): use the shared helper
- gateway/run.py: bump SessionDB init failure log DEBUG -> WARNING
(matches cli.py's existing correct behavior)
- cli.py (4 sites) + gateway/run.py (5 sites): replace bare
'Session database not available.' with format_session_db_unavailable()
Tests: 12 new tests in tests/test_hermes_state_wal_fallback.py + 1 new
test in tests/hermes_cli/test_kanban_db.py. Existing suites (state,
kanban, gateway, cli) remain green for all tests unrelated to pre-existing
failures on main.
Evidence: real-world user on NFSv3 mount (172.26.224.200:d2dfac12/home,
local_lock=none) reporting 'Session database not available.' on /resume;
'locking protocol' appears in 4 distinct log entries across backup,
kanban, TUI, and CLI paths in the same session.
closes#22032
teknium1 hit ModuleNotFoundError: No module named 'hermes_bootstrap' after
a code update, on both his Windows machine AND his Linux workstation. The
failure mode is real and affects every user who updates hermes by any path
OTHER than a fully-successful ``hermes update``.
## What happens
hermes_bootstrap.py is a top-level module registered via pyproject.toml's
``py-modules`` list (added by Brooklyn's Windows UTF-8 stdio work). It
must be registered in the venv's editable-install .pth file before Python
can find it as a bare ``import hermes_bootstrap``.
``hermes update`` handles this correctly: (1) git reset --hard, (2) clear
__pycache__, (3) uv pip install -e . (re-registers the package including
the new py-modules list), (4) restart.
BUT if any step AFTER (1) fails — network blip during pip install, PEP 668
on a system Python, venv locked, uv not in PATH, a crash mid-update — the
user is left with new code that references hermes_bootstrap and a venv
that doesn't know about it. Every hermes invocation after that crashes
with ModuleNotFoundError, including ``hermes update`` itself. No recovery
path without manual `uv pip install -e .`.
Also affects users who ``git pull`` the repo directly without running
hermes update — relatively common for developers.
## Fix
Wrap ``import hermes_bootstrap`` in a try/except ModuleNotFoundError
across all 6 entry points (hermes_cli/main, run_agent, gateway/run,
acp_adapter/entry, cli, batch_runner). On Windows, missing bootstrap
means the UTF-8 stdio setup doesn't run — degraded behavior (Unicode
chars may fail to print) but NOT a crash. POSIX is unaffected either way
since the bootstrap is a no-op there.
Once hermes is running again, the user can ``hermes update`` to fully
recover.
## Test update
tests/test_hermes_bootstrap.py::test_entry_point_imports_bootstrap
scans for the first top-level import in each entry point and asserts it
is hermes_bootstrap. Extended the check to accept a Try block whose body
is a lone Import of hermes_bootstrap — that's the recovery-friendly form
we just introduced.
Verified behavior by ``mv hermes_bootstrap.py hermes_bootstrap.py.bak``
and confirming ``python -c "import hermes_cli.main"`` succeeds. 82/82
tests pass (hermes_bootstrap + windows-native + windows-compat).
Closes the last Python-on-Windows UTF-8 exposure by making every
text-mode open() call explicit about its encoding.
Before: on Windows, bare open(path, 'r') defaults to the system
locale encoding (cp1252 on US-locale installs). That means reading
any config/yaml/markdown/json file with non-ASCII content either
crashes with UnicodeDecodeError or silently mis-decodes bytes.
After: all 89 affected call sites in production code now pass
encoding='utf-8' explicitly. Works identically on every platform
and every locale, no surprise behavior.
Mechanical sweep via:
ruff check --preview --extend-select PLW1514 --unsafe-fixes --fix --exclude 'tests,venv,.venv,node_modules,website,optional-skills, skills,tinker-atropos,plugins' .
All 89 fixes have the same shape: open(x) or open(x, mode) became
open(x, encoding='utf-8') or open(x, mode, encoding='utf-8'). Nothing
else changed. Every modified file still parses and the Windows/sandbox
test suite is still green (85 passed, 14 skipped, 0 failed across
tests/tools/test_code_execution_windows_env.py +
tests/tools/test_code_execution_modes.py + tests/tools/test_env_passthrough.py +
tests/test_hermes_bootstrap.py).
Scope notes:
- tests/ excluded: test fixtures can use locale encoding intentionally
(exercising edge cases). If we want to tighten tests later that's
a separate PR.
- plugins/ excluded: plugin-specific conventions may differ; plugin
authors own their code.
- optional-skills/ and skills/ excluded: skill scripts are user-authored
and we don't want to mass-edit them.
- website/ and tinker-atropos/ excluded: vendored / generated content.
46 files touched, 89 +/- lines (symmetric replacement). No behavior
change on POSIX or on Windows when the file is ASCII; bug fix on
Windows when the file contains non-ASCII.
Codebase-wide fix for Python-on-Windows UTF-8 footguns, complementing
the earlier execute_code sandbox fixes (which remain load-bearing for
when the sandbox explicitly scrubs child env).
Problem: Python on Windows has two long-standing text-encoding pitfalls:
1. sys.stdout/stderr are bound to the console code page (cp1252 on
US-locale installs) — print('café') crashes with UnicodeEncodeError.
2. Subprocess children don't know to use UTF-8 unless PYTHONUTF8 and/or
PYTHONIOENCODING are set in their env — so any Python we spawn
(linters, sandbox children, delegation workers) hits the same bug.
Solution: A tiny bootstrap module (hermes_bootstrap.py) imported as the
first statement of every Hermes entry point:
- hermes_cli/main.py (hermes / hermes-agent console_script)
- run_agent.py (hermes-agent direct)
- acp_adapter/entry.py (hermes-acp)
- gateway/run.py (messaging gateway)
- batch_runner.py (parallel batch mode)
- cli.py (legacy direct-launch CLI)
On Windows, the bootstrap:
- os.environ.setdefault('PYTHONUTF8', '1') (PEP 540 UTF-8 mode)
- os.environ.setdefault('PYTHONIOENCODING', 'utf-8')
- sys.stdout/stderr/stdin.reconfigure(encoding='utf-8', errors='replace')
Children inherit the env vars → they run in UTF-8 mode.
Current process's stdio is reconfigured → print('café') works now.
On POSIX (Linux/macOS), the bootstrap is a complete no-op. We don't
touch LANG, LC_*, or anything else — users who have intentionally
configured a non-UTF-8 locale aren't affected. POSIX systems are
already UTF-8 by default in 99% of modern setups, so there's nothing
to fix.
setdefault() (not overwrite) means users who explicitly set PYTHONUTF8=0
or PYTHONIOENCODING=cp1252 in their environment are respected.
What this does NOT fix: bare open(path, 'w') calls in the *parent*
process still default to locale encoding because PYTHONUTF8 is only
read at interpreter init. A ruff PLW1514 sweep (separate follow-up)
will add explicit encoding='utf-8' at those ~219 call sites for
belt-and-suspenders.
Tests (17): 16 passed, 1 skipped on Windows.
- Windows: env vars set, stdio reconfigured, child inherits UTF-8 mode
- POSIX: complete no-op (verified on fake POSIX + skipped on real
POSIX since we don't have a Linux box in this session)
- Idempotence: multiple calls safe
- Graceful degradation: non-reconfigurable streams don't crash
- User opt-out: explicit PYTHONUTF8=0 is respected
- Load order: every entry point's FIRST top-level import is
hermes_bootstrap, enforced by an AST-level parametrized test
pyproject.toml: added hermes_bootstrap to py-modules so it ships with
pip installs.
Follow-up to #15328's vision-unsupported retry branch in run_agent.py.
_strip_images_from_messages() previously deleted any message whose content
was entirely images. That's fine for synthetic user messages injected for
attachment delivery, but it breaks providers for tool-role messages — the
paired tool_call_id on the preceding assistant message ends up unmatched,
which OpenAI-compatible APIs reject with HTTP 400.
Fix: tool-role messages whose content becomes empty are replaced with a
plaintext placeholder that preserves the tool_call_id linkage. Only
non-tool messages are dropped. Added 10 tests covering the role-alternation
invariants + image-type coverage.
Image-rejection detector: expanded phrase list (image content not
supported / multimodal input / vision input / model does not support
image) and gated on 4xx status so transient 5xx errors never get
misinterpreted as 'server said no to images'. Detection is documented as
best-effort English phrase matching.
AUTHOR_MAP: mapped 3820588+ddupont808@users.noreply.github.com to
ddupont808 so release notes attribute the salvage correctly.
Tool handlers (e.g. computer_use capture) return a _multimodal envelope
dict when a screenshot is attached. The tool-message builder was passing
this raw dict as the `content` field of role:tool messages, which is an
illegal format — OpenAI-compatible APIs expect a string or a content-parts
list, not a plain Python dict, and would reject it with a 400/422 error.
Fix: unwrap _multimodal results to their `content` list
([{type:text,...},{type:image_url,...}]) in both the parallel and
sequential tool-call paths. The Anthropic adapter already handles content
lists natively; vision-capable OpenAI-compatible servers (mlx-vlm,
GPT-4o, etc.) accept image_url parts in tool messages directly.
Also add a _vision_supported adaptive fallback: on first image-rejection
error ("Only 'text' content type is supported." etc.) the agent strips all
image parts from the message history and retries with text only, so
text-only endpoints degrade gracefully without crashing the session.
Extends the cua-driver computer-use backend to drive backgrounded macOS
windows without stealing keyboard or mouse focus from the foreground app.
All changes target the cua-driver MCP backend and the shared dispatcher.
## cua_backend.py
**Window-aware capture**: capture() now calls list_windows + get_window_state
instead of the removed capture tool. Prefers structuredContent.windows
(MCP 2024-11-05+ cua-driver) for zero-parse window enumeration; falls back
to regex-parsed text for older builds. Stores the selected (pid, window_id)
as sticky context so subsequent action calls do not need a redundant round-trip.
**Action routing**: click/scroll/type_text/key all carry the sticky pid
(and window_id for element-indexed clicks). type_text routes through
type_text_chars (individual key events) rather than AX attribute write --
WebKit AXTextFields reject attribute writes from backgrounded processes.
**Key parsing**: _parse_key_combo splits cmd+s-style strings into
(key, [modifiers]) and routes to hotkey (modifier present) or
press_key (bare key) -- cua-driver actual tool names.
**set_value method**: new set_value(value, element) calls the cua-driver
set_value MCP tool. For AXPopUpButton / HTML select in a backgrounded Safari,
AXPress opens the native macOS popup which closes immediately when the app is
non-frontmost; set_value AX-presses the matching child option directly
(no menu required, no focus steal).
**focus_app**: reimplemented as a pure window-selector (enumerates
list_windows, sets sticky pid/window_id) without ever raising the window
or stealing focus.
**list_apps**: fixed tool name from listApps to list_apps; handles plain-text
response via regex when structured data is absent.
**Structured-content extraction**: _extract_tool_result now surfaces
structuredContent from MCP results, enabling the list_windows window array
without text parsing.
**Helpers**: _parse_windows_from_text, _parse_elements_from_tree,
_split_tree_text, _parse_key_combo extracted as module-level functions.
## schema.py
Added set_value to the action enum with a description explaining when to
prefer it over click (select/popup elements, sliders, no focus steal).
Added value field for set_value payloads.
## tool.py
Routed set_value action through _dispatch to backend.set_value.
Added set_value to _DESTRUCTIVE_ACTIONS (approval-gated).
Fixed MIME-type detection in _capture_response: cua-driver may return
JPEG; detect from base64 magic bytes (/9j/ -> image/jpeg, else image/png)
rather than hardcoding image/png.
## agent/display.py + run_agent.py
Guard _detect_tool_failure and result-preview logic against non-string
function_result values: multimodal tool results (dicts with _multimodal=True)
are not string-sliceable; treat them as successes and fall back to str()
for length/preview.
Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.
Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.
- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
Actions: capture (som/vision/ax), click, double_click, right_click,
middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
`content: [text, image_url]` parts) that flows through
handle_function_call into the tool message. Anthropic adapter converts
into native `tool_result` image blocks; OpenAI-compatible providers
get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
recent screenshots carry real image data; older ones become text
placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
instead of its base64 char length (~1MB would have registered as
~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
(curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
entries.
44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.
469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.
- `model_tools.py` provider-gating: the tool is available to every
provider. Providers without multi-part tool message support will see
text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
client-side eviction + compressor pruning cover the same cost ceiling
without a beta header.
- macOS only. cua-driver uses private SkyLight SPIs
(SLEventPostToPid, SLPSPostEventRecordTo,
_AXObserverAddNotificationAndCheckRemote) that can break on any macOS
update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
prints the Settings path.
Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.
When switching from a custom local provider (e.g. ollama-launch) to a
cloud provider, two bugs caused the CLI to misbehave:
1. _explicit_api_key/_explicit_base_url were only updated when the switch
result had non-empty values (guarded by `if result.api_key:` etc.).
If the previous provider set these to Ollama values ("ollama",
"http://127.0.0.1:11434/v1"), those stale values leaked into the next
turn's _ensure_runtime_credentials() call and were forwarded to the
new provider's API endpoint, causing authentication/routing failures.
Fix: unconditionally write result.api_key/base_url into the explicit
fields after every successful switch. An empty string is the correct
sentinel — it tells _ensure_runtime_credentials to re-resolve from the
auth store / config rather than forwarding a stale override.
2. In AIAgent.switch_model(), `self.base_url = base_url or self.base_url`
kept the old Ollama localhost URL whenever the incoming base_url was an
empty string. For providers that use a native SDK (not an OpenAI-compat
endpoint), the caller passes base_url="" and expects the agent to clear
the field — not silently inherit Ollama's address.
Fix: only update self.base_url when base_url is truthy.
3. _handle_model_picker_selection() was called from the prompt_toolkit
Enter key binding without any exception guard. Any unexpected error
in the model-selection code path propagated through prompt_toolkit's
key-binding dispatcher and caused the entire TUI to exit — which the
user sees as "the terminal exits when I switch providers".
Fix: wrap the call in try/except and close the picker on failure.
- Add pricing entries for Claude Opus 4.5/4.6/4.7, Sonnet 4.5/4.6, and
Haiku 4.5 with updated source URLs (platform.claude.com)
- Add _normalize_anthropic_model_name() to handle dot-notation variants
(e.g. claude-opus-4.7 → claude-opus-4-7) for pricing lookups
- Fix silent token loss: ensure session row exists before UPDATE in both
run_agent.py and hermes_state.py (INSERT OR IGNORE is idempotent)
- Log token persistence failures at DEBUG level instead of swallowing
them silently — makes undercounted analytics diagnosable
- Surface reasoning tokens in CLI /usage and TUI usage panel
- Add 'reasoning' and 'cost_status' fields to TUI Usage type
When empty-response terminal scaffolding fires on a tool-result turn,
_drop_trailing_empty_response_scaffolding left the live history ending at
a bare 'tool' message. The next user input then landed as [...tool, user],
a protocol-invalid sequence that OpenRouter/Opus and other providers
silently fail on (returns empty content). That retriggered the empty-retry
recovery every turn, and recovery flags never hit SQLite (no column for
them), so history kept looking broken on every reload.
Two fixes:
1. Scaffolding strip rewinds the orphan assistant(tool_calls)+tool pair
after popping sentinels. Only fires when scaffolding flags were
actually present, so mid-iteration tool loops are untouched.
2. _repair_message_sequence runs right before every API call as a
defensive belt: drops stray tool messages with unknown tool_call_ids,
merges consecutive user messages so no user input is lost. Does NOT
rewind assistant(tool_calls)+tool+user — that pattern is valid when
the user redirected before the model got its continuation turn.
Repro: session 20260507_044111_fa7e65. Opus-4.7/OpenRouter returned
content-less response after a 42KB execute_code output, nudge+retry
chain exhausted (no fallback configured), terminal sentinel appended,
scaffolding stripped leaving bare tool tail, user typed 'wtf happened..'
and landed as tool→user violation. Every subsequent turn collapsed in
<50ms with the same 3-retry empty chain because the API request itself
was malformed.
Verified live via HTTP mock: pre-fix reproduced 5 api_calls/0.15s exit
'empty_response_exhausted'; post-fix 1 api_call/0.10s exit
'text_response(finish_reason=stop)'. Three-turn session flows cleanly
through the scenario. Full run_agent suite: 1242 passed (0 regressions,
2 pre-existing concurrent_interrupt failures unrelated).
Enables plugins to transform LLM output text after generation,
useful for vocabulary/personality transformation without burning
inference tokens.
Follows same pattern as transform_tool_result and transform_terminal_output:
- First non-empty string result wins
- Fail-open: exceptions logged as warnings, agent continues
- Signature: (response_text, session_id, model, platform)
Replaces the per-directory shadow-repo design with a single shared shadow
git store at ~/.hermes/checkpoints/store/. Object DB is now deduplicated
across every working directory the agent has ever touched; a dozen
worktrees of the same project cost near-zero in additional disk.
Why
---
Pre-v2 design had three compounding problems that let ~/.hermes/checkpoints/
grow to multi-GB on active machines:
1. Each working directory got its own full shadow git repo — no object
dedup across projects or across worktrees of the same project.
2. _prune() was a documented no-op: max_snapshots only limited the
/rollback listing. Loose objects accumulated forever.
3. Defaults: enabled=True, auto_prune=False — users paid the disk cost
without ever asking for /rollback.
Field report on a single workstation: 847 MB across 47 shadow repos,
mostly redundant clones of the hermes-agent source tree.
Changes
-------
- tools/checkpoint_manager.py: full rewrite. Single bare store, per-project
refs (refs/hermes/<hash>), per-project indexes (store/indexes/<hash>),
per-project metadata (store/projects/<hash>.json with workdir +
created_at + last_touch). On first v2 init, any pre-v2 per-directory
shadow repos are auto-migrated into legacy-<timestamp>/ so the new
store starts clean. _prune() now actually rewrites the per-project ref
to the last max_snapshots commits and runs git gc --prune=now. New
_enforce_size_cap() drops oldest commits round-robin across projects
when the store exceeds max_total_size_mb. _drop_oversize_from_index()
filters any single file larger than max_file_size_mb out of the snapshot.
- hermes_cli/checkpoints.py: new 'hermes checkpoints' CLI
(status / list / prune / clear / clear-legacy) for managing the store
outside a session.
- hermes_cli/config.py: flipped defaults — enabled=False, max_snapshots=20,
auto_prune=True. Added max_total_size_mb=500, max_file_size_mb=10.
Tightened DEFAULT_EXCLUDES (added target/, *.so/*.dylib/*.dll,
*.mp4/*.mov, *.zip/*.tar.gz, .worktrees/, .mypy_cache/, etc.).
- run_agent.py / cli.py / gateway/run.py: thread the new kwargs through
AIAgent and the startup auto_prune hooks.
- Tests rewritten to match v2 storage while keeping backwards-compat
coverage for the pre-v2 prune path (per-directory shadow repos under
base/ are still swept correctly for anyone mid-migration).
- Docs updated: user-guide/checkpoints-and-rollback.md explains the
shared store, new defaults, migration, and the new CLI;
reference/cli-commands.md documents 'hermes checkpoints'.
E2E validated
-------------
- Legacy migration: pre-v2 shadow repos auto-archived into legacy-<ts>/.
- Object dedup: two projects with an identical shared.py blob resolve to
7 total objects in the store (v1 would have stored the blob twice).
- max_snapshots=3 actually enforced: after 6 commits, list shows 3.
- Orphan prune: deleting a project's workdir + 'hermes checkpoints prune
--retention-days 0' removes its ref, index, and metadata; GC reclaims
the objects.
- max_file_size_mb=1 excludes a 2 MB weights.bin while keeping the
tracked source code files.
- hermes checkpoints {status,prune,clear,clear-legacy} all work from the
CLI without an agent running.
Breaking / migration
--------------------
No in-place data migration — legacy per-directory shadow repos are moved
into legacy-<timestamp>/ on first run. Old /rollback history is still
accessible by inspecting the archive with git; run
'hermes checkpoints clear-legacy' to reclaim the space when ready. Users
relying on /rollback must now set checkpoints.enabled=true (or pass
--checkpoints) explicitly.
Introduces providers/ package — single source of truth for every
inference provider. Adding a simple api-key provider now requires one
providers/<name>.py file with zero edits anywhere else.
What this PR ships:
- providers/ package (ProviderProfile ABC + 33 profiles across 4 api_modes)
- ProviderProfile declarative fields: name, api_mode, aliases, display_name,
env_vars, base_url, models_url, auth_type, fallback_models, hostname,
default_headers, fixed_temperature, default_max_tokens, default_aux_model
- 4 overridable hooks: prepare_messages, build_extra_body,
build_api_kwargs_extras, fetch_models
- chat_completions.build_kwargs: profile path via _build_kwargs_from_profile,
legacy flag path retained for lmstudio/tencent-tokenhub (which have
session-aware reasoning probing that doesn't map cleanly to hooks yet)
- run_agent.py: profile path for all registered providers; legacy path
variable scoping fixed (all flags defined before branching)
- Auto-wires: auth.PROVIDER_REGISTRY, models.CANONICAL_PROVIDERS,
doctor health checks, config.OPTIONAL_ENV_VARS, model_metadata._URL_TO_PROVIDER
- GeminiProfile: thinking_config translation (native + openai-compat nested)
- New tests/providers/ (79 tests covering profile declarations, transport
parity, hook overrides, e2e kwargs assembly)
Deltas vs original PR (salvaged onto current main):
- Added profiles: alibaba-coding-plan, azure-foundry, minimax-oauth
(were added to main since original PR)
- Skipped profiles: lmstudio, tencent-tokenhub stay on legacy path (their
reasoning_effort probing has no clean hook equivalent yet)
- Removed lmstudio alias from custom profile (it's a separate provider now)
- Skipped openrouter/custom from PROVIDER_REGISTRY auto-extension
(resolve_provider special-cases them; adding breaks runtime resolution)
- runtime_provider: profile.api_mode only as fallback when URL detection
finds nothing (was breaking minimax /v1 override)
- Preserved main's legacy-path improvements: deepseek reasoning_content
preserve, gemini Gemma skip, OpenRouter response caching, Anthropic 1M
beta recovery, etc.
- Kept agent/copilot_acp_client.py in place (rejected PR's relocation —
main has 7 fixes landed since; relocation would revert them)
- _API_KEY_PROVIDER_AUX_MODELS alias kept for backward compat with existing
test imports
Co-authored-by: kshitijk4poor <82637225+kshitijk4poor@users.noreply.github.com>
Closes#14418
Each auxiliary model must be resolved with its own provider so that
provider-specific paths (e.g. Bedrock static table, OpenRouter API)
are invoked for the correct client, not inherited from the main model.
When the main model is Bedrock, passing self.provider unconditionally
to get_model_context_length() for the aux model caused the Bedrock
static table hard-intercept (step 1b) to fire for non-Bedrock models,
returning BEDROCK_DEFAULT_CONTEXT_LENGTH=128K instead of the model's
real context window — triggering a false compression warning every session.
Fix: pass _aux_cfg_provider when explicitly set, falling back to
self.provider only when the aux provider is unset or "auto".
Closes#12977
Related: #13807, #17460
The reasoning-box extraction loop in run_conversation() walked backwards
through the entire message history looking for any assistant message
with a non-empty 'reasoning' field. When the current turn produced
no reasoning (e.g. the provider returned reasoning_content=null for a
trivial response), the loop walked past the current turn and showed
reasoning from a prior turn — stale text from minutes or hours ago
displayed as if it belonged to the current reply.
Fix: stop the walk at the user message that started the current turn.
That picks the most recent reasoning WITHIN the turn (correct for
tool-calling turns where reasoning lands on the tool-call step and
the final-answer step has reasoning=None — common on Claude thinking,
DeepSeek v4, Codex Responses), and returns None cleanly when the
current turn genuinely had no reasoning.
Co-authored-by: happy5318 <happy5318@users.noreply.github.com>
* revert(gateway): remove stale-code self-check and auto-restart
Removes the _detect_stale_code / _trigger_stale_code_restart mechanism
introduced in #17648 and iterated in #19740. On every incoming message
the gateway compared the boot-time git HEAD SHA to the current SHA on
disk, and if they differed it would reply with
Gateway code was updated in the background --
restarting this gateway so your next message runs
on the new code. Please retry in a moment.
and then kick off a graceful restart. This is unwanted behaviour:
users who run a long-lived gateway and do their own ad-hoc git
operations on the checkout end up with their chat interrupted and
the current message dropped every time HEAD moves, with no way to
opt out.
If an operator really needs the old protection against stale
sys.modules after "hermes update", the SIGKILL-survivor sweep in
hermes update (hermes_cli/main.py, also tagged #17648) already
handles the supervisor-respawn case on its own.
Removed:
gateway/run.py:
- _STALE_CODE_SENTINELS, _GIT_SHA_CACHE_TTL_SECS
- _read_git_head_sha(), _compute_repo_mtime() module helpers
- class-level _boot_wall_time / _boot_repo_mtime / _boot_git_sha /
_stale_code_restart_triggered defaults
- __init__ boot-snapshot block (_boot_*, _cached_current_sha*,
_repo_root_for_staleness, _stale_code_notified)
- _current_git_sha_cached(), _detect_stale_code(),
_trigger_stale_code_restart() methods
- stale-code check + user-facing restart notice at the top of
_handle_message()
tests/gateway/test_stale_code_self_check.py (deleted, 412 lines)
No new logic added. Zero remaining references to any removed
symbol. Gateway test suite passes the same 4589 tests it passed
before; the 3 pre-existing unrelated failures (discord free-channel,
feishu bot admission, teams typing) are unchanged by this commit.
* fix(agent): stateful streaming scrubber for reasoning-block leaks (#17924)
Per-delta _strip_think_blocks ran at _fire_stream_delta and destroyed
downstream state. When MiniMax-M2.7 / DeepSeek / Qwen3 streamed a tag
split across deltas (delta1='<think>', delta2='Let me check'), the
regex case-2 match erased delta1 entirely, so CLI/gateway state
machines never learned a block was open and leaked delta2 as content.
Raw consumers (ACP, api_server, TTS) had no downstream defense at all.
Replace the per-delta regex with a stateful StreamingThinkScrubber
that survives delta boundaries:
- Closed <tag>X</tag> pairs always stripped (matches _strip_think_blocks
case 1).
- Unterminated open at block boundary enters a block; content
discarded until close tag arrives. At end-of-stream, held
content is dropped.
- Orphan close tags stripped without boundary gating.
- Partial tags at delta boundaries held back until resolved.
- Block-boundary rule (start-of-stream, after \n, or
whitespace-only since last \n) preserves prose that mentions
tag names.
Reset at turn start alongside the existing context scrubber; flush at
turn end so a benign '<' held back at end-of-stream reaches the UI.
E2E-verified on live OpenRouter->MiniMax-m2 streams: closed pairs
strip cleanly, first word of post-block content is preserved, pure
content passes through unchanged. Stefan's screenshot case (#17924)
— 'Let me check' getting chopped to ' me check' — no longer happens.
Final _strip_think_blocks calls on completed strings (final_response,
replay, compression) are preserved; only the streaming per-delta call
site switched to the scrubber.
MCP servers commonly emit JSON Schema `pattern` (e.g. `\\d{4}-\\d{2}-\\d{2}`
for date-time params) and `format` keywords. llama.cpp's
`json-schema-to-grammar` converter rejects regex escape classes
(\\d/\\w/\\s) and most format values, returning HTTP 400
"parse: error parsing grammar: unknown escape at \\d" — the whole request
fails.
Cloud providers (OpenAI, Anthropic, OpenRouter, Gemini) accept these
keywords fine and use them as prompting hints. Stripping unconditionally
loses useful hints for every cloud user to fix a llama.cpp-only bug.
Approach: classify the llama.cpp grammar-parse 400 in the error
classifier, and on match do a one-shot in-place strip of pattern/format
from `self.tools`, then retry. Follows the existing
`thinking_signature` recovery pattern. Cloud users hit zero overhead;
llama.cpp users pay one failed request per session.
Changes
- agent/error_classifier.py: new `FailoverReason.llama_cpp_grammar_pattern`
+ narrow HTTP-400 branch matching "error parsing grammar",
"json-schema-to-grammar", or "unable to generate parser ... template".
- tools/schema_sanitizer.py: new `strip_pattern_and_format()` helper —
reactive, walks schema nodes, skips property names (search_files.pattern
survives). Returns strip count for logging.
- run_agent.py: new one-shot recovery block in the retry loop. Strips,
logs, continues. Falls through to normal retry if nothing to strip.
- tests: 4 classifier tests (3 variants + 1 non-400 negative), 7 strip
tests including the property-name preservation and idempotency checks.
Co-authored-by: Chris Danis <cdanis@gmail.com>
The `used` property was reading `self._used` without holding the lock,
while `consume()`, `refund()`, and `remaining` all properly acquire
`self._lock` before accessing `_used`. This means a concurrent call to
`used` during `consume()` or `refund()` could observe a partially-
updated value, leading to incorrect iteration budget metrics reported
to the gateway, or in extreme cases a ValueError from CPython's list
implementation when the internal array resizes during iteration.
Fix: acquire the lock in `used` just like `remaining` does.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Gemini's OpenAI-compatibility endpoint strictly requires the `name` field
on `role: tool` messages — it returns HTTP 400 ("Request contains an
invalid argument") when the function name is missing. OpenAI/Anthropic/
ollama tolerate the absence, so the gap stays invisible until the
conversation accumulates a tool turn and the user routes it through Gemini
(direct API or via ollama-cloud proxy).
Fix: add a `_get_tool_call_name_static()` helper alongside the existing
`_get_tool_call_id_static()`, and populate `name` at every site that
constructs a `role: tool` message — the pre-call sanitizer stub, the
tool-call args repair marker, both interrupt-skip paths, both
result-append paths (parallel + sequential), the invalid-tool-name
recovery, the invalid-JSON-args recovery, and the exception fallback.
Each call site was already in scope of the function name (`function_name`,
`skipped_name`, `name`, or a dict tool_call), so the change is local —
no new lookups, no behavior change for providers that already worked.
Fixes#16478
The background memory/skill review fork had two user-visible issues:
1. max_iterations=8 was too tight for multi-step reviews. A review that
needs to skill_view one or two candidate skills, add a memory entry,
and patch a skill routinely blew the budget — surfacing an 'Iteration
budget exhausted (8/8)' warning to the user and leaving the review
half-finished.
2. Mid-review lifecycle messages leaked into the user's terminal past the
existing quiet_mode + redirect_stdout/stderr guards. _emit_status and
_emit_warning route through _vprint(force=True) -> _print_fn /
status_callback, which bypass sys.stdout entirely. The stdout redirect
only catches raw print() calls.
Changes:
- Bump the review fork's max_iterations from 8 to 16.
- Set review_agent.suppress_status_output = True on the fork. This
short-circuits _vprint unconditionally so _emit_status/_emit_warning
emissions (iteration-budget warnings, rate-limit retries, compression
messages) never reach the user. The only user-visible output remains
the compact final summary line ('💾 Self-improvement review: ...')
which is printed via self._safe_print on the *main* agent (outside
the fork's redirect/suppress scope).
Summarizer filter is already correct — _summarize_background_review_actions
only surfaces tool calls with data.get('success') is truthy, so failed
attempts and reasoning text never reach the summary line.
Add 'xiaomi' to the _anthropic_preserve_dots() provider whitelist and
'xiaomimimo.com' to the URL-based fallback check. Without this,
normalize_model_name() converts mimo-v2.5 to mimo-v2-5, which the
Xiaomi API rejects with HTTP 400.
Fixes#16156
Ollama serves Qwen3 thinking inside the content field as <think>...</think>
blocks rather than in the API-level reasoning_content field. This means
_has_structured was False for these responses, so an empty-looking reply
after a tool call triggered the nudge instead of the prefill continuation,
causing a double-response loop.
Fix: detect <think>/<thinking>/<reasoning> in final_response and:
1. Skip the nudge when thinking is present (model is still reasoning)
2. Include _has_inline_thinking in _has_structured so prefill kicks in
Per-request OpenAI-wire clients (used by both non-streaming and
streaming chat-completions paths in _interruptible_api_call) should
not run the SDK's built-in retry loop: the agent's outer loop owns
retries with credential rotation, provider fallback, and backoff that
the SDK can't see.
Leaving SDK retries on (default 2) compounds with our outer retries
and lets a single hung provider request stretch to ~3x the per-call
timeout before our stale detector reports it.
Shared/primary clients and Anthropic / Bedrock paths are unaffected
(they don't go through here).
Salvage of #15811 core improvement — the timeout push-down in the
original PR required scaffolding that has since been refactored on
main, so only the max_retries=0 change is preserved.
Co-authored-by: QifengKuang <k2767567815@gmail.com>
Tighten the provenance semantics added in #19618: skills a user asks a
foreground agent to write via skill_manage(create) now stay invisible to
the curator. Only skills the background self-improvement review fork
sediments through skill_manage get the created_by=agent marker.
- tools/skill_provenance.py — new ContextVar module mirroring the
_approval_session_key pattern: set_current_write_origin / reset /
get / is_background_review. Default origin is 'foreground'; the
review fork sets 'background_review'.
- run_agent.py — run_conversation() binds the ContextVar from
self._memory_write_origin at the top of each call. The review fork
runs on its own thread (fresh context), so foreground and review
contexts never cross-contaminate.
- tools/skill_manager_tool.py — skill_manage(action='create') now
only calls mark_agent_created() when is_background_review(). All
other cases (foreground create, patch, edit, write_file, delete)
continue as before.
- tests: test_skill_provenance.py (6 tests covering the ContextVar
surface), split test_full_create_via_dispatcher into foreground
vs. review-fork variants, curator status tests now mark-first.
Why: the agent routinely edits existing user skills on the user's
behalf; those writes must never flip provenance. And when a user
explicitly asks the foreground agent to create a skill, that skill
belongs to the user. The curator should only be cleaning up after
its own autonomous sediment from the review nudge loop.
Preflight compression can run synchronously before the first model call when a loaded session exceeds the active context threshold. Gateway users saw no visible progress while the compression LLM call was in flight, which can look like a dropped message during long compactions.\n\nEmit the existing lifecycle status through _emit_status before starting preflight compression so CLI, gateway, and WebUI status callbacks all get immediate feedback.\n\nAdds a regression assertion for the preflight path.
Enable OpenRouter's response caching feature (beta) via X-OpenRouter-Cache
headers. When enabled, identical API requests return cached responses for
free (zero billing), reducing both latency and cost.
Configuration via config.yaml:
openrouter:
response_cache: true # default: on
response_cache_ttl: 300 # 1-86400 seconds
Changes:
- Add openrouter config section to DEFAULT_CONFIG (response_cache + TTL)
- Add build_or_headers() in auxiliary_client.py that builds attribution
headers plus optional cache headers based on config
- Replace inline _OR_HEADERS dicts with build_or_headers() at all 5 sites:
run_agent.py __init__, _apply_client_headers_for_base_url(), and
auxiliary_client.py _try_openrouter() + _to_async_client()
- Add _check_openrouter_cache_status() method to AIAgent that reads
X-OpenRouter-Cache-Status from streaming response headers and logs
HIT/MISS status
- Document in cli-config.yaml.example
- Add 28 tests (22 unit + 6 integration)
Ref: https://openrouter.ai/docs/guides/features/response-caching
When a provider's credential pool has a single entry in 429-cooldown,
resolve_provider_client returns None and AIAgent.__init__ raises a
misleading RuntimeError suggesting the API key is missing — even when
valid fallback_providers are configured.
This patch makes __init__ iterate the fallback chain before raising,
mirroring the existing in-flight fallback logic in the request loop.
If a fallback resolves, the agent initializes against it and sets
_fallback_activated=True so _restore_primary_runtime can pick the
primary back up after cooldown.
Closes#17929
Prevents ghost sessions from accumulating in state.db when the TUI/web
dashboard is opened and closed without sending a message.
Changes:
- run_agent.py: Add _ensure_db_session() gate method, called at
run_conversation() entry. Remove eager create_session() from __init__.
Handle compression rotation flag correctly.
- tui_gateway/server.py: Remove eager db.create_session() in
_start_agent_build(). Add post-first-message pending_title re-apply.
- hermes_state.py: Extract _insert_session_row() shared helper (DRY).
Add prune_empty_ghost_sessions() for one-time migration.
- cli.py: One-time ghost session prune on startup. Fix _pending_title
to call _ensure_db_session() before set_session_title().
- hermes_cli/main.py: Guard TUI exit summary on message_count > 0.
- tests: Update test_860_dedup to call _ensure_db_session() before
direct _flush_messages_to_session_db() calls.
Closes: ghost session clutter in hermes sessions list and web dashboard.
DeepSeek V4 Pro tightened thinking-mode validation and rejects empty-string
reasoning_content with HTTP 400:
The reasoning content in the thinking mode must be passed back to the API.
run_agent.py injected "" at three fallback sites — the tool-call pad in
_build_assistant_message and both injection branches of
_copy_reasoning_content_for_api (cross-provider poison guard + unconditional
thinking pad). All three now emit " " (single space), which satisfies the
non-empty check on V4 Pro without leaking fabricated reasoning.
Also upgrades stale empty-string placeholders on replay: sessions persisted
before this change have reasoning_content="" pinned at creation time; when
the active provider enforces thinking-mode echo, the replay path now rewrites
"" -> " " so existing users don't 400 on their first V4 Pro turn after
updating. Non-thinking providers still round-trip "" verbatim.
Updates 9 existing assertions + adds 2 regression tests (stale-placeholder
upgrade, non-thinking verbatim preservation).
Refs #15250, #17400.
Closes#17341.
The user-visible /compress banner and the post-compression last_prompt_tokens
writeback both counted only the raw message transcript (chars/4). With a 15KB
system prompt and 30 tool schemas (~26KB), a 4-message transcript that looks
like ~45 tokens to the transcript-only estimator is really ~10.5K tokens of
request pressure — a 234x gap.
Two user-facing consequences:
- Banner shows 'Compressing … (~45 tokens)…' while compression is actually
firing on 10K+ tokens of real pressure, confusing users about why
compression triggered (reported by @codecovenant on X; #6217).
- Post-compression last_prompt_tokens writeback omits tool schemas, so the
next should_compress() check compares real usage against a stale
underestimate — compression triggers late, potentially past the model's
context limit on small-context models (#14695).
Swap estimate_messages_tokens_rough() for estimate_request_tokens_rough()
at every user-visible banner and at the post-compression writeback.
estimate_request_tokens_rough() already existed for exactly this purpose
and includes system prompt + tool schemas.
Touched call sites:
- run_agent.py: post-compression last_prompt_tokens writeback, post-tool
call should_compress() fallback when provider usage is missing
- cli.py: /compress banner + summary
- gateway/run.py: gateway /compress banner + summary
- tui_gateway/server.py: TUI /compress status + summary
- acp_adapter/server.py: ACP /compact before/after
Left intentionally alone:
- Session-hygiene fallback and the 'no agent' /status path in gateway/run.py
— no agent instance is in scope to query for system prompt/tools, and the
existing 30-50% overestimate wobble on hygiene is safety-accepted.
- Verbose-mode 'Request size' logging — informational only, already counts
system prompt via api_messages[0].
Also relabels the feedback line from 'Rough transcript estimate' to
'Approx request size' so the metric label matches what it actually measures.
Credits: diagnoses from @devilardis (#14695) and @Jackten (#6217);
user report @codecovenant on X (2026-04-30).
Closes#14695Closes#6217
When the self-improvement background review fires after a turn, it runs
in a bg thread and emits a ' 💾 <summary>' line to announce what it
saved to memory or skills. Two problems made this invisible to users
even when the review successfully modified a skill:
1. The print went through `_cprint` (prompt_toolkit's print_formatted_text)
on a bg thread while the CLI's PromptSession was live. Direct
print_formatted_text races with the input-area redraw and the line
can land behind/above the prompt, scrolled off without the user
seeing it.
2. The message said only '💾 Skill created.' / '💾 Memory updated'
with no indication that the self-improvement loop was the one doing
this. Users who did catch the line couldn't tell the background
review from some other agent action.
Fixes:
- `_cprint` now detects when it's called from a non-app thread with a
running prompt_toolkit Application, and routes through
`run_in_terminal` via `loop.call_soon_threadsafe`. That pauses the
input, prints the line above the prompt, and redraws — the normal
prompt_toolkit contract for bg-thread output. Direct-print fallback
preserved for the no-app / same-thread / import-error paths. Affects
every bg-thread emission, not just the review summary (curator
summaries and auxiliary failure prints benefit too).
- The summary now reads ' 💾 Self-improvement review: <summary>' in
both the CLI and the gateway `background_review_callback` path, so
the origin is unambiguous.
Tests:
- New `tests/cli/test_cprint_bg_thread.py` covers all five routing
branches (no app, app-not-running, cross-thread schedule, same-thread
direct, app-loop-attribute-error, import-error).
- New case in `tests/run_agent/test_background_review.py` asserts the
attributed prefix shows up in both `_safe_print` and
`background_review_callback`.
Live E2E: exercised _cprint from a bg thread inside a real Application
event loop; confirmed get_app_or_none() sees the app, call_soon_threadsafe
schedules run_in_terminal, and the inner _pt_print runs.
Builds on #16855 (@lsdsjy) which fixed DeepSeek v4 reasoning_content
replay via model_extra fallback + capturing tool_calls at method entry.
Kimi / Moonshot thinking mode enforces the same echo-back contract and
hits the same 400 when a tool-call turn is persisted without
reasoning_content.
- _build_assistant_message: pad branch now uses _needs_thinking_reasoning_pad()
(DeepSeek OR Kimi) instead of _needs_deepseek_tool_reasoning() alone.
- Extract _needs_thinking_reasoning_pad() and reuse it in
_copy_reasoning_content_for_api so both sites share one predicate.
- tests/run_agent/test_deepseek_reasoning_content_echo.py: add
TestBuildAssistantMessagePadsStrictProviders parametrized over DeepSeek
(attr=None, attr-absent), Kimi (attr=None), Moonshot (via base_url),
and an OpenRouter negative control that must NOT pad. Proven to fail
2/5 cases on Kimi/Moonshot without this change.
- scripts/release.py: add AUTHOR_MAP entries for lsdsjy and season179.
Refs #17400.
Co-authored-by: season179 <season.saw@gmail.com>
Three fixes bundled for curator reliability on existing installs and
broken/partial installs:
1. run_agent.py: defer `import fire` into the __main__ block. `fire` is
only used by `fire.Fire(main)` when running run_agent.py directly as
a CLI — it is NOT needed for library usage. Importing it at module
top made `from run_agent import AIAgent` from a daemon thread (e.g.
the curator's forked review agent) crash with ModuleNotFoundError
on broken/partial installs where `fire` isn't present.
2. hermes_cli/config.py: add version 22 → 23 migration that writes the
`curator` + `auxiliary.curator` sections to config.yaml with their
defaults, only filling keys the user hasn't overridden. Existing
configs from before PR #16049 / the April 2026 `auxiliary.curator`
unification had neither section on disk, so users couldn't see or
edit the settings in their config.yaml (runtime deep-merge papered
over it at read time, but the file never reflected reality).
3. hermes_cli/config.py: `ensure_hermes_home()` now pre-creates
`~/.hermes/logs/curator/` alongside cron/sessions/logs/memories on
every CLI launch. Managed-mode (NixOS) variant mkdir's it
defensively after the activation-script existence checks, since the
activation script may not know about this subpath.
4. agent/curator.py: `_reports_root()` mkdir's the dir at call time as
belt-and-suspenders for entry paths that bypass both
ensure_hermes_home() and the v23 migration (gateway-only installs,
bare library use).
E2E validated in isolated HERMES_HOME: fresh install gets full defaults
seeded; partial-override config keeps user's `enabled: false` and
custom `interval_hours` while filling the missing keys; re-running the
migration is a no-op.
Long-lived Gateway processes were sending duplicate tool names to
providers that enforce uniqueness:
- DeepSeek: 'Tool names must be unique.'
- Xiaomi MiMo: 'tools contains duplicate names: lcm_expand'
- Moonshot/Kimi: 'function name lcm_grep is duplicated'
TUI was unaffected because TUI runs with quiet_mode=False and skips the
cache entirely.
Root cause (two layered bugs)
- model_tools.get_tool_definitions(quiet_mode=True) memoizes its result
in _tool_defs_cache. The cache-hit path returned list(cached) (safe),
but the FIRST uncached call stored and returned the SAME object.
run_agent.py mutates self.tools (memory + LCM context-engine schemas)
in-place, so the very first agent init in a Gateway process
poisoned the cache, and every subsequent init appended LCM schemas
again on top of the already-polluted list.
- run_agent.py's context-engine injection (lcm_grep / lcm_describe /
lcm_expand) had no dedup, unlike the memory-tools injection right
above it which already skips already-present names.
Fix (defense in depth, per the issue's suggested fix)
- model_tools.get_tool_definitions: on the uncached branch, cache the
computed list but return list(result) to the caller. Same pattern as
the cache-hit path.
- run_agent.py: build _existing_tool_names from self.tools and skip
schemas whose names are already present, mirroring the memory-tools
block. This also defends against plugin paths that may register the
same schemas via ctx.register_tool().
Tests (tests/test_get_tool_definitions_cache_isolation.py)
- test_first_uncached_call_returns_fresh_list \u2014 pins the fix; without
it, first-call alias caused all the symptoms.
- test_cache_hit_returns_fresh_list \u2014 pre-existing behavior stays.
- test_caller_mutation_does_not_poison_cache \u2014 simulates run_agent
appending lcm_grep / lcm_expand to the returned list and asserts the
next call doesn't see them.
- test_repeated_caller_mutation_does_not_accumulate \u2014 reproduces the
long-lived Gateway accumulation pattern across 5 agent inits.
- test_non_quiet_mode_does_not_use_cache \u2014 sanity, explains why TUI
was fine.
5/5 pass on the new file; 23/23 still pass on tests/test_model_tools.py.
When a user sets model.context_length in config.yaml, the value was only
used for Hermes' internal compression decisions (context_compressor) but
NOT for Ollama's num_ctx parameter. Ollama auto-detects context from GGUF
metadata (often 256K+) and allocates that much VRAM regardless of the
user's config — causing OOM on smaller GPUs like the P100 (16GB).
Root cause: two separate context values existed independently:
- context_compressor.context_length = config value (e.g. 65536) ✓
- _ollama_num_ctx = GGUF metadata value (e.g. 256000) ✗ ignored config
Changes:
1. Cap Ollama num_ctx to config context_length (run_agent.py)
When model.context_length is explicitly set and no explicit
ollama_num_ctx override exists, cap the auto-detected GGUF value
to the user's context_length. This is the core fix — it prevents
Ollama from allocating more VRAM than the user budgeted.
2. Pass config_context_length through all secondary call sites
Several paths called get_model_context_length() without the config
override, falling through to the 256K default fallback:
- cli.py: @-reference expansion and /model switch display
- gateway/run.py: @-reference expansion and /model switch display
- tui_gateway/server.py: @-reference expansion
- hermes_cli/model_switch.py: resolve_display_context_length()
3. Normalize root-level context_length in config (hermes_cli/config.py)
_normalize_root_model_keys() now migrates root-level context_length
into the model section, matching existing behavior for provider and
base_url. Users who wrote `context_length: 65536` at the YAML root
instead of under `model:` had it silently ignored.
4. Fix misleading comments (agent/model_metadata.py)
DEFAULT_FALLBACK_CONTEXT is 256K (CONTEXT_PROBE_TIERS[0]), not 128K
as two comments stated.
Tests: 3 new tests for root-level context_length normalization.
All existing context_length tests pass (96 tests).
Keep context-1m-2025-08-07 in OAuth requests by default so 1M-capable
subscriptions retain full context. When Anthropic rejects a request with
400 'long context beta is not yet available for this subscription',
disable the beta for the rest of the session, rebuild the client, and
retry once.
Addresses #17680 (thanks @JayGwod for the clean reproduction) without
forcing every OAuth user off the 1M context window.
Changes:
- agent/error_classifier.py: new FailoverReason.oauth_long_context_beta_forbidden;
pattern matches 400 + 'long context beta' + 'not yet available'. Narrow
enough that the existing 429 tier-gate pattern keeps its own reason.
- agent/anthropic_adapter.py: _common_betas_for_base_url,
build_anthropic_client, build_anthropic_kwargs gain drop_context_1m_beta
kwarg. Default=False (1M stays). OAuth OAUTH_ONLY_BETAS unchanged.
- agent/transports/anthropic.py: build_kwargs forwards the flag.
- run_agent.py: self._oauth_1m_beta_disabled flag, retry-once guard,
recovery branch next to the image-shrink path. _rebuild_anthropic_client
honors the flag. The main build_kwargs call site threads it through for
fast-mode extra_headers.
- hermes_cli/doctor.py, hermes_cli/models.py: sibling OAuth /v1/models
probes get the same reactive retry — previously they'd falsely report
the Anthropic API as unreachable for affected subscriptions.
Tests: 2190 tests/agent/ + 94 adjacent integration tests pass. New unit
tests cover the classifier pattern (including the collision guard against
the 429 tier-gate) and the drop_context_1m_beta adapter behavior (default
keeps 1M, flag strips only 1M while preserving every other beta).
feat(gateway): refine Platform._missing_ and platform-connected dispatch
Restricts plugin-name acceptance to bundled plugin scan + registry
(no arbitrary string -> enum-pollution), pulls per-platform connectivity
checks into a _PLATFORM_CONNECTED_CHECKERS lambda map with a clean
_is_platform_connected method, and adds tests covering the checker map,
plugin platform interface, and IRC setup wizard.
Closes remaining functional gaps and adds documentation.
webhook.py: Cross-platform delivery now checks the plugin registry
for unknown platform names instead of hardcoding 15 names in a tuple.
Plugin platforms can receive webhook-routed deliveries.
prompt_builder: Platform hints (system prompt LLM guidance) now fall
back to the plugin registry's platform_hint field. Plugin platforms
can tell the LLM 'you're on IRC, no markdown.'
PlatformEntry: Added platform_hint field for LLM guidance injection.
IRC adapter: Added acquire_scoped_lock/release_scoped_lock in
connect/disconnect to prevent two profiles from using the same IRC
identity. Added platform_hint for IRC-specific LLM guidance.
Removed dead token-empty-warning extension for plugin platforms
(plugin adapters handle their own env vars via check_fn).
website/docs/developer-guide/adding-platform-adapters.md:
- Added 'Plugin Path (Recommended)' section with full code examples,
PLUGIN.yaml template, config.yaml examples, and a table showing all
18 integration points the plugin system handles automatically
- Renamed built-in checklist to clarify it's for core contributors
gateway/platforms/ADDING_A_PLATFORM.md:
- Added Plugin Path section pointing to the reference implementation
and full docs guide
- Clarified built-in path is for core contributors only
Each AIAgent.__init__() was unconditionally starting a daemon thread to
pre-warm the OpenRouter model metadata cache. In gateway mode a new
AIAgent is created for every incoming message, so one OS thread leaked
per request. After ~1 000 messages the process hit the Linux thread
limit and raised RuntimeError: can't start new thread for all subsequent
requests.
Add a module-level threading.Event (_openrouter_prewarm_done) that is
set before the thread is started. Subsequent AIAgent instantiations
skip the spawn entirely; fetch_model_metadata() is cached for 1 hour so
the single background call is sufficient.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fixes#6672
Memory providers now receive on_session_switch() whenever AIAgent.session_id
rotates mid-process — /resume, /branch, /reset, /new, and context
compression. Before this, providers that cached per-session state in
initialize() (Hindsight's _session_id, _document_id, accumulated
_session_turns, _turn_counter) kept writing into the old session's
record after the agent had moved on.
MemoryProvider ABC
------------------
- New optional hook on_session_switch(new_session_id, *,
parent_session_id='', reset=False, **kwargs) with no-op default for
backward compat. reset=True signals /reset or /new — providers should
flush accumulated per-session buffers. reset=False for /resume,
/branch, compression where the logical conversation continues.
MemoryManager
-------------
- on_session_switch() fans the hook out to every registered provider.
Isolated try/except per provider — one bad provider can't block others.
- Empty/None new_session_id is a no-op to avoid corrupting provider state
during shutdown paths.
run_agent.py
------------
- _sync_external_memory_for_turn now passes session_id=self.session_id
into sync_all() and queue_prefetch_all(). Providers with defensive
session_id updates in sync_turn (Hindsight already had this at
plugins/memory/hindsight/__init__.py:1199) now actually receive the
current id.
- Compression block at ~L8884 already notified the context engine of
the rollover; now also calls
_memory_manager.on_session_switch(reason='compression').
cli.py
------
- new_session() fires reset=True, reason='new_session' so providers
flush buffers.
- _handle_resume_command fires reset=False, reason='resume' with the
previous session as parent_session_id.
- _handle_branch_command fires reset=False, reason='branch' with the
parent session_id already captured for the DB parent link.
gateway/run.py
--------------
- _handle_resume_command now evicts the cached AIAgent, mirroring
/branch and /reset. The next message rebuilds a fresh agent whose
memory provider initialize() runs with the correct session_id —
matches the pattern the gateway already uses for provider state
cross-session transitions.
Hindsight reference implementation
----------------------------------
- plugins/memory/hindsight/__init__.py adds on_session_switch that:
updates _session_id, mints a fresh _document_id (prevents
vectorize-io/hindsight#1303 overwrite), and clears _session_turns /
_turn_counter / _turn_index so in-flight batches don't flush under
the new document id. parent_session_id only overwritten when provided
(avoids clobbering on a bare switch).
Tests
-----
- tests/agent/test_memory_session_switch.py: new dedicated file. ABC
default no-op, manager fan-out, failure isolation, empty-id no-op,
session_id propagation through sync_all/queue_prefetch_all, Hindsight
state transitions for every reset/non-reset case, parent preservation.
- tests/cli/test_branch_command.py: new test verifying /branch fires
the hook with correct parent_session_id + reset=False + reason.
- tests/gateway/test_resume_command.py: new test verifying /resume
evicts the cached agent.
- tests/run_agent/test_memory_sync_interrupted.py: updated existing
assertions to account for the session_id kwarg on sync_all and
queue_prefetch_all.
E2E verified (real imports, tmp HERMES_HOME):
- /resume: session_id updates, doc_id fresh, buffers cleared, parent set
- /branch: session_id forks, parent links to original
- /new: reset=True clears accumulated state
- compression: reason='compression' propagated, lineage preserved
- Empty id: no-op, state preserved
- Legacy provider without on_session_switch: no crash
Reported by @nicoloboschi (Hindsight maintainer); related scope-widening
comment by @kidonng extending coverage to compression.
MiniMax's /anthropic endpoint documents cache_control support (0.1x read
pricing, 5-min TTL) for MiniMax-M2.7, M2.5, M2.1, M2. PR #12846 gated
third-party Anthropic-wire caching on 'claude' in model name, which left
MiniMax's own model family re-paying full input tokens every turn.
Opt in explicitly via provider id (minimax / minimax-cn) or host match
(api.minimax.io / api.minimaxi.com). Narrow allowlist mirroring the
existing Qwen/Alibaba branch below; leaves room for a capability-based
surface (ProviderConfig.supports_anthropic_cache) if a third provider
needs it.
Closes#17332
Completes the cfg_get migration started in PR #17304. Covers the
remaining hermes_cli/ and plugins/ config-access sites that the first
PR intentionally left opportunistic.
Migrated (33 sites across 14 files):
hermes_cli/setup.py 13 sites (terminal.*, agent.*, display.*, compression.*, tts.*)
hermes_cli/tools_config.py 7 sites (tts.*, browser.*, web.*, platform_toolsets.*)
hermes_cli/plugins_cmd.py 3 sites (plugins.*, memory.*, context.*)
plugins/memory/honcho/cli.py 3 sites (hosts.*)
hermes_cli/web_server.py 1 site (dashboard.*)
hermes_cli/skills_config.py 1 site (platform_disabled)
hermes_cli/plugins.py 1 site (plugins.disabled)
hermes_cli/status.py 1 site (terminal.backend)
hermes_cli/mcp_config.py 1 site (mcp_servers.*)
hermes_cli/webhook.py 1 site (platforms.webhook)
plugins/memory/__init__.py 1 site (memory.provider)
plugins/memory/hindsight/ 1 site (banks.hermes)
plugins/memory/holographic/ 1 site (plugins.hermes-memory-store)
run_agent.py 1 site (auxiliary.compression)
The helper supports non-literal keys too, so e.g.
cfg.get('hosts', {}).get(HOST, {})
becomes
cfg_get(cfg, 'hosts', HOST, default={})
Migration bugs caught and fixed during this PR:
1. An AST-based batch rewrite naïvely captured the first word token in
a chain, which corrupted 'self._config.get(...).get(...)' into
'self.cfg_get(_config, ...)' (dropping 'self.', creating a broken
method call). Plugins/memory/hindsight caught it via its test suite.
Fixed manually to 'cfg_get(self._config, ...)'.
2. Import-extension heuristic rewrote multi-line parenthesized imports
('from X import (\n A,\n B,\n)') as
'from X import cfg_get, (' — syntactically broken. Fixed by inserting
cfg_get as the first name inside the parentheses.
Combined with PR #17304, the cfg_get migration now covers:
PR #17304 (first batch): 20 sites in tools/ + gateway/
PR #17317 (this one): 33 sites in hermes_cli/ + plugins/ + run_agent.py
Total: 53 sites migrated. Remaining ~8 sites are either:
- Function-call chains (e.g. '_load_stt_config().get(...).get(...)')
that would need double-evaluation or a local binding to migrate
cleanly — intentionally deferred.
- JSON response-navigation (e.g. 'response_data.get('data',{}).get('web'))
which is unrelated to config access and shouldn't use cfg_get.
Verified:
- 412/412 tests/plugins/ pass (including the hindsight test that caught
the self.X regex bug before commit)
- 3181/3189 tests/hermes_cli/ pass (8 pre-existing failures on main,
verified by git-stash comparison)
- Live 'hermes status' and 'hermes config' render correctly (exercise
the migrated terminal.backend, tts.provider, browser.cloud_provider,
compression.threshold, display.tool_progress sites)
- Live 'hermes chat': 1 turn + /quit, zero errors in 11-line log window
No semantic changes — cfg_get was already proven to be a 1:1 match for
the original .get("X",{}).get("Y",default) pattern in PR #17304.
The background skill-review prompts (_SKILL_REVIEW_PROMPT and the **Skills**
half of _COMBINED_REVIEW_PROMPT) steered the reviewer toward passive
behavior — most passes concluded 'Nothing to save.' even when the session
produced real lessons. User-preference corrections (style, format,
legibility, verbosity) were especially lost: they were read as memory
signals only, so skills never carried the fix.
This rewrite changes the stance:
- **Active-update bias.** The reviewer now treats inaction as a missed
learning opportunity. 'Nothing to save.' remains an explicit escape
but is no longer framed as the most-common outcome.
- **User-preference corrections are first-class skill signals.** Style,
tone, format, legibility, verbosity complaints — and the actual
phrasings users use ('stop doing X', 'this is too verbose', 'I hate
when you Y', 'remember this') — now warrant patching the skill that
governs the task, not just writing to memory.
- **Loaded-skill-first preference order.** When a skill was loaded via
/skill-name or skill_view during the session, the reviewer patches
THAT one first. It was in play; it's the right place.
- **Four-step ladder: patch-loaded → patch-umbrella → support-file →
create.** Support files are explicitly enumerated as three kinds:
* references/<topic>.md — session-specific detail OR condensed
knowledge banks (quoted research, API docs excerpts, domain notes)
* templates/<name>.<ext> — starter files to copy and modify
* scripts/<name>.<ext> — statically re-runnable actions
- **Name-veto for CREATE.** New skill names MUST be class-level — no PR
numbers, error strings, codenames, library-alone names, or session
artifacts ('fix-X / debug-Y / audit-Z-today'). If the proposed name
only fits today's task, fall back to one of the patch/support-file
options.
- **Memory scope clarified.** 'who the user is and what the current
situation and state of your operations are' — MEMORY.md is
situational/state, USER.md is identity/preferences.
- **Curator handoff.** Reviewer flags overlap; the background curator
handles consolidation at scale. Single-session reviewer doesn't
attempt umbrella-rebalancing.
Tests: tests/run_agent/test_review_prompt_class_first.py upgraded to
assert the new behavioral contracts (active bias, user-correction
signals, loaded-skill-first, support-file kinds, name-veto, memory
framing, curator handoff). 17 tests, all pass.
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
- Remove dead _lmstudio_loaded_context attribute from run_agent.py (set
but never read — the loaded context is pushed to context_compressor.update_model
which is the actual consumer)
- Cache empty reasoning options with 60s TTL to avoid per-turn HTTP probe
for non-reasoning LM Studio models. Non-empty results cached permanently.
- Extract _lmstudio_server_root(), _lmstudio_request_headers(), and
_lmstudio_fetch_raw_models() shared helpers in models.py — eliminates
URL-strip + auth-header + HTTP-call duplication across probe_lmstudio_models,
ensure_lmstudio_model_loaded, and lmstudio_model_reasoning_options
- Revert runtime_provider.py base_url precedence change: preserve the
established contract (saved config.base_url > env var > default) for all
api_key providers
- Remove unnecessary config version bump 22→23
- Fix TUI test: relax target_model assertion to avoid module-cache flake
- AUTHOR_MAP: added rugved@lmstudio.ai → rugvedS07
CopilotACPClient communicates via subprocess stdio and returns a plain
SimpleNamespace from _create_chat_completion(). The streaming path tries
to iterate this as a stream, crashing with:
TypeError: 'types.SimpleNamespace' object is not iterable
Mirror the existing ACP exclusion pattern (used for Responses API upgrade)
to disable streaming when provider is copilot-acp or base_url starts with
acp:// or acp+tcp://.
Based on PR #9428 by @ningfangbin and issue #16271 by @Joseph19820124.
Fixes#16271
* perf(startup): lazy-import OpenAI, Anthropic, Firecrawl, account_usage
Four heavy SDK/module imports are now deferred off the hot startup path.
Net savings on cold module imports:
cli 1200 → 958 ms (-242)
run_agent 1220 → 901 ms (-319)
tools.web_tools 711 → 423 ms (-288)
agent.anthropic_adapter 230 → 15 ms (-215)
agent.auxiliary_client 253 → 68 ms (-185)
Four independent changes in one PR since they all use the same pattern
and share the same risk profile (heavy SDK import → lazy proxy or
function-local import):
1. tools/web_tools.py:
'from firecrawl import Firecrawl' moved into _get_firecrawl_client(),
which is only called when backend='firecrawl'. Users on Exa/Tavily/
Parallel pay zero firecrawl cost.
2. cli.py + gateway/run.py:
'from agent.account_usage import ...' moved into the /limits handlers.
account_usage transitively pulls the OpenAI SDK chain; only needed
when the user runs /limits.
3. agent/anthropic_adapter.py:
'try: import anthropic as _anthropic_sdk' replaced with a cached
'_get_anthropic_sdk()' accessor. The three usage sites
(build_anthropic_client, build_anthropic_bedrock_client,
read_claude_code_credentials_from_keychain) now resolve via the
accessor. All pre-existing test patches of
'agent.anthropic_adapter._anthropic_sdk' keep working because the
accessor respects any value already in module globals.
4. agent/auxiliary_client.py AND run_agent.py:
'from openai import OpenAI' replaced with an '_OpenAIProxy()' module-
level object that looks like the OpenAI class but imports the SDK on
first call/isinstance check. This preserves:
- 15+ in-module OpenAI(...) construction sites in auxiliary_client
and the single site in run_agent's _create_openai_client (Python's
function-scope name lookup finds the proxy, forwards the call);
- 'patch("agent.auxiliary_client.OpenAI", ...)' and
'patch("run_agent.OpenAI", ...)' test patterns used by 28+ test
files (patch replaces the module attribute as usual).
Tried two alternatives first:
- 'from openai._client import OpenAI' — doesn't skip openai/__init__.py
(the audit's hypothesis here was wrong).
- Module-level __getattr__ — works for external access but Python
function-scope name resolution skips __getattr__, so in-module
OpenAI(...) calls NameError.
Note: 'openai' still loads on 'import cli' because
cli.py -> neuter_async_httpx_del() -> openai._base_client, and
run_agent.py -> code_execution_tool.py (module-level
build_execute_code_schema) -> _load_config() -> 'from cli import
CLI_CONFIG'. Deferring those is a separate, larger change — out of scope
for this PR. The savings above all come from avoiding the openai/*,
anthropic/*, and firecrawl/* top-level type-tree imports on paths that
don't need them.
Verified:
- 302/302 tests in tests/agent/{test_anthropic_adapter,
test_bedrock_1m_context, test_minimax_provider, test_anthropic_keychain}
pass. Two pre-existing failures on main unchanged.
- 106/106 tests/agent/test_auxiliary_client.py pass (1 pre-existing fail).
- 97/97 tests/run_agent/test_create_openai_client_kwargs_isolation.py,
test_plugin_context_engine_init.py, test_invalid_context_length_warning.py,
test_api_max_retries_config.py,
tests/hermes_cli/test_gemini_provider.py, test_ollama_cloud_provider.py
pass (1 pre-existing fail).
- Live hermes chat smoke: 2 turns + /model switch + tool calls, zero
errors in the 57-line agent.log window.
- Module-level import of run_agent + auxiliary_client + anthropic_adapter
no longer pulls 'anthropic' or 'firecrawl' at all.
* fix(gateway): restore top-level account_usage import for test-patch surface
CI caught two failures in tests/gateway/test_usage_command.py that I
missed locally:
AttributeError: 'module' object at gateway.run has no attribute 'fetch_account_usage'
The test uses monkeypatch.setattr('gateway.run.fetch_account_usage', ...)
to inject a fake account-fetch call. Moving the import inside the
handler deleted that module-level attribute, breaking the patch surface.
Restoring the top-level import in gateway/run.py gives up the ~230 ms
gateway-boot savings from that one lazy, but:
1. the gateway is a long-running daemon — boot cost is paid once per
install, not per turn;
2. the other four lazy-imports (firecrawl, openai, anthropic, cli's
account_usage) remain in place and still account for the bulk of
the savings reported in the PR body;
3. preserving the patch surface keeps the established
'gateway.run.fetch_account_usage' monkeypatch pattern working
without touching tests.
Verified: tests/gateway/test_usage_command.py — 8 passed, 0 failed.
Full targeted sweep (2336 tests across agent/gateway/hermes_cli/run_agent):
2332 passed, 4 failed — all 4 pre-existing on main.
---------
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Mechanical cleanup across 43 files — removes 46 unused imports
(F401) and 14 unused local variables (F841) detected by
`ruff check --select F401,F841`. Net: -49 lines.
Also fixes a latent NameError in rl_cli.py where `get_hermes_home()`
was called at module line 32 before its import at line 65 — the
module never imported successfully on main. The ruff audit surfaced
this because it correctly saw the symbol as imported-but-unused
(the call happened before the import ran); the fix moves the import
to the top of the file alongside other stdlib imports.
One `# noqa: F401` kept in hermes_cli/status.py for `subprocess`:
tests monkeypatch `hermes_cli.status.subprocess` as a regression
guard that systemctl isn't called on Termux, so the name must
exist at module scope even though the module body doesn't reference
it. Docstring explains the reason.
Also fixes an invalid `# noqa:` directive in
gateway/platforms/discord.py:308 that lacked a rule code.
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
* fix(anthropic): remove Claude Code fingerprinting from OAuth Messages API path
OAuth requests now identify as Hermes on the wire. Removed:
- "You are Claude Code, Anthropic's official CLI for Claude." system
prompt prepend
- Hermes Agent → Claude Code / Nous Research → Anthropic
system-prompt substitutions
- mcp_ tool-name prefix on outgoing tool schemas + message history
- Matching mcp_ strip on inbound tool_use blocks (strip_tool_prefix path
removed from AnthropicTransport.normalize_response, + all 5 call
sites in run_agent.py and auxiliary_client.py)
- user-agent: claude-cli/<v> (external, cli) and x-app: cli headers on
the Messages API client
Added:
- OAuth path strips context-1m-2025-08-07 — Anthropic rejects OAuth
requests carrying it with HTTP 400 'This authentication style is
incompatible with the long context beta header.'
Kept (auth plumbing, not identity spoofing):
- _is_oauth_token classifier and is_oauth flag threading
- Bearer vs x-api-key auth routing
- _OAUTH_ONLY_BETAS (claude-code-20250219, oauth-2025-04-20) — backend
requires these on the OAuth-gated Messages endpoint
- _OAUTH_CLIENT_ID (Claude Code's) — Anthropic doesn't issue OAuth
creds to third parties; this is the only way the login flow works
- claude-cli/<v> User-Agent on the OAuth token exchange + refresh
endpoints at platform.claude.com/v1/oauth/token — bare requests get
Cloudflare 1010 blocked
Verified live against api.anthropic.com with a fresh sk-ant-oat01-*
token:
- claude-haiku-4-5 simple message: HTTP 200, 'OK' response
- claude-haiku-4-5 tool call: HTTP 200, stop_reason=tool_use, tool
named 'terminal' (no mcp_ prefix) round-tripped correctly
- Outgoing wire: no user-agent, no x-app, real Hermes identity in
system prompt, real tool name in schema
Closes/supersedes #16820 (mcp_ PascalCase normalization patch — no longer
needed since the mcp_ round-trip is gone).
* fix(anthropic): resolve_anthropic_token() reads credential pool first
Close the gap where ~/.hermes/auth.json → credential_pool.anthropic
(where hermes login + dashboard PKCE flow write OAuth tokens) was not
in resolve_anthropic_token()'s source list.
Before: users who authed via hermes login got the token written into
the pool, but legacy fallback code paths (auxiliary_client, models
catalog fetch, explicit-runtime path) that call resolve_anthropic_token()
saw None and raised 'No Anthropic credentials found' — even though the
token was sitting in auth.json.
New priority 1: pool.select() with env-sourced entries skipped. Skipping
env:* entries preserves the existing env-var priority logic further
down the chain (static env OAuth → refreshable Claude Code upgrade via
_prefer_refreshable_claude_code_token).
Surfaced while writing the hermes-agent-dev skill playbook for
'finding a live OAuth token for an E2E test'.
---------
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Adds a pre-call sanitizer that detects assistant messages containing only
reasoning (reasoning / reasoning_content, no visible content, no
tool_calls) and drops them from the API copy. Adjacent user messages
left behind are merged so role alternation is preserved for the
provider.
Mirrors Claude Code's approach in src/utils/messages.ts
(filterOrphanedThinkingOnlyMessages + mergeAdjacentUserMessages). We
drop the whole turn rather than fabricate stub text (the '.' /
'(continued)' pattern from contributor PRs #11098, #13010, #16842 that
were rejected because they put words in the model's mouth).
The stored conversation history (self.messages) is never mutated — only
the per-call api_messages copy. Users still see the reasoning block in
the CLI/gateway transcript; only the wire copy is cleaned. Session
persistence keeps the full trace.
Two call sites covered:
- Main agent loop, after _sanitize_api_messages (catches every turn).
- Iteration-limit-summary fallback path.
Tests: tests/run_agent/test_thinking_only_sanitizer.py — 25 cases
covering detection (string/list content, whitespace-only, tool_calls,
reasoning_details list form), drop behavior, adjacent-user merge
(string+string, list+list, mixed), non-mutation of input dicts, and
system-message handling.
E2E live-tested against 5 providers with a poisoned history (empty
assistant message + reasoning_content): OpenRouter→Anthropic/OpenAI/
DeepSeek-R1/Qwen, native Gemini. All 5 accepted the cleaned request.
Happy-path regression (5/5) confirms the sanitizer is a noop when no
thinking-only turn exists.
Related: #16823 (wontfix — stub-text approach rejected).
Co-authored-by: teknium1 <teknium@users.noreply.github.com>
Registers tencent-tokenhub (https://tokenhub.tencentmaas.com/v1) as a
new API-key provider with model tencent/hy3-preview (256K context).
- PROVIDER_REGISTRY entry + TOKENHUB_API_KEY / TOKENHUB_BASE_URL env vars
- Aliases: tencent, tokenhub, tencent-cloud, tencentmaas
- openai_chat transport with is_tokenhub branch for top-level
reasoning_effort (Hy3 is a reasoning model)
- tencent/hy3-preview:free added to OpenRouter curated list
- 60+ tests (provider registry, aliases, runtime resolution,
credentials, model catalog, URL mapping, context length)
- Docs: integrations/providers.md, environment-variables.md,
model-catalog.json
Author: simonweng <simonweng@tencent.com>
Salvaged from PR #16860 onto current main (resolved conflicts with
#16935 Azure Anthropic env-var hint tests and the --provider choices=
list removal in chat_parser).
Follow-up to #15328's vision-unsupported retry branch in run_agent.py.
_strip_images_from_messages() previously deleted any message whose content
was entirely images. That's fine for synthetic user messages injected for
attachment delivery, but it breaks providers for tool-role messages — the
paired tool_call_id on the preceding assistant message ends up unmatched,
which OpenAI-compatible APIs reject with HTTP 400.
Fix: tool-role messages whose content becomes empty are replaced with a
plaintext placeholder that preserves the tool_call_id linkage. Only
non-tool messages are dropped. Added 10 tests covering the role-alternation
invariants + image-type coverage.
Image-rejection detector: expanded phrase list (image content not
supported / multimodal input / vision input / model does not support
image) and gated on 4xx status so transient 5xx errors never get
misinterpreted as 'server said no to images'. Detection is documented as
best-effort English phrase matching.
AUTHOR_MAP: mapped 3820588+ddupont808@users.noreply.github.com to
ddupont808 so release notes attribute the salvage correctly.
Tool handlers (e.g. computer_use capture) return a _multimodal envelope
dict when a screenshot is attached. The tool-message builder was passing
this raw dict as the `content` field of role:tool messages, which is an
illegal format — OpenAI-compatible APIs expect a string or a content-parts
list, not a plain Python dict, and would reject it with a 400/422 error.
Fix: unwrap _multimodal results to their `content` list
([{type:text,...},{type:image_url,...}]) in both the parallel and
sequential tool-call paths. The Anthropic adapter already handles content
lists natively; vision-capable OpenAI-compatible servers (mlx-vlm,
GPT-4o, etc.) accept image_url parts in tool messages directly.
Also add a _vision_supported adaptive fallback: on first image-rejection
error ("Only 'text' content type is supported." etc.) the agent strips all
image parts from the message history and retries with text only, so
text-only endpoints degrade gracefully without crashing the session.
Extends the cua-driver computer-use backend to drive backgrounded macOS
windows without stealing keyboard or mouse focus from the foreground app.
All changes target the cua-driver MCP backend and the shared dispatcher.
## cua_backend.py
**Window-aware capture**: capture() now calls list_windows + get_window_state
instead of the removed capture tool. Prefers structuredContent.windows
(MCP 2024-11-05+ cua-driver) for zero-parse window enumeration; falls back
to regex-parsed text for older builds. Stores the selected (pid, window_id)
as sticky context so subsequent action calls do not need a redundant round-trip.
**Action routing**: click/scroll/type_text/key all carry the sticky pid
(and window_id for element-indexed clicks). type_text routes through
type_text_chars (individual key events) rather than AX attribute write --
WebKit AXTextFields reject attribute writes from backgrounded processes.
**Key parsing**: _parse_key_combo splits cmd+s-style strings into
(key, [modifiers]) and routes to hotkey (modifier present) or
press_key (bare key) -- cua-driver actual tool names.
**set_value method**: new set_value(value, element) calls the cua-driver
set_value MCP tool. For AXPopUpButton / HTML select in a backgrounded Safari,
AXPress opens the native macOS popup which closes immediately when the app is
non-frontmost; set_value AX-presses the matching child option directly
(no menu required, no focus steal).
**focus_app**: reimplemented as a pure window-selector (enumerates
list_windows, sets sticky pid/window_id) without ever raising the window
or stealing focus.
**list_apps**: fixed tool name from listApps to list_apps; handles plain-text
response via regex when structured data is absent.
**Structured-content extraction**: _extract_tool_result now surfaces
structuredContent from MCP results, enabling the list_windows window array
without text parsing.
**Helpers**: _parse_windows_from_text, _parse_elements_from_tree,
_split_tree_text, _parse_key_combo extracted as module-level functions.
## schema.py
Added set_value to the action enum with a description explaining when to
prefer it over click (select/popup elements, sliders, no focus steal).
Added value field for set_value payloads.
## tool.py
Routed set_value action through _dispatch to backend.set_value.
Added set_value to _DESTRUCTIVE_ACTIONS (approval-gated).
Fixed MIME-type detection in _capture_response: cua-driver may return
JPEG; detect from base64 magic bytes (/9j/ -> image/jpeg, else image/png)
rather than hardcoding image/png.
## agent/display.py + run_agent.py
Guard _detect_tool_failure and result-preview logic against non-string
function_result values: multimodal tool results (dicts with _multimodal=True)
are not string-sliceable; treat them as successes and fall back to str()
for length/preview.
Background macOS desktop control via cua-driver MCP — does NOT steal the
user's cursor or keyboard focus, works with any tool-capable model.
Replaces the Anthropic-native `computer_20251124` approach from the
abandoned #4562 with a generic OpenAI function-calling schema plus SOM
(set-of-mark) captures so Claude, GPT, Gemini, and open models can all
drive the desktop via numbered element indices.
- `tools/computer_use/` package — swappable ComputerUseBackend ABC +
CuaDriverBackend (stdio MCP client to trycua/cua's cua-driver binary).
- Universal `computer_use` tool with one schema for all providers.
Actions: capture (som/vision/ax), click, double_click, right_click,
middle_click, drag, scroll, type, key, wait, list_apps, focus_app.
- Multimodal tool-result envelope (`_multimodal=True`, OpenAI-style
`content: [text, image_url]` parts) that flows through
handle_function_call into the tool message. Anthropic adapter converts
into native `tool_result` image blocks; OpenAI-compatible providers
get the parts list directly.
- Image eviction in convert_messages_to_anthropic: only the 3 most
recent screenshots carry real image data; older ones become text
placeholders to cap per-turn token cost.
- Context compressor image pruning: old multimodal tool results have
their image parts stripped instead of being skipped.
- Image-aware token estimation: each image counts as a flat 1500 tokens
instead of its base64 char length (~1MB would have registered as
~250K tokens before).
- COMPUTER_USE_GUIDANCE system-prompt block — injected when the toolset
is active.
- Session DB persistence strips base64 from multimodal tool messages.
- Trajectory saver normalises multimodal messages to text-only.
- `hermes tools` post-setup installs cua-driver via the upstream script
and prints permission-grant instructions.
- CLI approval callback wired so destructive computer_use actions go
through the same prompt_toolkit approval dialog as terminal commands.
- Hard safety guards at the tool level: blocked type patterns
(curl|bash, sudo rm -rf, fork bomb), blocked key combos (empty trash,
force delete, lock screen, log out).
- Skill `apple/macos-computer-use/SKILL.md` — universal (model-agnostic)
workflow guide.
- Docs: `user-guide/features/computer-use.md` plus reference catalog
entries.
44 new tests in tests/tools/test_computer_use.py covering schema
shape (universal, not Anthropic-native), dispatch routing, safety
guards, multimodal envelope, Anthropic adapter conversion, screenshot
eviction, context compressor pruning, image-aware token estimation,
run_agent helpers, and universality guarantees.
469/469 pass across tests/tools/test_computer_use.py + the affected
agent/ test suites.
- `model_tools.py` provider-gating: the tool is available to every
provider. Providers without multi-part tool message support will see
text-only tool results (graceful degradation via `text_summary`).
- Anthropic server-side `clear_tool_uses_20250919` — deferred;
client-side eviction + compressor pruning cover the same cost ceiling
without a beta header.
- macOS only. cua-driver uses private SkyLight SPIs
(SLEventPostToPid, SLPSPostEventRecordTo,
_AXObserverAddNotificationAndCheckRemote) that can break on any macOS
update. Pin with HERMES_CUA_DRIVER_VERSION.
- Requires Accessibility + Screen Recording permissions — the post-setup
prints the Settings path.
Supersedes PR #4562 (pyautogui/Quartz foreground backend, Anthropic-
native schema). Credit @0xbyt4 for the original #3816 groundwork whose
context/eviction/token design is preserved here in generic form.
Streaming-only providers (glm, MiniMax, gpt-5.x via aigw, Anthropic via
openai-compat shims) emit reasoning through delta.reasoning_content
chunks that get accumulated into the local reasoning_text string — but
never land on the assistant message object as a top-level attribute. The
prior guard at _build_assistant_message only wrote reasoning_content
when the SDK exposed hasattr(msg, 'reasoning_content'), so these
providers persisted the chain-of-thought under the internal 'reasoning'
key and omitted the protocol-standard field.
The poison was silent until the user later switched to a DeepSeek-v4 or
Kimi thinking model, at which point replay failed with HTTP 400:
'The reasoning_content in the thinking mode must be passed back to the
API.' One reported session store accumulated 4,031 poisoned messages
across 1,101 files (#16844).
Fix: add an additive fallback that promotes the already-sanitized
reasoning_text to reasoning_content when no earlier branch wrote it AND
reasoning text was actually captured. Layered on top of the existing
SDK-attr branch and DeepSeek ''-pad (#15250) rather than replacing them,
so every existing behavior is preserved:
- SDK-exposed reasoning_content (OpenAI/Moonshot/DeepSeek SDK) still
wins.
- DeepSeek tool-call ''-pad still fires when the SDK exposes the attr
but the value is None.
- Non-thinking turns with no reasoning leave the field absent, so
_copy_reasoning_content_for_api's cross-provider leak guard (#15748),
promote-from-'reasoning' tier, and thinking-pad tier remain live at
replay time.
- No empty '' gets eagerly written on every assistant turn (which would
have bypassed the read-side ladder and triggered empty thinking-block
insertion in the Anthropic adapter).
Tests: three new TestBuildAssistantMessage cases covering the streaming
promotion path, SDK precedence, and field-absent-when-no-reasoning
invariant.
Credit @Sanjays2402 for the original diagnosis and patch in #16884;
this is a scoped rework that preserves the existing read-side
compensation code as defense in depth.
Refs #16844, #16884, #15250, #15353, #15748.
A misconfigured auxiliary.compression.model is a user-fixable problem that silent recovery would hide. The previous retry-on-main logic transparently swallowed aux-model failures whenever the fallback succeeded, leaving the user's broken config in place and racking up future failures.
Track the aux-model failure on the compressor alongside the existing fallback-placeholder fields:
- _last_aux_model_failure_model: str | None
- _last_aux_model_failure_error: str | None
Both are set at the moment the aux model errors (captured before summary_model is cleared for retry), regardless of whether the retry succeeds. Cleared at compress() start and on on_session_reset() so a clean run doesn't leak stale warnings.
Surface at three places:
- gateway hygiene auto-compress: ℹ note to the platform adapter (thread_id preserved)
- gateway /compress command: ℹ line appended to the reply
- CLI via _emit_warning: deduped on (model, error) so repeat compactions don't spam
Distinct from the existing ⚠️ dropped-turns warning — different severity, different emoji, explicit 'context is intact' reassurance.
Same layering concern as the persisted-assistant scrub already removed:
_emit_interim_assistant_message and the final_response return path were
mutating model output broadly. Streaming scrubber covers real leaks
delta-by-delta; these post-stream scrubs were redundant.
Reviewer pushback on the original boundary-hardening commits — three
overreach points pulled plugin-specific policy into shared core paths:
1. gateway/run.py hardcoded a '## Honcho Context' literal split for
vision-LLM output. Plugin-format heading in framework code; could
truncate legitimate output naturally containing that header.
Drop the literal split; keep generic sanitize_context (the wrapper
strip is plugin-agnostic). Plugin-specific cleanup belongs at the
provider boundary, not the shared gateway path.
2. run_agent.run_conversation scrubbed user_message and
persist_user_message before the conversation loop. User text is
sacred — if a user types a literal <memory-context> tag we must
not silently delete it. The producer (build_memory_context_block)
is the only legitimate emitter; user input should never need the
reverse op.
3. _build_assistant_message scrubbed model output before persistence.
Same hazard: would silently mutate legitimate documentation/code
the model emits containing the literal markers. The streaming
scrubber catches real leaks delta-by-delta before content is
concatenated; persist-time scrub was redundant belt-and-suspenders.
4. _fire_stream_delta stripped leading newlines from every delta unless
a paragraph break flag was set. Mid-stream '\n' is legitimate
markdown — lists, code fences, paragraph breaks — and chunk
boundaries are arbitrary. Narrow lstrip to the very first delta
of the stream only (so stale provider preamble still gets cleaned
on turn start, but mid-stream formatting survives).
Plus: build_memory_context_block now logs a warning when its defensive
sanitize_context strips something — surfaces buggy providers returning
pre-wrapped text instead of silently double-fencing.
Net architectural change: scrub surface collapses from 8 sites to 3
(StreamingContextScrubber on output deltas, plugin→backend send,
build_memory_context_block input-validation). Plugin-specific strings
stay out of shared runtime paths. User input and persisted assistant
output are no longer mutated.
Tests: rescoped TestMemoryContextSanitization (helper-correctness only,
no source-inspection of removed call sites), updated vision tests to
drop '## Honcho Context' literal-split assertions, updated
_build_assistant_message persistence test to assert preservation.
Added: cross-turn scrubber reset, build_memory_context_block warn-on-
violation, mid-stream newline preservation (plain + code fence).
sanitize_context() uses a non-greedy block regex that needs both
<memory-context> open and close tags present in a single string. When a
provider streams the fenced memory block across multiple deltas (typical
for recalled-context leaks — the payload often arrives in 10+ 1-80 char
chunks), the per-delta sanitize stripped the lone open/close tags via
_FENCE_TAG_RE but let the payload in between flow straight to the UI.
Adds StreamingContextScrubber: a small stateful scrubber that tracks
open/close tag pairs across deltas, holds back partial-tag tails at
chunk boundaries, and discards span contents wholesale (including the
system-note line that fragments across deltas).
Wired into _fire_stream_delta; reset per user turn; benign trailing
partial-tag tails are flushed at the end of each model call. Mid-span
interruption (provider drops closing tag) drops the orphaned content
rather than leaking it — truncated answer > leaked memory.
Follow-up to #13672 (@dontcallmejames).
The auto-lowered-threshold warning only named the compression model,
making it confusing when the main and aux models are configured with
the same slug but end up with different resolved context lengths (e.g.
OpenRouter's stepfun/step-3.5-flash catalog value vs. a main-model
context_length override). Users couldn't tell whether the warning
reflected two different models or a context-resolution mismatch.
Now includes both 'model (provider)' labels. The aux provider falls
back to the client's base_url hostname when the configured provider
is 'auto', so users see where compression is actually being called.
PR #13734 fixed the concurrent-tool-executor vector (ThreadPoolExecutor
workers didn't inherit the CLI's TLS approval callback). Two vectors
remained that could still land in the deadlocking input() fallback:
1. _spawn_background_review spawns a raw threading.Thread with no
approval callback installed, so any dangerous-command guard the
review agent trips falls back to input() -> deadlock against the
parent's prompt_toolkit TUI (same class as delegate_task subagents,
fixed in 023b1bff1 / #15491). Install a _bg_review_auto_deny
callback at thread start, clear on finally.
2. prompt_dangerous_approval's fallback unconditionally spawned a
daemon thread calling input() when approval_callback was None.
That fallback can never succeed under prompt_toolkit because the
user's Enter goes to pt's raw-mode stdin capture. Detect an active
pt Application via get_app_or_none() and fail closed (deny + log)
instead, so future threads that forget to install a callback
degrade gracefully instead of hanging 60s invisibly.
Regression guards:
- tests/run_agent/test_background_review.py verifies the review
worker thread sees a callable auto-deny callback mid-run and that
the slot is cleared in the finally block.
- tests/tools/test_approval.py TestFailClosedUnderPromptToolkit
verifies prompt_dangerous_approval returns 'deny' fast under a
mocked pt Application, and that a real callback still wins over
the guard.
When tools execute concurrently via ThreadPoolExecutor, worker threads
could not see the thread-local approval/sudo callbacks registered by
the CLI. This caused dangerous-command prompts to fall back to plain
input(), which deadlocks against prompt_toolkit's raw terminal mode.
Capture parent-thread callbacks before launching workers, register
them locally in each _run_tool thread, and clear them on exit.
Mirrors the existing fix pattern from cli.py run_agent() for the
main agent worker thread (GHSA-qg5c-hvr5-hjgr / #13617).
The background skill/memory review agent was created without toolset
restrictions, inheriting the full default tool set. This allowed it to
use terminal, send_message, delegate_task, and other tools outside its
intended scope, potentially performing unrelated side effects after
skill creation.
Restrict the review agent to only memory and skills toolsets by passing
enabled_toolsets=['memory', 'skills'] during AIAgent construction.
Fixes#15204
* feat(image-input): native multimodal routing based on model vision capability
Attach user-sent images as OpenAI-style content parts on the user turn when
the active model supports native vision, so vision-capable models see real
pixels instead of a lossy text description from vision_analyze.
Routing decision (agent/image_routing.py::decide_image_input_mode):
agent.image_input_mode = auto | native | text (default: auto)
In auto mode:
- If auxiliary.vision.provider/model is explicitly configured, keep the
text pipeline (user paid for a dedicated vision backend).
- Else if models.dev reports supports_vision=True for the active
provider/model, attach natively.
- Else fall back to text (current behaviour).
Call sites updated: gateway/run.py (all messaging platforms), tui_gateway
(dashboard/Ink), cli.py (interactive /attach + drag-drop).
run_agent.py changes:
- _prepare_anthropic_messages_for_api now passes image parts through
unchanged when the model supports vision — the Anthropic adapter
translates them to native image blocks. Previous behaviour
(vision_analyze → text) only runs for non-vision Anthropic models.
- New _prepare_messages_for_non_vision_model mirrors the same contract
for chat.completions and codex_responses paths, so non-vision models
on any provider get text-fallback instead of failing at the provider.
- New _model_supports_vision() helper reads models.dev caps.
vision_analyze description rewritten: positions it as a tool for images
NOT already visible in the conversation (URLs, tool output, deeper
inspection). Prevents the model from redundantly calling it on images
already attached natively.
Config default: agent.image_input_mode = auto.
Tests: 35 new (test_image_routing.py + test_vision_aware_preprocessing.py),
all existing tests that reference _prepare_anthropic_messages_for_api
still pass (198 targeted + new tests green).
* feat(image-input): size-cap + resize oversized images, charge image tokens in compressor
Two follow-ups that make the native image routing safer for long / heavy
sessions:
1) Oversize handling in build_native_content_parts:
- 20 MB ceiling per image (matches vision_tools._MAX_BASE64_BYTES,
the most restrictive provider — Gemini inline data).
- Delegates to vision_tools._resize_image_for_vision (Pillow-based,
already battle-tested) to downscale to 5 MB first-try.
- If Pillow is missing or resize still overshoots, the image is
dropped and reported back in skipped[]; caller falls back to text
enrichment for that image.
2) Image-token accounting in context_compressor:
- New _IMAGE_TOKEN_ESTIMATE = 1600 (matches Claude Code's constant;
within the realistic range for Anthropic/GPT-4o/Gemini billing).
- _content_length_for_budget() helper: sums text-part lengths and
charges _IMAGE_CHAR_EQUIVALENT (1600 * 4 chars) per image/image_url/
input_image part. Base64 payload inside image_url is NOT counted
as chars — dimensions don't matter, only image-presence.
- Both tail-cut sites (_prune_old_tool_results L527 and
_find_tail_cut_by_tokens L1126) now call the helper so multi-image
conversations don't slip past compression budget.
Tests: 9 new in test_image_routing.py (oversize triggers resize,
resize-fails-returns-None, oversize-skipped-reported), 11 new in
test_compressor_image_tokens.py (flat charge per image, multiple images,
Responses-API / Anthropic-native / OpenAI-chat shapes, no-inflation on
raw base64, bounds-check on the constant, integration test that an
image-heavy tail actually gets trimmed).
* fix(image-input): replace blanket 20MB ceiling with empirically-verified per-provider limits
The previous commit imposed a hardcoded 20 MB base64 ceiling on all
providers, triggering auto-resize on anything larger. This was wrong in
both directions:
* Too loose for Anthropic — actual limit is 5 MB (returns HTTP 400
'image exceeds 5 MB maximum' above that).
* Too strict for OpenAI / Codex / OpenRouter — accept 49 MB+ without
complaint (empirically verified April 2026 with progressive PNG
sizes).
New behaviour:
* _PROVIDER_BASE64_CEILING table: only anthropic and bedrock have a
ceiling (5 MB, since bedrock-on-Claude shares Anthropic's decoder).
* Providers NOT in the table get no ceiling — images attach at native
size and we trust the provider to return its own error if it
disagrees. A provider-specific 400 message is clearer than us
guessing wrong and silently degrading image quality.
* build_native_content_parts() gains a keyword-only provider arg;
gateway/CLI/TUI pass the active provider so Anthropic users get
auto-resize protection while OpenAI users don't pay it.
* Resize target dropped from 5 MB to 4 MB to slide safely under
Anthropic's boundary with header overhead.
Empirical measurements (direct API, no Hermes in the loop):
image b64 anthropic openrouter/gpt5.5 codex-oauth/gpt5.5
0.19 MB ✓ ✓ ✓
12.37 MB ✗ 400 5MB ✓ ✓
23.85 MB ✗ 400 5MB ✓ ✓
49.46 MB ✗ 413 ✓ ✓
Tests: rewrote TestOversizeHandling (5 tests): no-ceiling pass-through,
Anthropic resize fires, Anthropic skip on resize-fail, build_native_parts
routes ceiling by provider, unknown provider gets no ceiling. All 52
targeted tests pass.
* refactor(image-input): attempt native, shrink-and-retry on provider reject
Replace proactive per-provider size ceilings with a reactive shrink path
on the provider's actual rejection. All providers now attempt native
full-size attachment first; if the provider returns an image-too-large
error, the agent silently shrinks and retries once.
Why the previous design was wrong: hardcoding provider ceilings
(anthropic=5MB, others=unlimited) meant OpenAI users on a 10MB image
paid no tax, but Anthropic users lost quality on anything >5MB even
though the empirical behaviour at provider-reject time is the same
(shrink + retry). Baking the table into the routing layer also
requires updating Hermes every time a provider's limit changes.
Reactive design:
- image_routing.py: _file_to_data_url encodes native size, no ceiling.
build_native_content_parts drops its provider kwarg.
- error_classifier.py: new FailoverReason.image_too_large + pattern
match ("image exceeds", "image too large", etc.) checked BEFORE
context_overflow so Anthropic's 5MB rejection lands in the right
bucket.
- run_agent.py: new _try_shrink_image_parts_in_messages walks api
messages in-place, re-encodes oversized data: URL image parts
through vision_tools._resize_image_for_vision to fit under 4MB,
handles both chat.completions (dict image_url) and Responses
(string image_url) shapes, ignores http URLs (provider-fetched).
New image_shrink_retry_attempted flag in the retry loop fires the
shrink exactly once per turn after credential-pool recovery but
before auth retries.
E2E verified live against Anthropic claude-sonnet-4-6:
- 17.9MB PNG (23.9MB b64) attached at native size
- Anthropic returns 400 "image exceeds 5 MB maximum"
- Agent logs '📐 Image(s) exceeded provider size limit — shrank and
retrying...'
- Retry succeeds, correct response delivered in 6.8s total.
Tests: 12 new (8 shrink-helper shapes + 4 classifier signals),
replaces 5 proactive-ceiling tests with 3 simpler 'native attach works'
tests. 181 targeted tests pass. test_enum_members_exist in
test_error_classifier.py updated for the new enum value.
Adds a short always-on pointer to the system prompt: when the user asks
about configuring, setting up, troubleshooting, or using Hermes Agent
itself, load the hermes-agent skill via skill_view(name='hermes-agent')
and fall back to https://hermes-agent.nousresearch.com/docs via
web_extract. Keeps sessions without skill_view loaded useful too — the
docs URL + web_extract is enough to answer most questions.
The guidance is appended right after DEFAULT_AGENT_IDENTITY (or SOUL.md)
so it ships regardless of which toolset profile is active. Footprint is
~560 chars, behind the existing prompt cache.
On provider switches mid-session (e.g. MiniMax -> DeepSeek), the source
assistant turn carries a 'reasoning' field written by the prior provider
but no 'reasoning_content' key. _copy_reasoning_content_for_api would
promote that foreign 'reasoning' to 'reasoning_content' on the outbound
DeepSeek request, leaking a cross-provider chain of thought and in
practice causing HTTP 400.
DeepSeek's own _build_assistant_message always pins reasoning_content=''
at creation time for tool-call turns, so the shape (reasoning set,
reasoning_content absent, tool_calls present) is unreachable from
same-provider DeepSeek history — it can only come from a prior provider.
Pad with '' in that case instead of promoting.
Healthy same-provider 'reasoning' promotion (no tool_calls, or on
providers that do not require the empty-string pin) is unchanged.
When _compress_context rotates session_id (compression split), fire
on_session_start(new_sid, boundary_reason="compression",
old_session_id=<old>) on the active context engine. Plugin engines
(e.g. hermes-lcm) use this to preserve DAG lineage across the rollover
instead of re-initializing fresh per-session state.
Built-in ContextCompressor.on_session_start accepts **kwargs and ignores
them — no behavior change for default users.
Closes hermes-lcm#68 symptom: after Hermes compressed and minted a new
physical session, LCM was treating the split as a fresh /new and losing
continuity (compression_count: 1, store_messages: 0, dag_nodes: 0).
Credit: @Tosko4 (PR #13370) — minimized scope to the boundary_reason
signal only; the broader session-lifecycle refactor will be taken in
separate PRs if justified by concrete plugin need.
Temporary background review agents can initialize Hindsight-backed memory clients, but close() alone skips provider teardown. Shut the memory provider down before closing so aiohttp sessions do not leak at process exit.
Made-with: Cursor
The background memory/skill review (_spawn_background_review) has always
forked a new AIAgent passing only model and provider, then relied on
AIAgent.__init__ to re-resolve credentials from env vars. This works for
users with keys in ~/.hermes/.env but silently falls back to env-var
auto-resolution in all cases, which fails for OAuth-only providers,
session-scoped creds, and credential-pool setups where auth can't be
reconstructed from env.
This used to be invisible -- failures were swallowed via logger.debug().
PR 8a2506af4 (Apr 24) surfaced auxiliary failures to the user, which
made the stale bug visible as:
"Auxiliary background review failed: No LLM provider configured"
Fix: pass api_key, base_url, api_mode, and credential_pool from the
parent's live runtime into the fork -- matching how every other
auxiliary path (compression, memory flush, vision, session search)
already inherits the parent's credentials via _current_main_runtime().
Follow-up to PR #16053 (/btw as /background alias). Cleans up the
plumbing added exclusively for the old ephemeral /btw handler and
repairs a broken btw bypass that landed between my refactor and this
follow-up.
run_agent.py:
- Remove persist_session kwarg, instance attr, and _persist_session
short-circuit. Only /btw ever passed persist_session=False; with
/btw gone the default (always persist) is the only behavior anyone
ever wanted.
gateway/run.py:
- Remove the unreachable 'if _cmd_def_inner.name == "btw"' block
(PR #16059). Canonical name for a /btw message is 'background' after
alias resolution — the comparison could never be true, and it called
_handle_btw_command which no longer exists. The /background branch
above it already dispatches /btw correctly.
tests/gateway/test_running_agent_session_toggles.py:
- Fix test_btw_dispatches_mid_run to mock _handle_background_command
(the real dispatch target for /btw) instead of the deleted
_handle_btw_command.
The background skill-review prompt (spawned after N user turns) now instructs
the reviewer to SURVEY existing skills first, identify the CLASS of task, and
PREFER updating/generalizing an existing skill over creating a new narrow one.
This reduces near-duplicate skill accumulation at the source. Catches the
common failure mode where repeated tasks of the same class each spawn their
own specific skill ("fix-my-tauri-error", "fix-my-electron-error") instead
of a single class-level skill ("desktop-app-build-troubleshooting").
Applied to both _SKILL_REVIEW_PROMPT and the **Skills** half of
_COMBINED_REVIEW_PROMPT. Memory-only review prompt unchanged.
Groundwork for the Curator feature (issue #7816) — the creation-side fix.
Curator handles the retirement/consolidation side in a follow-up PR.
Tests assert the behavioral instructions are present (survey, class, update-
over-create, overlap-flagging, opt-out clause) rather than snapshotting the
full prompt text.
Nous Portal multiplexes multiple upstream providers (DeepSeek, Kimi,
MiMo, Hermes) behind one endpoint. Before this fix, any 429 on any of
those models recorded a cross-session file breaker that blocked EVERY
model on Nous for the cooldown window -- even though the caller's
own RPM/RPH/TPM/TPH buckets were healthy. Users hit a DeepSeek V4 Pro
capacity error, restarted, switched to Kimi 2.6, and still got
'Nous Portal rate limit active -- resets in 46m 53s'.
Nous already emits the full x-ratelimit-* header suite on every
response (captured by rate_limit_tracker into agent._rate_limit_state).
We now gate the breaker on that data: trip it only when either the
429's own headers or the last-known-good state show a bucket with
remaining == 0 AND a reset window >= 60s. Upstream-capacity 429s
(healthy buckets everywhere, but upstream out of capacity) fall
through to normal retry/fallback and the breaker is never written.
Note: the in-memory 'restart TUI/gateway to clear' workaround
circulated in Discord does NOT work -- the breaker is file-backed at
~/.hermes/rate_limits/nous.json. The workaround for users still
affected by a bad state file is to delete it.
Reported in Discord by CrazyDok1 and KYSIV (Apr 2026).
Azure OpenAI exposes an OpenAI-compatible endpoint at
`{resource}.openai.azure.com/openai/v1` that accepts the standard
`openai` Python client. Two issues prevented gpt-5.x models from working:
1. `_max_tokens_param()` only sent `max_completion_tokens` for
`api.openai.com` URLs. Azure also requires `max_completion_tokens`
for gpt-5.x models.
2. The `codex_responses` upgrade gate unconditionally upgraded gpt-5.x
to Responses API. Azure does NOT support the Responses API — it serves
gpt-5.x on the regular `/chat/completions` path, causing a 404.
Fix: add `_is_azure_openai_url()` that matches `openai.azure.com` URLs.
- `_max_tokens_param()` now returns `max_completion_tokens` for Azure.
- The `codex_responses` upgrade gate skips Azure so gpt-5.x stays on
`chat_completions` where Azure actually serves it.
- The fallback-provider api_mode picker also recognises Azure and stays
on chat_completions.
- Tests cover max_tokens routing, api_mode behaviour, and URL detection.
gpt-4.x models on Azure are unaffected (already used chat_completions +
max_tokens, which Azure accepts for those models).
Salvage of PR #10086 — rewritten against current main where the
codex_responses upgrade gate gained copilot-acp / explicit-api_mode
exclusions.
Azure OpenAI requires an `api-version` query parameter on every request.
When users include it in the base_url (e.g. `?api-version=2025-04-01-preview`),
the OpenAI SDK silently drops it during URL construction, causing 404 errors.
Extract query params from base_url and pass them via `default_query` so the
SDK appends them to every request. This is a generic solution that works for
any custom endpoint requiring query parameters, not just Azure.
No-op for URLs without query params — fully backward compatible.
Fixes#15779. Custom-provider per-model context_length (`custom_providers[].models.<id>.context_length`) is now honored across every resolution path, not just agent startup. Also adds 256K as the top probe tier and default fallback.
## What changed
New helper `hermes_cli.config.get_custom_provider_context_length()` — single source of truth for the per-model override lookup, with trailing-slash-insensitive base-url matching.
`agent.model_metadata.get_model_context_length()` gains an optional `custom_providers=` kwarg (step 0b — runs after explicit `config_context_length` but before every other probe).
Wired through five call sites that previously either duplicated the lookup or ignored it entirely:
- `run_agent.py` startup — refactored to use the new helper (dedups legacy inline loop, keeps invalid-value warning)
- `AIAgent.switch_model()` — re-reads custom_providers from live config on every /model switch
- `hermes_cli.model_switch.resolve_display_context_length()` — new `custom_providers=` kwarg
- `gateway/run.py` /model confirmation (picker callback + text path)
- `gateway/run.py` `_format_session_info` (/info)
## Context probe tiers
`CONTEXT_PROBE_TIERS = [256_000, 128_000, 64_000, 32_000, 16_000, 8_000]` — was `[128_000, ...]`. `DEFAULT_FALLBACK_CONTEXT` follows tier[0], so unknown models now default to 256K. The stale `128000` literal in the OpenRouter metadata-miss path is replaced with `DEFAULT_FALLBACK_CONTEXT` for consistency.
## Repro (from #15779)
```yaml
custom_providers:
- name: my-custom-endpoint
base_url: https://example.invalid/v1
model: gpt-5.5
models:
gpt-5.5:
context_length: 1050000
```
`/model gpt-5.5 --provider custom:my-custom-endpoint` → previously "Context: 128,000", now "Context: 1,050,000".
## Tests
- `tests/hermes_cli/test_custom_provider_context_length.py` — new file, 19 tests covering the helper, step-0b integration, and the 256K tier invariants
- `tests/hermes_cli/test_model_switch_context_display.py` — added regression tests for #15779 through the display resolver
- `tests/gateway/test_session_info.py` — updated default-fallback assertion (128K → 256K)
- `tests/agent/test_model_metadata.py` — updated tier assertions for the new top tier
Previously _copy_reasoning_content_for_api only padded reasoning_content
when the assistant message had tool_calls. DeepSeek V4 thinking mode
requires the field on every assistant turn, including plain text replies
without tool_calls.
- Remove the 'source_msg.get("tool_calls") and' guard
- Update test: plain assistant turns now get padded for DeepSeek/Kimi
Fixes#15213
When a user sends /stop during a streaming API call, the outer poll loop
detects _interrupt_requested and closes the HTTP connection. However, the
inner _call() thread catches the connection error and enters its retry
loop — opening a FRESH connection without checking the interrupt flag.
On slow providers like ollama-cloud, each retry attempt blocks for the
full stream-read timeout (120s+). With 3 retry attempts this caused
510+ second delays between /stop and actual response — the agent appeared
completely unresponsive despite the stop being acknowledged.
Fix: add an _interrupt_requested check at the top of the streaming retry
loop so the agent exits immediately instead of retrying.
Also fix log truncation: all session key logging in gateway/run.py used
[:20] or [:30] slices, which truncated 'agent:main:telegram:dm:5690190437'
(33 chars) to 'agent:main:telegram:' — losing the identifying chat type
and user ID. Replace with full keys to make logs debuggable.
Reported by user Sidharth Pulipaka via Telegram on ollama-cloud provider.
The AIAgent.flush_memories pre-compression save, the gateway
_flush_memories_for_session, and everything feeding them are
obsolete now that the background memory/skill review handles
persistent memory extraction.
Problems with flush_memories:
- Pre-dates the background review loop. It was the only memory-save
path when introduced; the background review now fires every 10 user
turns on CLI and gateway alike, which is far more frequent than
compression or session reset ever triggered flush.
- Blocking and synchronous. Pre-compression flush ran on the live agent
before compression, blocking the user-visible response.
- Cache-breaking. Flush built a temporary conversation prefix
(system prompt + memory-only tool list) that diverged from the live
conversation's cached prefix, invalidating prompt caching. The
gateway variant spawned a fresh AIAgent with its own clean prompt
for each finalized session — still cache-breaking, just in a
different process.
- Redundant. Background review runs in the live conversation's
session context, gets the same content, writes to the same memory
store, and doesn't break the cache. Everything flush_memories
claimed to preserve is already covered.
What this removes:
- AIAgent.flush_memories() method (~248 LOC in run_agent.py)
- Pre-compression flush call in _compress_context
- flush_memories call sites in cli.py (/new + exit)
- GatewayRunner._flush_memories_for_session + _async_flush_memories
(and the 3 call sites: session expiry watcher, /new, /resume)
- 'flush_memories' entry from DEFAULT_CONFIG auxiliary tasks,
hermes tools UI task list, auxiliary_client docstrings
- _memory_flush_min_turns config + init
- #15631's headroom-deduction math in
_check_compression_model_feasibility (headroom was only needed
because flush dragged the full main-agent system prompt along;
the compression summariser sends a single user-role prompt so
new_threshold = aux_context is safe again)
- The dedicated test files and assertions that exercised
flush-specific paths
What this renames (with read-time backcompat on sessions.json):
- SessionEntry.memory_flushed -> SessionEntry.expiry_finalized.
The session-expiry watcher still uses the flag to avoid re-running
finalize/eviction on the same expired session; the new name
reflects what it now actually gates. from_dict() reads
'expiry_finalized' first, falls back to the legacy 'memory_flushed'
key so existing sessions.json files upgrade seamlessly.
Supersedes #15631 and #15638.
Tested: 383 targeted tests pass across run_agent/, agent/, cli/,
and gateway/ session-boundary suites. No behavior regressions —
background memory review continues to handle persistent memory
extraction on both CLI and gateway.
_check_compression_model_feasibility calls get_model_context_length
without provider=, so Codex OAuth users get 1,050,000 (from models.dev
for 'openai') instead of the actual 272,000 limit. This happens because
_infer_provider_from_url maps chatgpt.com → 'openai' (not 'openai-codex'),
skipping the Codex-specific resolution branch entirely.
Result: compression threshold set at 85% of 1.05M = 892K — conversations
never trigger compression, the context grows unbounded, and when gateway
hygiene eventually forces compression, the Codex endpoint drops the
oversized streaming request ('peer closed connection without sending
complete message body').
Fix: forward self.provider to get_model_context_length so provider-
specific resolution branches (Codex OAuth 272K, Copilot live /models,
Nous suffix-match) fire correctly.
Reported by user on GPT 5.5 via Codex OAuth Pro (paste.rs/vsra3).
When the auxiliary compression model's context is smaller than the main
model's compression threshold, _check_compression_model_feasibility
auto-lowers the session threshold. Previously it set:
new_threshold = aux_context
This let the raw message list grow to exactly aux_context tokens. But
compression and flush_memories actually send system_prompt + tool_schemas
+ messages to the aux model. With 50+ tools that overhead is 25-30K
tokens, so the full request overflowed aux with HTTP 400.
Subtract a headroom estimate from aux_context before setting the new
threshold: the actual tool-schema token count (from
estimate_request_tokens_rough) plus a 12K allowance for the system
prompt (not yet built at __init__ time) and flush-instruction overhead.
Clamp to MINIMUM_CONTEXT_LENGTH so the session still starts even with
an unusually heavy tool schema.
This fixes the 'flush_memories overflow on busy toolsets' path that
Teknium flagged — where main and aux can be nominally the same model
but still 400 because the threshold left no room for the request
overhead. Same fix also protects the normal compression summarisation
request on the same binding aux.
Tests: two new regression tests cover the headroom reservation and the
MINIMUM_CONTEXT_LENGTH floor. Two existing tests updated for the new
(lower) threshold values now that empty-tools still produces a 12K
static headroom deduction.
The memory-flush fallback for api_mode='codex_responses' was unconditionally
adding `temperature` to codex_kwargs before calling _run_codex_stream. The
Responses API does not accept temperature on any supported backend:
- chatgpt.com/backend-api/codex rejects it outright
- api.openai.com + gpt-5/o-series reasoning models reject it
- Copilot Responses rejects it on reasoning models
The CodexAuxiliaryClient adapter and the codex_responses transport both
correctly omit temperature — the flush fallback was the only path putting
it back. On errors from the primary aux path (e.g. expired OAuth token),
users saw `⚠ Auxiliary memory flush failed: HTTP 400: Unsupported parameter:
temperature`.
Reported by Garik [NOUS] on GPT-5.5 via Codex OAuth Pro.
Extracts _needs_kimi_tool_reasoning() for symmetry with the existing
_needs_deepseek_tool_reasoning() helper, so _copy_reasoning_content_for_api
uses the same detection logic as _build_assistant_message. Future changes
to either provider's signals now only touch one function.
Adds tests/run_agent/test_deepseek_reasoning_content_echo.py covering:
- All 3 DeepSeek detection signals (provider, model, host)
- Poisoned history replay (empty string fallback)
- Plain assistant turns NOT padded
- Explicit reasoning_content preserved
- Reasoning field promoted to reasoning_content
- Existing Kimi/Moonshot detection intact
- Non-thinking providers left alone
21 tests, all pass.
DeepSeek V4 thinking mode requires reasoning_content on every
assistant message that includes tool_calls. When this field is
missing from persisted history, replaying the session causes
HTTP 400: 'The reasoning_content in the thinking mode must be
passed back to the API.'
Two-part fix (refs #15250):
1. _copy_reasoning_content_for_api: Merge the Kimi-only and
DeepSeek detection into a single needs_tool_reasoning_echo
check. This handles already-poisoned persisted sessions by
injecting an empty reasoning_content on replay.
2. _build_assistant_message: Store reasoning_content='' on new
DeepSeek tool-call messages at creation time, preventing
future session poisoning at the source.
Additional fix:
3. _handle_max_iterations: Add missing call to
_copy_reasoning_content_for_api in the max-iterations flush
path (previously only main loop and flush_memories had it).
Detection covers:
- provider == 'deepseek'
- model name containing 'deepseek' (case-insensitive)
- base URL matching api.deepseek.com (for custom provider)
``run_conversation`` was calling ``memory_manager.sync_all(
original_user_message, final_response)`` at the end of every turn
where both args were present. That gate didn't consider the
``interrupted`` local flag, so an external memory backend received
partial assistant output, aborted tool chains, or mid-stream resets as
durable conversational truth. Downstream recall then treated the
not-yet-real state as if the user had seen it complete, poisoning the
trust boundary between "what the user took away from the turn" and
"what Hermes was in the middle of producing when the interrupt hit".
Extracted the inline sync block into a new private method
``AIAgent._sync_external_memory_for_turn(original_user_message,
final_response, interrupted)`` so the interrupt guard is a single
visible check at the top of the method instead of hidden in a
boolean-and at the call site. That also gives tests a clean seam to
assert on — the pre-fix layout buried the logic inside the 3,000-line
``run_conversation`` function where no focused test could reach it.
The new method encodes three independent skip conditions:
1. ``interrupted`` → skip entirely (the #15218 fix). Applies even
when ``final_response`` and ``original_user_message`` happen to
be populated — an interrupt may have landed between a streamed
reply and the next tool call, so the strings on disk are not
actually the turn the user took away.
2. No memory manager / no final_response / no user message →
preserve existing skip behaviour (nothing new for providerless
sessions, system-initiated refreshes, tool-only turns that never
resolved, etc.).
3. Sync_all / queue_prefetch_all exceptions → swallow. External
memory providers are strictly best-effort; a misconfigured or
offline backend must never block the user from seeing their
response.
The prefetch side-effect is gated on the same interrupt flag: the
user's next message is almost certainly a retry of the same intent,
and a prefetch keyed on the interrupted turn would fire against stale
context.
### Tests (16 new, all passing on py3.11 venv)
``tests/run_agent/test_memory_sync_interrupted.py`` exercises the
helper directly on a bare ``AIAgent`` (``__new__`` pattern that the
interrupt-propagation tests already use). Coverage:
- Interrupted turn with full-looking response → no sync (the fix)
- Interrupted turn with long assistant output → no sync (the interrupt
could have landed mid-stream; strings-on-disk lie)
- Normal completed turn → sync_all + queue_prefetch_all both called
with the right args (regression guard for the positive path)
- No final_response / no user_message / no memory manager → existing
pre-fix skip paths still apply
- sync_all raises → exception swallowed, prefetch still attempted
- queue_prefetch_all raises → exception swallowed after sync succeeded
- 8-case parametrised matrix across (interrupted × final_response ×
original_user_message) asserts sync fires iff interrupted=False AND
both strings are non-empty
Closes#15218
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extends _repair_tool_call_arguments() to cover the most common local-model
JSON corruption pattern: llama.cpp/Ollama backends emit literal tabs and
newlines inside JSON string values (memory save summaries, file contents,
etc.). Previously fell through to '{}' replacement, losing the call.
Adds two repair passes:
- Pass 0: json.loads(strict=False) + re-serialise to canonical wire form
- Pass 4: escape 0x00-0x1F control chars inside string values, then retry
Ports the core utility from #12068 / PR #12093 without the larger plumbing
change (that PR also replaced json.loads at 8 call sites; current main's
_repair_tool_call_arguments is already the single chokepoint, so the
upgrade happens transparently for every existing caller).
Credit: @truenorth-lj for the original utility design.
4 new regression tests covering literal newlines, tabs, re-serialisation
to strict=True-valid output, and the trailing-comma + control-char
combination case.
When the streaming path (chat completions) assembled tool call deltas and
detected malformed JSON arguments, it set has_truncated_tool_args=True but
passed the broken args through unchanged. This triggered the truncation
handler which returned a partial result and killed the session (/new required).
_many_ malformations are repairable: trailing commas, unclosed brackets,
Python None, empty strings. _repair_tool_call_arguments() already existed
for the pre-API-request path but wasn't called during streaming assembly.
Now when JSON parsing fails during streaming assembly, we attempt repair
via _repair_tool_call_arguments() before flagging as truncated. If repair
succeeds (returns valid JSON), the tool call proceeds normally. Only truly
unrepairable args fall through to the truncation handler.
This prevents the most common session-killing failure mode for models like
GLM-5.1 that produce trailing commas or unclosed brackets.
Tests: 12 new streaming assembly repair tests, all 29 existing repair
tests still passing.
When a session is split by context compression mid-tool-call, an assistant
message may end up with truncated/invalid JSON in tool_calls[*].function.arguments.
On the next turn this is replayed verbatim and providers reject the entire request
with HTTP 400 invalid_tool_call_format, bricking the conversation in a loop that
cannot recover without manual session quarantine.
This patch adds a defensive sanitizer that runs immediately before
client.chat.completions.create() in AIAgent.run_conversation():
- Validates each assistant tool_calls[*].function.arguments via json.loads
- Replaces invalid/empty arguments with '{}'
- Injects a synthetic tool response (or prepends a marker to the existing one)
so downstream messages keep valid tool_call_id pairing
- Logs each repair with session_id / message_index / preview for observability
Defense in depth: corruption can originate from compression splits, manual edits,
or plugin bugs. Sanitizing at the send chokepoint catches all sources.
Adds 7 unit tests covering: truncated JSON, empty string, None, non-string args,
existing matching tool response (no duplicate injection), non-assistant messages
ignored, multiple repairs.
Fixes#15236
Three interrupt-recovery sites in run_agent.py rebuilt self._anthropic_client
with build_anthropic_client(self._anthropic_api_key, ...) unconditionally.
When provider=bedrock + api_mode=anthropic_messages (AnthropicBedrock SDK
path), self._anthropic_api_key is the sentinel 'aws-sdk' — build_anthropic_client
doesn't accept that and the rebuild either crashed or produced a non-functional
client.
Extract a _rebuild_anthropic_client() helper that dispatches to
build_anthropic_bedrock_client(region) when provider='bedrock', falling back
to build_anthropic_client() for native Anthropic and other anthropic_messages
providers (MiniMax, Kimi, Alibaba, etc.). Three inline rebuild sites now call
the helper.
Partial salvage of #14680 by @bsgdigital — only the _rebuild_anthropic_client
helper. The normalize_model_name Bedrock-prefix piece was subsumed by #14664,
and the aux client aws_sdk branch was subsumed by #14770 (both in the same
salvage PR as this commit).
## Problem
When a pooled HTTPS connection to the Bedrock runtime goes stale (NAT
timeout, VPN flap, server-side TCP RST, proxy idle cull), the next
Converse call surfaces as one of:
* botocore.exceptions.ConnectionClosedError / ReadTimeoutError /
EndpointConnectionError / ConnectTimeoutError
* urllib3.exceptions.ProtocolError
* A bare AssertionError raised from inside urllib3 or botocore
(internal connection-pool invariant check)
The agent loop retries the request 3x, but the cached boto3 client in
_bedrock_runtime_client_cache is reused across retries — so every
attempt hits the same dead connection pool and fails identically.
Only a process restart clears the cache and lets the user keep working.
The bare-AssertionError variant is particularly user-hostile because
str(AssertionError()) is an empty string, so the retry banner shows:
⚠️ API call failed: AssertionError
📝 Error:
with no hint of what went wrong.
## Fix
Add two helpers to agent/bedrock_adapter.py:
* is_stale_connection_error(exc) — classifies exceptions that
indicate dead-client/dead-socket state. Matches botocore
ConnectionError + HTTPClientError subtrees, urllib3
ProtocolError / NewConnectionError, and AssertionError
raised from a frame whose module name starts with urllib3.,
botocore., or boto3.. Application-level AssertionErrors are
intentionally excluded.
* invalidate_runtime_client(region) — per-region counterpart to
the existing reset_client_cache(). Evicts a single cached
client so the next call rebuilds it (and its connection pool).
Wire both into the Converse call sites:
* call_converse() / call_converse_stream() in
bedrock_adapter.py (defense-in-depth for any future caller)
* The two direct client.converse(**kwargs) /
client.converse_stream(**kwargs) call sites in run_agent.py
(the paths the agent loop actually uses)
On a stale-connection exception, the client is evicted and the
exception re-raised unchanged. The agent's existing retry loop then
builds a fresh client on the next attempt and recovers without
requiring a process restart.
## Tests
tests/agent/test_bedrock_adapter.py gets three new classes (14 tests):
* TestInvalidateRuntimeClient — per-region eviction correctness;
non-cached region returns False.
* TestIsStaleConnectionError — classifies botocore
ConnectionClosedError / EndpointConnectionError /
ReadTimeoutError, urllib3 ProtocolError, library-internal
AssertionError (both urllib3.* and botocore.* frames), and
correctly ignores application-level AssertionError and
unrelated exceptions (ValueError, KeyError).
* TestCallConverseInvalidatesOnStaleError — end-to-end: stale
error evicts the cached client, non-stale error (validation)
leaves it alone, successful call leaves it cached.
All 116 tests in test_bedrock_adapter.py pass.
Signed-off-by: Andre Kurait <andrekurait@gmail.com>
Try to activate fallback model after errors was calling get_model_context_length()
without the config_context_length parameter, causing it to fall through to
DEFAULT_FALLBACK_CONTEXT (128K) even when config.yaml has an explicit
model.context_length value (e.g. 204800 for MiniMax-M2.7).
This mirrors the fix already present in switch_model() at line 1988, which
correctly passes config_context_length. The fallback path was missed.
Fixes: context_length forced to 128K on fallback activation
ssl.SSLError (and its subclass ssl.SSLCertVerificationError) inherits from
OSError *and* ValueError via Python's MRO. The is_local_validation_error
check used isinstance(api_error, (ValueError, TypeError)) to detect
programming bugs that should abort immediately — but this inadvertently
caught ssl.SSLError, treating a TLS transport failure as a non-retryable
client error.
The error classifier already maps SSLCertVerificationError to
FailoverReason.timeout with retryable=True (its type name is in
_TRANSPORT_ERROR_TYPES), but the inline isinstance guard was overriding
that classification and triggering an unnecessary abort.
Fix: add ssl.SSLError to the exclusion list alongside the existing
UnicodeEncodeError carve-out so TLS errors fall through to the
classifier's retryable path.
Closes#14367
Claude-style and some Anthropic-tuned models occasionally emit tool
names as class-like identifiers: TodoTool_tool, Patch_tool,
BrowserClick_tool, PatchTool. These failed strict-dict lookup in
valid_tool_names and triggered the 'Unknown tool' self-correction
loop, wasting a full turn of iteration and tokens.
_repair_tool_call already handled lowercase / separator / fuzzy
matches but couldn't bridge the CamelCase-to-snake_case gap or the
trailing '_tool' suffix that Claude sometimes tacks on. Extend it
with two bounded normalization passes:
1. CamelCase -> snake_case (via regex lookbehind).
2. Strip trailing _tool / -tool / tool suffix (case-insensitive,
applied twice so TodoTool_tool reduces all the way: strip
_tool -> TodoTool, snake -> todo_tool, strip 'tool' -> todo).
Cheap fast-paths (lowercase / separator-normalized) still run first
so the common case stays zero-cost. Fuzzy match remains the last
resort unchanged.
Tests: tests/run_agent/test_repair_tool_call_name.py covers the
three original reports (TodoTool_tool, Patch_tool, BrowserClick_tool),
plus PatchTool, WriteFileTool, ReadFile_tool, write-file_Tool,
patch-tool, and edge cases (empty, None, '_tool' alone, genuinely
unknown names).
18 new tests + 17 existing arg-repair tests = 35/35 pass.
Closes#14784
Extracts pool-rotation-room logic into `_pool_may_recover_from_rate_limit`
so single-credential pools no longer block the eager-fallback path on 429.
The existing check `pool is not None and pool.has_available()` lets
fallback fire only after the pool marks every entry as exhausted. With
exactly one credential in the pool (the common shape for Gemini OAuth,
Vertex service accounts, and any personal-key setup), `has_available()`
flips back to True as soon as the cooldown expires — Hermes retries
against the same entry, hits the same daily-quota 429, and burns the
retry budget in a tight loop before ever reaching the configured
`fallback_model`. Observed in the wild as 4+ hours of 429 noise on a
single Gemini key instead of falling through to Vertex as configured.
Rotation is only meaningful with more than one credential — gate on
`len(pool.entries()) > 1`. Multi-credential pools keep the current
wait-for-rotation behaviour unchanged.
Fixes#11314. Related to #8947, #10210, #7230. Narrower scope than
open PRs #8023 (classifier change) and #11492 (503/529 credential-pool
bypass) — this addresses the single-credential 429 case specifically
and does not conflict with either.
Tests: 6 new unit tests in tests/run_agent/test_provider_fallback.py
covering (a) None pool, (b) single-cred available, (c) single-cred in
cooldown, (d) 2-cred available rotates, (e) multi-cred all cooling-down
falls back, (f) many-cred available rotates. All 18 tests in the file
pass.
Previously _handle_credential_pool_error handled 401, 402, and 429
but silently ignored 403. When a provider returns 403 for a revoked or
unauthorised credential (e.g. Nous agent_key invalidated by a newer
login), the pool was never rotated and every subsequent request
continued to use the same failing credential.
Treat 403 the same as 402: immediately mark the current credential
exhausted and rotate to the next pool entry, since a Forbidden response
will not resolve itself with a retry.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When using GitHub Copilot as provider, HTTP 401 errors could cause
Hermes to silently fall back to the next model in the chain instead
of recovering. This adds a one-shot retry mechanism that:
1. Re-resolves the Copilot token via the standard priority chain
(COPILOT_GITHUB_TOKEN -> GH_TOKEN -> GITHUB_TOKEN -> gh auth token)
2. Rebuilds the OpenAI client with fresh credentials and Copilot headers
3. Retries the failed request before falling back
The fix handles the common case where the gho_* OAuth token remains
valid but the httpx client state becomes stale (e.g. after startup
race conditions or long-lived sessions).
Key design decisions:
- Always rebuild client even if token string unchanged (recovers stale state)
- Uses _apply_client_headers_for_base_url() for canonical header management
- One-shot flag guard prevents infinite 401 loops (matches existing pattern
used by Codex/Nous/Anthropic providers)
- No token exchange via /copilot_internal/v2/token (returns 404 for some
account types; direct gho_* auth works reliably)
Tests: 3 new test cases covering end-to-end 401->refresh->retry,
client rebuild verification, and same-token rebuild scenarios.
Docs: Updated providers.md with Copilot auth behavior section.
json.JSONDecodeError inherits from ValueError. The agent loop's
non-retryable classifier at run_agent.py ~L10782 treated any
ValueError/TypeError as a local programming bug and short-circuited
retry. Without a carve-out, a transient JSONDecodeError from a
provider that returned a malformed response body, a truncated stream,
or a router-layer corruption would fail the turn immediately.
Add JSONDecodeError to the existing UnicodeEncodeError exclusion
tuple so the classified-retry logic (which already handles 429/529/
context-overflow/etc.) gets to run on bad-JSON errors.
Tests (tests/run_agent/test_jsondecodeerror_retryable.py):
- JSONDecodeError: NOT local validation
- UnicodeEncodeError: NOT local validation (existing carve-out)
- bare ValueError: IS local validation (programming bug)
- bare TypeError: IS local validation (programming bug)
- source-level assertion that run_agent.py still carries the carve-out
(guards against accidental revert)
Closes#14782
Two related paths where Codex auth failures silently swallowed the
fallback chain instead of switching to the next provider:
1. cli.py — _ensure_runtime_credentials() calls resolve_runtime_provider()
before each turn. When provider is explicitly configured (not "auto"),
an AuthError from token refresh is re-raised and printed as a bold-red
error, returning False before the agent ever starts. The fallback chain
was never tried. Fix: on AuthError, iterate fallback_providers and
switch to the first one that resolves successfully.
2. run_agent.py — inside the codex_responses validity gate (inner retry
loop), response.status in {"failed","cancelled"} with non-empty output
items was treated as a valid response and broke out of the retry loop,
reaching _normalize_codex_response() outside the fallback machinery.
That function raises RuntimeError on status="failed", which propagates
to the outer except with no fallback logic. Fix: detect terminal status
codes before the output_items check and set response_invalid=True so
the existing fallback chain fires normally.
- Load prompt_caching.cache_ttl in AIAgent (5m default, 1h opt-in)
- Document DEFAULT_CONFIG and developer guide example
- Add unit tests for default, 1h, and invalid TTL fallback
Made-with: Cursor
Manual /compress crashed with 'LCMEngine' object has no attribute
'_align_boundary_forward' when any context-engine plugin was active.
The gateway handler reached into _align_boundary_forward and
_find_tail_cut_by_tokens on tmp_agent.context_compressor, but those
are ContextCompressor-specific — not part of the generic ContextEngine
ABC — so every plugin engine (LCM, etc.) raised AttributeError.
- Add optional has_content_to_compress(messages) to ContextEngine ABC
with a safe default of True (always attempt).
- Override it in the built-in ContextCompressor using the existing
private helpers — preserves exact prior behavior for 'compressor'.
- Rewrite gateway /compress preflight to call the ABC method, deleting
the private-helper reach-in.
- Add focus_topic to the ABC compress() signature. Make _compress_context
retry without focus_topic on TypeError so older strict-sig plugins
don't crash on manual /compress <focus>.
- Regression test with a fake ContextEngine subclass that only
implements the ABC (mirrors LCM's surface).
Reported by @selfhostedsoul (Discord, Apr 22).
Closes#11616.
The agent's API retry loop hardcoded max_retries = 3, so users with
fallback providers on flaky primaries burned through ~3 × provider
timeout (e.g. 3 × 180s = 9 minutes) before their fallback chain got a
chance to kick in.
Expose a new config key:
agent:
api_max_retries: 3 # default unchanged
Set it to 1 for fast failover when you have fallback providers, or
raise it if you prefer longer tolerance on a single provider. Values
< 1 are clamped to 1 (single attempt, no retry); non-integer values
fall back to the default.
This wraps the Hermes-level retry loop only — the OpenAI SDK's own
low-level retries (max_retries=2 default) still run beneath this for
transient network errors.
Changes:
- hermes_cli/config.py: add agent.api_max_retries default 3 with comment.
- run_agent.py: read self._api_max_retries in AIAgent.__init__; replace
hardcoded max_retries = 3 in the retry loop with self._api_max_retries.
- cli-config.yaml.example: documented example entry.
- hermes_cli/tips.py: discoverable tip line.
- tests/run_agent/test_api_max_retries_config.py: 4 tests covering
default, override, clamp-to-one, and invalid-value fallback.
NormalizedResponse and ToolCall now have backward-compat properties
so the agent loop can read them directly without the shim:
ToolCall: .type, .function (returns self), .call_id, .response_item_id
NormalizedResponse: .reasoning_content, .reasoning_details,
.codex_reasoning_items
This eliminates the 35-line shim and its 4 call sites in run_agent.py.
Also changes flush_memories guard from hasattr(response, 'choices')
to self.api_mode in ('chat_completions', 'bedrock_converse') so it
works with raw boto3 dicts too.
WS1 items 3+4 of Cycle 2 (#14418).
Follow-up for #13862 — the post-init api_mode upgrade at __init__ (direct OpenAI /
gpt-5-requires-responses path) runs AFTER the eager transport warm. Clear the cache
so the stale chat_completions entry is evicted.
Cosmetic: correctness was already fine since _get_transport() keys by current
api_mode, but this avoids leaving unused cache state behind.
Consolidate 4 per-transport lazy singleton helpers (_get_anthropic_transport,
_get_codex_transport, _get_chat_completions_transport, _get_bedrock_transport)
into one generic _get_transport(api_mode) with a shared dict cache.
Collapse the 65-line main normalize block (3 api_mode branches, each with
its own SimpleNamespace shim) into 7 lines: one _get_transport() call +
one _nr_to_assistant_message() shared shim. The shim extracts provider_data
fields (codex_reasoning_items, reasoning_details, call_id, response_item_id)
into the SimpleNamespace shape downstream code expects.
Wire chat_completions and bedrock_converse normalize through their transports
for the first time — these were previously falling into the raw
response.choices[0].message else branch.
Remove 8 dead codex adapter imports that have zero callers after PRs 1-6.
Transport lifecycle improvements:
- Eagerly warm transport cache at __init__ (surfaces import errors early)
- Invalidate transport cache on api_mode change (switch_model, fallback
activation, fallback restore, transport recovery) — prevents stale
transport after mid-session provider switch
run_agent.py: -32 net lines (11,988 -> 11,956).
PR 7 of the provider transport refactor.
Port from openclaw/openclaw#67318. Some open models (notably Gemma
variants served via OpenRouter) emit tool calls as XML blocks inside
assistant content instead of via the structured tool_calls field:
<function name="read_file"><parameter name="path">/tmp/x</parameter></function>
<tool_call>{"name":"x"}</tool_call>
<function_calls>[{...}]</function_calls>
Left unstripped, this raw XML leaked to gateway users (Discord, Telegram,
Matrix, Feishu, Signal, WhatsApp, etc.) and the CLI, since hermes-agent's
existing reasoning-tag stripper handled only <think>/<thinking>/<thought>
variants.
Extend _strip_think_blocks (run_agent.py) and _strip_reasoning_tags
(cli.py) to cover:
* <tool_call>, <tool_calls>, <tool_result>
* <function_call>, <function_calls>
* <function name="..."> ... </function> (Gemma-style)
The <function> variant is boundary-gated (only strips when the tag sits
at start-of-line or after sentence punctuation AND carries a name="..."
attribute) so prose mentions like 'Use <function> declarations in JS'
are preserved. Dangling <function name="..."> with no close is
intentionally left visible — matches OpenClaw's asymmetry so a truncated
streaming tail still reaches the user.
Tests: 9 new cases in TestStripThinkBlocks (run_agent) + 9 in new file
tests/run_agent/test_strip_reasoning_tags_cli.py. Covers Qwen-style
<tool_call>, Gemma-style <function name="...">, multi-line payloads,
prose preservation, stray close tags, dangling open tags, and mixed
reasoning+tool_call content.
Note: this port covers the post-streaming final-text path, which is what
gateway adapters and CLI display consume. Extending the per-delta stream
filter in gateway/stream_consumer.py to hide these tags live as they
stream is a separate follow-up; for now users may see raw XML briefly
during a stream before the final cleaned text replaces it.
Refs: openclaw/openclaw#67318
When the streaming connection dropped AFTER user-visible text was
delivered but a tool call was in flight, we stubbed the turn with a
'⚠ Stream stalled mid tool-call; Ask me to retry' warning — costing
an iteration and breaking the flow. Users report this happening
increasingly often on long SSE streams through flaky provider routes.
Fix: in the existing inner stream-retry loop, relax the
deltas_were_sent short-circuit. If a tool call was in flight
(partial_tool_names populated) AND the error is a transient connection
error (timeout, RemoteProtocolError, SSE 'connection lost', etc.),
silently retry instead of bailing out. Fire a brief 'Connection
dropped mid tool-call; reconnecting…' marker so the user understands
the preamble is about to be re-streamed.
Researched how Claude Code (tombstone + non-streaming fallback),
OpenCode (blind Effect.retry wrapping whole stream), and Clawdbot
(4-way gate: stopReason==error + output==0 + !hadPotentialSideEffects)
handle this. Chose the narrow Clawdbot-style gate: retry only when
(a) a tool call was actually in flight (otherwise the existing
stub-with-recovered-text is correct for pure-text stalls) and
(b) the error is transient. Side-effect safety is automatic — no
tool has been dispatched within this single API call yet.
UX trade-off: user sees preamble text twice on retry (OpenCode-style).
Strictly better than a lost action with a 'retry manually' message.
If retries exhaust, falls through to the existing stub-with-warning
path so the user isn't left with zero signal.
Tests: 3 new tests in TestSilentRetryMidToolCall covering
(1) silent retry recovers tool call; (2) exhausted retries fall back
to stub; (3) text-only stalls don't trigger retry. 30/30 pass.
* fix(plugins): auto-coerce user-installed memory plugins to kind=exclusive
User-installed memory provider plugins at $HERMES_HOME/plugins/<name>/
were being dispatched to the general PluginManager, which has no
register_memory_provider method on PluginContext. Every startup logged:
Failed to load plugin 'mempalace': 'PluginContext' object has no
attribute 'register_memory_provider'
Bundled memory providers were already skipped via skip_names={memory,
context_engine} in discover_and_load, but user-installed ones weren't.
Fix: _parse_manifest now scans the plugin's __init__.py source for
'register_memory_provider' or 'MemoryProvider' (same heuristic as
plugins/memory/__init__.py:_is_memory_provider_dir) and auto-coerces
kind to 'exclusive' when the manifest didn't declare one explicitly.
This routes the plugin to plugins/memory discovery instead of the
general loader.
The escape hatch: if a manifest explicitly declares kind: standalone,
the heuristic doesn't override it.
Reported by Uncle HODL on Discord.
* fix(nous): actionable CLI message when Nous 401 refresh fails
Mirrors the Anthropic 401 diagnostic pattern. When Nous returns 401
and the credential refresh (_try_refresh_nous_client_credentials)
also fails, the user used to see only the raw APIError. Now prints:
🔐 Nous 401 — Portal authentication failed.
Response: <truncated body>
Most likely: Portal OAuth expired, account out of credits, or
agent key revoked.
Troubleshooting:
• Re-authenticate: hermes login --provider nous
• Check credits / billing: https://portal.nousresearch.com
• Verify stored credentials: $HERMES_HOME/auth.json
• Switch providers temporarily: /model <model> --provider openrouter
Addresses the common 'my hermes model hangs' pattern where the user's
Portal OAuth expired and the CLI gave no hint about the next step.
Adds schema v7 'api_call_count' column. run_agent.py increments it by 1
per LLM API call, web_server analytics SQL aggregates it, frontend uses
the real counter instead of summing sessions.
The 'API Calls' card on the analytics dashboard previously displayed
COUNT(*) from the sessions table — the number of conversations, not
LLM requests. Each session makes 10-90 API calls through the tool loop,
so the reported number was ~30x lower than real.
Salvaged from PR #10140 (@kshitijk4poor). The cache-token accuracy
portions of the original PR were deferred — per-provider analytics is
the better path there, since cache_write_tokens and actual_cost_usd
are only reliably available from a subset of providers (Anthropic
native, Codex Responses, OpenRouter with usage.include).
Tests:
- schema_version v7 assertion
- migration v2 -> v7 adds api_call_count column with default 0
- update_token_counts increments api_call_count by provided delta
- absolute=True sets api_call_count directly
- /api/analytics/usage exposes total_api_calls in totals
- Add configurable retain_tags / retain_source / retain_user_prefix /
retain_assistant_prefix knobs for native Hindsight.
- Thread gateway session identity (user_name, chat_id, chat_name,
chat_type, thread_id) through AIAgent and MemoryManager into
MemoryProvider.initialize kwargs so providers can scope and tag
retained memories.
- Hindsight attaches the new identity fields as retain metadata,
merges per-call tool tags with configured default tags, and uses
the configurable transcript labels for auto-retained turns.
Co-authored-by: Abner <abner.the.foreman@agentmail.to>
Fourth and final transport — completes the transport layer with all four
api_modes covered. Wraps agent/bedrock_adapter.py behind the ProviderTransport
ABC, handles both raw boto3 dicts and already-normalized SimpleNamespace.
Wires all transport methods to production paths in run_agent.py:
- build_kwargs: _build_api_kwargs bedrock branch
- validate_response: response validation, new bedrock_converse branch
- finish_reason: new bedrock_converse branch in finish_reason extraction
Based on PR #13467 by @kshitijk4poor, with one adjustment: the main normalize
loop does NOT add a bedrock_converse branch to invoke normalize_response on
the already-normalized response. Bedrock's normalize_converse_response runs
at the dispatch site (run_agent.py:5189), so the response already has the
OpenAI-compatible .choices[0].message shape by the time the main loop sees
it. Falling through to the chat_completions else branch is correct and
sidesteps a redundant NormalizedResponse rebuild.
Transport coverage — complete:
| api_mode | Transport | build_kwargs | normalize | validate |
|--------------------|--------------------------|:------------:|:---------:|:--------:|
| anthropic_messages | AnthropicTransport | ✅ | ✅ | ✅ |
| codex_responses | ResponsesApiTransport | ✅ | ✅ | ✅ |
| chat_completions | ChatCompletionsTransport | ✅ | ✅ | ✅ |
| bedrock_converse | BedrockTransport | ✅ | ✅ | ✅ |
17 new BedrockTransport tests pass. 117 transport tests total pass.
160 bedrock/converse tests across tests/agent/ pass. Full tests/run_agent/
targeted suite passes (885/885 + 15 skipped; the 1 remaining failure is the
pre-existing test_concurrent_interrupt flake on origin/main).
Third concrete transport — handles the default 'chat_completions' api_mode used
by ~16 OpenAI-compatible providers (OpenRouter, Nous, NVIDIA, Qwen, Ollama,
DeepSeek, xAI, Kimi, custom, etc.). Wires build_kwargs + validate_response to
production paths.
Based on PR #13447 by @kshitijk4poor, with fixes:
- Preserve tool_call.extra_content (Gemini thought_signature) via
ToolCall.provider_data — the original shim stripped it, causing 400 errors
on multi-turn Gemini 3 thinking requests.
- Preserve reasoning_content distinctly from reasoning (DeepSeek/Moonshot) so
the thinking-prefill retry check (_has_structured) still triggers.
- Port Kimi/Moonshot quirks (32000 max_tokens, top-level reasoning_effort,
extra_body.thinking) that landed on main after the original PR was opened.
- Keep _qwen_prepare_chat_messages_inplace alive and call it through the
transport when sanitization already deepcopied (avoids a second deepcopy).
- Skip the back-compat SimpleNamespace shim in the main normalize loop — for
chat_completions, response.choices[0].message is already the right shape
with .content/.tool_calls/.reasoning/.reasoning_content/.reasoning_details
and per-tool-call .extra_content from the OpenAI SDK.
run_agent.py: -239 lines in _build_api_kwargs default branch extracted to the
transport. build_kwargs now owns: codex-field sanitization, Qwen portal prep,
developer role swap, provider preferences, max_tokens resolution (ephemeral >
user > NVIDIA 16384 > Qwen 65536 > Kimi 32000 > anthropic_max_output), Kimi
reasoning_effort + extra_body.thinking, OpenRouter/Nous/GitHub reasoning,
Nous product attribution tags, Ollama num_ctx, custom-provider think=false,
Qwen vl_high_resolution_images, request_overrides.
39 new transport tests (8 build_kwargs, 5 Kimi, 4 validate, 4 normalize
including extra_content regression, 3 cache stats, 3 basic). Tests/run_agent/
targeted suite passes (885/885 + 15 skipped; the 1 remaining failure is the
test_concurrent_interrupt flake present on origin/main).
Add ResponsesApiTransport wrapping codex_responses_adapter.py behind the
ProviderTransport ABC. Auto-registered via _discover_transports().
Wire ALL Codex transport methods to production paths in run_agent.py:
- build_kwargs: main _build_api_kwargs codex branch (50 lines extracted)
- normalize_response: main loop + flush + summary + retry (4 sites)
- convert_tools: memory flush tool override
- convert_messages: called internally via build_kwargs
- validate_response: response validation gate
- preflight_kwargs: request sanitization (2 sites)
Remove 7 dead legacy wrappers from AIAgent (_responses_tools,
_chat_messages_to_responses_input, _normalize_codex_response,
_preflight_codex_api_kwargs, _preflight_codex_input_items,
_extract_responses_message_text, _extract_responses_reasoning_text).
Keep 3 ID manipulation methods still used by _build_assistant_message.
Update 18 test call sites across 3 test files to call adapter functions
directly instead of through deleted AIAgent wrappers.
24 new tests. 343 codex/responses/transport tests pass (0 failures).
PR 4 of the provider transport refactor.
* feat(models): hide OpenRouter models that don't advertise tool support
Port from Kilo-Org/kilocode#9068.
hermes-agent is tool-calling-first — every provider path assumes the
model can invoke tools. Models whose OpenRouter supported_parameters
doesn't include 'tools' (e.g. image-only or completion-only models)
cannot be driven by the agent loop and fail at the first tool call.
Filter them out of fetch_openrouter_models() so they never appear in
the model picker (`hermes model`, setup wizard, /model slash command).
Permissive when the field is missing — OpenRouter-compatible gateways
(Nous Portal, private mirrors, older snapshots) don't always populate
supported_parameters. Treat missing as 'unknown → allow' rather than
silently emptying the picker on those gateways. Only hide models
whose supported_parameters is an explicit list that omits tools.
Tests cover: tools present → kept, tools absent → dropped, field
missing → kept, malformed non-list → kept, non-dict item → kept,
empty list → dropped.
* feat(delegate): cross-agent file state coordination for concurrent subagents
Prevents mangled edits when concurrent subagents touch the same file
(same process, same filesystem — the mangle scenario from #11215).
Three layers, all opt-out via HERMES_DISABLE_FILE_STATE_GUARD=1:
1. FileStateRegistry (tools/file_state.py) — process-wide singleton
tracking per-agent read stamps and the last writer globally.
check_stale() names the sibling subagent in the warning when a
non-owning agent wrote after this agent's last read.
2. Per-path threading.Lock wrapped around the read-modify-write
region in write_file_tool and patch_tool. Concurrent siblings on
the same path serialize; different paths stay fully parallel.
V4A multi-file patches lock in sorted path order (deadlock-free).
3. Delegate-completion reminder in tools/delegate_tool.py: after a
subagent returns, writes_since(parent, child_start, parent_reads)
appends '[NOTE: subagent modified files the parent previously
read — re-read before editing: ...]' to entry.summary when the
child touched anything the parent had already seen.
Complements (does not replace) the existing path-overlap check in
run_agent._should_parallelize_tool_batch — batch check prevents
same-file parallel dispatch within one agent's turn (cheap prevention,
zero API cost), registry catches cross-subagent and cross-turn
staleness at write time (detection).
Behavior is warning-only, not hard-failing — matches existing project
style. Errors surface naturally: sibling writes often invalidate the
old_string in patch operations, which already errors cleanly.
Tests: tests/tools/test_file_state_registry.py — 16 tests covering
registry state transitions, per-path locking, per-path-not-global
locking, writes_since filtering, kill switch, and end-to-end
integration through the real read_file/write_file/patch handlers.
Adds role='leaf'|'orchestrator' to delegate_task. With max_spawn_depth>=2,
an orchestrator child retains the 'delegation' toolset and can spawn its
own workers; leaf children cannot delegate further (identical to today).
Default posture is flat — max_spawn_depth=1 means a depth-0 parent's
children land at the depth-1 floor and orchestrator role silently
degrades to leaf. Users opt into nested delegation by raising
max_spawn_depth to 2 or 3 in config.yaml.
Also threads acp_command/acp_args through the main agent loop's delegate
dispatch (previously silently dropped in the schema) via a new
_dispatch_delegate_task helper, and adds a DelegateEvent enum with
legacy-string back-compat for gateway/ACP/CLI progress consumers.
Config (hermes_cli/config.py defaults):
delegation.max_concurrent_children: 3 # floor-only, no upper cap
delegation.max_spawn_depth: 1 # 1=flat (default), 2-3 unlock nested
delegation.orchestrator_enabled: true # global kill switch
Salvaged from @pefontana's PR #11215. Overrides vs. the original PR:
concurrency stays at 3 (PR bumped to 5 + cap 8 — we keep the floor only,
no hard ceiling); max_spawn_depth defaults to 1 (PR defaulted to 2 which
silently enabled one level of orchestration for every user).
Co-authored-by: pefontana <fontana.pedro93@gmail.com>
Reported during the TUI v2 blitz test: switching from openrouter to
anthropic via `/model <name> --provider anthropic` appeared to succeed,
but the next turn kept hitting openrouter — the provider the user was
deliberately moving away from.
Two gaps caused this:
1. `Agent.switch_model` reset `_fallback_activated` / `_fallback_index`
but left `_fallback_chain` intact. The chain was seeded from
`fallback_providers:` at agent init for the *original* primary, so
when the new primary returned 401 (invalid/expired Anthropic key),
`_try_activate_fallback()` picked the old provider back up without
informing the user. Prune entries matching either the old primary
(user is moving away) or the new primary (redundant) whenever the
primary provider actually changes.
2. `_apply_model_switch` persisted `HERMES_MODEL` but never updated
`HERMES_INFERENCE_PROVIDER`. Any ambient re-resolution of the runtime
(credential pool refresh, compressor rebuild, aux clients) falls
through to that env var in `resolve_requested_provider`, so it kept
reporting the original provider even after an in-memory switch.
Adds three regression tests: fallback-chain prune on primary change,
no-op on same-provider model swap, and env-var sync on explicit switch.
The 💾 Cache footer was gated on `self._use_prompt_caching`, which is
only True for Anthropic marker injection (native Anthropic, OpenRouter
Claude, Anthropic-wire gateways, Qwen on OpenCode/Alibaba). Providers
with automatic server-side prefix caching — OpenAI, Kimi, DeepSeek,
Qwen on OpenRouter — return `prompt_tokens_details.cached_tokens` too,
but users couldn't see their cache % because the display path never
fired for them. Result: people couldn't tell their cache was working or
broken without grepping agent.log.
`canonical_usage` from `normalize_usage()` already unifies all three
API shapes (Anthropic / Codex Responses / OpenAI chat completions) into
`cache_read_tokens` and `cache_write_tokens`. Drop the gate and read
from there — now the footer fires whenever the provider reported any
cached or written tokens, regardless of whether hermes injected markers.
Also removes duplicated branch-per-API-shape extraction code.
Qwen models on OpenCode, OpenCode Go, and direct DashScope accept
Anthropic-style cache_control markers on OpenAI-wire chat completions,
but hermes only injected markers for Claude-named models. Result: zero
cache hits on every turn, full prompt re-billed — a community user
reported burning through their OpenCode Go subscription on Qwen3.6.
Extend _anthropic_prompt_cache_policy to return (True, False) — envelope
layout, not native — for the Alibaba provider family when the model name
contains 'qwen'. Envelope layout places markers on inner content blocks
(matching pi-mono's 'alibaba' cacheControlFormat) and correctly skips
top-level markers on tool-role messages (which OpenCode rejects).
Non-Qwen models on these providers (GLM, Kimi) keep their existing
behaviour — they have automatic server-side caching and don't need
client markers.
Upstream reference: pi-mono #3392 / #3393 documented this contract for
opencode-go Qwen models.
Adds 7 regression tests covering Qwen3.5/3.6/coder on each affected
provider plus negative cases for GLM/Kimi/OpenRouter-Qwen.
Two call sites still used a raw substring check to identify ollama.com:
hermes_cli/runtime_provider.py:496:
_is_ollama_url = "ollama.com" in base_url.lower()
run_agent.py:6127:
if fb_base_url_hint and "ollama.com" in fb_base_url_hint.lower() ...
Same bug class as GHSA-xf8p-v2cg-h7h5 (OpenRouter substring leak), which
was fixed in commit dbb7e00e via base_url_host_matches() across the
codebase. The earlier sweep missed these two Ollama sites. Self-discovered
during April 2026 security-advisory triage; filed as GHSA-76xc-57q6-vm5m.
Impact is narrow — requires a user with OLLAMA_API_KEY configured AND a
custom base_url whose path or look-alike host contains 'ollama.com'.
Users on default provider flows are unaffected. Filed as a draft advisory
to use the private-fork flow; not CVE-worthy on its own.
Fix is mechanical: replace substring check with base_url_host_matches
at both sites. Same helper the rest of the codebase uses.
Tests: 67 -> 71 passing. 7 new host-matcher cases in
tests/test_base_url_hostname.py (path injection, lookalike host,
localtest.me subdomain, ollama.ai TLD confusion, localhost, genuine
ollama.com, api.ollama.com subdomain) + 4 call-site tests in
tests/hermes_cli/test_runtime_provider_resolution.py verifying
OLLAMA_API_KEY is selected only when base_url actually targets
ollama.com.
Fixes GHSA-76xc-57q6-vm5m
Kimi/Moonshot endpoints require explicit parameters that Hermes was not
sending, causing 'Response truncated due to output length limit' errors
and inconsistent reasoning behavior.
Root cause analysis against Kimi CLI source (MoonshotAI/kimi-cli,
packages/kosong/src/kosong/chat_provider/kimi.py):
1. max_tokens: Kimi's API defaults to a very low value when omitted.
Reasoning tokens share the output budget — the model exhausts it on
thinking alone. Send 32000, matching Kimi CLI's generate() default.
2. reasoning_effort: Kimi CLI sends this as a top-level parameter (not
inside extra_body). Hermes was not sending it at all because
_supports_reasoning_extra_body() returns False for non-OpenRouter
endpoints.
3. extra_body.thinking: Kimi CLI uses with_thinking() which sets
extra_body.thinking={"type":"enabled"} alongside reasoning_effort.
This is a separate control from the OpenAI-style reasoning extra_body
that Hermes sends for OpenRouter/GitHub. Without it, the Kimi gateway
may not activate reasoning mode correctly.
Covers api.kimi.com (Kimi Code) and api.moonshot.ai/cn (Moonshot).
Tests: 6 new test cases for max_tokens, reasoning_effort, and
extra_body.thinking under various configs.
Full AST-based scan of all .py files to find every case where a module
or name is imported locally inside a function body but is already
available at module level. This is the second pass — the first commit
handled the known cases from the lint report; this one catches
everything else.
Files changed (19):
cli.py — 16 removals: time as _time/_t/_tmod (×10),
re / re as _re (×2), os as _os, sys,
partial os from combo import,
from model_tools import get_tool_definitions
gateway/run.py — 8 removals: MessageEvent as _ME /
MessageType as _MT (×3), os as _os2,
MessageEvent+MessageType (×2), Platform,
BasePlatformAdapter as _BaseAdapter
run_agent.py — 6 removals: get_hermes_home as _ghh,
partial (contextlib, os as _os),
cleanup_vm, cleanup_browser,
set_interrupt as _sif (×2),
partial get_toolset_for_tool
hermes_cli/main.py — 4 removals: get_hermes_home, time as _time,
logging as _log, shutil
hermes_cli/config.py — 1 removal: get_hermes_home as _ghome
hermes_cli/runtime_provider.py
— 1 removal: load_config as _load_bedrock_config
hermes_cli/setup.py — 2 removals: importlib.util (×2)
hermes_cli/nous_subscription.py
— 1 removal: from hermes_cli.config import load_config
hermes_cli/tools_config.py
— 1 removal: from hermes_cli.config import load_config, save_config
cron/scheduler.py — 3 removals: concurrent.futures, json as _json,
from hermes_cli.config import load_config
batch_runner.py — 1 removal: list_distributions as get_all_dists
(kept print_distribution_info, not at top level)
tools/send_message_tool.py
— 2 removals: import os (×2)
tools/skills_tool.py — 1 removal: logging as _logging
tools/browser_camofox.py
— 1 removal: from hermes_cli.config import load_config
tools/image_generation_tool.py
— 1 removal: import fal_client
environments/tool_context.py
— 1 removal: concurrent.futures
gateway/platforms/bluebubbles.py
— 1 removal: httpx as _httpx
gateway/platforms/whatsapp.py
— 1 removal: import asyncio
tui_gateway/server.py — 2 removals: from datetime import datetime,
import time
All alias references (_time, _t, _tmod, _re, _os, _os2, _json, _ghh,
_ghome, _sif, _ME, _MT, _BaseAdapter, _load_bedrock_config, _httpx,
_logging, _log, get_all_dists) updated to use the top-level names.
Sweep ~74 redundant local imports across 21 files where the same module
was already imported at the top level. Also includes type fixes and lint
cleanups on the same branch.
Add agent/transports/types.py with three shared dataclasses:
- NormalizedResponse: content, tool_calls, finish_reason, reasoning, usage, provider_data
- ToolCall: id, name, arguments, provider_data (per-tool-call protocol metadata)
- Usage: prompt_tokens, completion_tokens, total_tokens, cached_tokens
Add normalize_anthropic_response_v2() to anthropic_adapter.py — wraps the
existing v1 function and maps its output to NormalizedResponse. One call site
in run_agent.py (the main normalize branch) uses v2 with a back-compat shim
to SimpleNamespace for downstream code.
No ABC, no registry, no streaming, no client lifecycle. Those land in PR 3
with the first concrete transport (AnthropicTransport).
46 new tests:
- test_types.py: dataclass construction, build_tool_call, map_finish_reason
- test_anthropic_normalize_v2.py: v1-vs-v2 regression tests (text, tools,
thinking, mixed, stop reasons, mcp prefix stripping, edge cases)
Part of the provider transport refactor (PR 2 of 9).
The mid-run steer marker was '[USER STEER (injected mid-run, not tool
output): <text>]'. Replaced with a plain two-newline-prefixed
'User guidance: <text>' suffix.
Rationale: the marker lives inside the tool result's content string
regardless of whether the tool returned JSON, plain text, an MCP
result, or a plugin result. The bracketed tag read like structured
metadata that some tools (terminal, execute_code) could confuse with
their own output formatting. A plain labelled suffix works uniformly
across every content shape we produce.
Behavior unchanged:
- Still injected into the last tool-role message's content.
- Still preserves multimodal (Anthropic) content-block lists by
appending a text block.
- Still drained at both sites added in #12959 and #13205 — per-tool
drain between individual calls, and pre-API-call drain at the top
of each main-loop iteration.
Checked Codex's equivalent (pending_input / inject_user_message_without_turn
in codex-rs/core): they record mid-turn user input as a real role:user
message via record_user_prompt_and_emit_turn_item(). That's cleaner for
their Responses-API model but not portable to Chat Completions where
role alternation after tool_calls is strict. Embedding the guidance in
the last tool result remains the correct placement for us.
Validation: all 21 tests in tests/run_agent/test_steer.py pass.
Aslaaen's fix in the original PR covered _detect_api_mode_for_url and the
two openai/xai sites in run_agent.py. This finishes the sweep: the same
substring-match false-positive class (e.g. https://api.openai.com.evil/v1,
https://proxy/api.openai.com/v1, https://api.anthropic.com.example/v1)
existed in eight more call sites, and the hostname helper was duplicated
in two modules.
- utils: add shared base_url_hostname() (single source of truth).
- hermes_cli/runtime_provider, run_agent: drop local duplicates, import
from utils. Reuse the cached AIAgent._base_url_hostname attribute
everywhere it's already populated.
- agent/auxiliary_client: switch codex-wrap auto-detect, max_completion_tokens
gate (auxiliary_max_tokens_param), and custom-endpoint max_tokens kwarg
selection to hostname equality.
- run_agent: native-anthropic check in the Claude-style model branch
and in the AIAgent init provider-auto-detect branch.
- agent/model_metadata: Anthropic /v1/models context-length lookup.
- hermes_cli/providers.determine_api_mode: anthropic / openai URL
heuristics for custom/unknown providers (the /anthropic path-suffix
convention for third-party gateways is preserved).
- tools/delegate_tool: anthropic detection for delegated subagent
runtimes.
- hermes_cli/setup, hermes_cli/tools_config: setup-wizard vision-endpoint
native-OpenAI detection (paired with deduping the repeated check into
a single is_native_openai boolean per branch).
Tests:
- tests/test_base_url_hostname.py covers the helper directly
(path-containing-host, host-suffix, trailing dot, port, case).
- tests/hermes_cli/test_determine_api_mode_hostname.py adds the same
regression class for determine_api_mode, plus a test that the
/anthropic third-party gateway convention still wins.
Also: add asslaenn5@gmail.com → Aslaaen to scripts/release.py AUTHOR_MAP.
Requests through Vercel AI Gateway now carry referrerUrl / appName /
User-Agent attribution so traffic shows up in the gateway's analytics.
Adds _AI_GATEWAY_HEADERS in auxiliary_client and a new
ai-gateway.vercel.sh branch in _apply_client_headers_for_base_url.
Follow-up for salvaged PR #3185:
- run_agent.py: pass self.api_key to query_ollama_num_ctx() so Ollama
behind an auth proxy (same issue class as the LM Studio fix) can be
probed successfully.
- scripts/release.py AUTHOR_MAP: map @tannerfokkens-maker's local-hostname
commit email.
When /steer is sent during an API call (model thinking), the steer text
sits in _pending_steer until after the next tool batch — which may never
come if the model returns a final response. In that case the steer is
only delivered as a post-run follow-up, defeating the purpose.
Add a pre-API-call drain at the top of the main loop: before building
api_messages, check _pending_steer and inject into the last tool result
in the messages list. This ensures steers sent during model thinking are
visible on the very next API call.
If no tool result exists yet (first iteration), the steer is restashed
for the post-tool drain to pick up — injecting into a user message would
break role alternation.
Three new tests cover the pre-API-call drain: injection into last tool
result, restash when no tool message exists, and backward scan past
non-tool messages.
Kimi's gateway selects the correct temperature server-side based on the
active mode (thinking -> 1.0, non-thinking -> 0.6). Sending any
temperature value — even the previously "correct" one — conflicts with
gateway-managed defaults.
Replaces the old approach of forcing specific temperature values (0.6
for non-thinking, 1.0 for thinking) with an OMIT_TEMPERATURE sentinel
that tells all call sites to strip the temperature key from API kwargs
entirely.
Changes:
- agent/auxiliary_client.py: OMIT_TEMPERATURE sentinel, _is_kimi_model()
prefix check (covers all kimi-* models), _fixed_temperature_for_model()
returns sentinel for kimi models. _build_call_kwargs() strips temp.
- run_agent.py: _build_api_kwargs, flush_memories, and summary generation
paths all handle the sentinel by popping/omitting temperature.
- trajectory_compressor.py: _effective_temperature_for_model returns None
for kimi (sentinel mapped), direct client calls use kwargs dict to
conditionally include temperature.
- mini_swe_runner.py: same sentinel handling via wrapper function.
- 6 test files updated: all 'forces temperature X' assertions replaced
with 'temperature not in kwargs' assertions.
Net: -76 lines (171 added, 247 removed).
Inspired by PR #13137 (@kshitijk4poor).
Extract 12 Codex Responses API format-conversion and normalization functions
from run_agent.py into agent/codex_responses_adapter.py, following the
existing pattern of anthropic_adapter.py and bedrock_adapter.py.
run_agent.py: 12,550 → 11,865 lines (-685 lines)
Functions moved:
- _chat_content_to_responses_parts (multimodal content conversion)
- _summarize_user_message_for_log (multimodal message logging)
- _deterministic_call_id (cache-safe fallback IDs)
- _split_responses_tool_id (composite ID splitting)
- _derive_responses_function_call_id (fc_ prefix conversion)
- _responses_tools (schema format conversion)
- _chat_messages_to_responses_input (message format conversion)
- _preflight_codex_input_items (input validation)
- _preflight_codex_api_kwargs (API kwargs validation)
- _extract_responses_message_text (response text extraction)
- _extract_responses_reasoning_text (reasoning extraction)
- _normalize_codex_response (full response normalization)
All functions are stateless module-level functions. AIAgent methods remain
as thin one-line wrappers. Both module-level helpers are re-exported from
run_agent.py for backward compatibility with existing test imports.
Includes multimodal inline image support (PR #12969) that the original PR
was missing.
Based on PR #12975 by @kshitijk4poor.
Follow-up for PR #12252 salvage:
- Extract 75-line inline repair block to _repair_tool_call_arguments()
module-level helper for testability and readability
- Remove redundant 'import re as _re' (re already imported at line 33)
- Bound the while-True excess-delimiter removal loop to 50 iterations
- Add 17 tests covering all 6 repair stages
- Add sirEven to AUTHOR_MAP in release.py
Cherry-picked from PR #12252 by @sirEven.
Models like GLM-5.1 via Ollama can produce malformed tool_call arguments
(truncated JSON, trailing commas, Python None). The existing except
Exception: pass silently passes broken args to the API, which rejects
them with HTTP 400, crashing the session.
Adds a multi-stage repair pipeline at the pre-send normalization point:
1. Empty/whitespace-only → {}
2. Python None literal → {}
3. Strip trailing commas
4. Auto-close unclosed brackets
5. Remove excess closing delimiters
6. Last resort: replace with {} (logged at WARNING)
Cherry-picked from PR #12481 by @Sanjays2402.
Reasoning models (GLM-5.1, QwQ, DeepSeek R1) inflate completion_tokens
with internal thinking tokens. The compression trigger summed
prompt_tokens + completion_tokens, causing premature compression at ~42%
actual context usage instead of the configured 50% threshold.
Now uses only prompt_tokens — completion tokens don't consume context
window space for the next API call.
- 3 new regression tests
- Added AUTHOR_MAP entry for @Sanjays2402
Closes#12026
OpenAI-compatible clients (Open WebUI, LobeChat, etc.) can now send vision
requests to the API server. Both endpoints accept the canonical OpenAI
multimodal shape:
Chat Completions: {type: text|image_url, image_url: {url, detail?}}
Responses: {type: input_text|input_image, image_url: <str>, detail?}
The server validates and converts both into a single internal shape that the
existing agent pipeline already handles (Anthropic adapter converts,
OpenAI-wire providers pass through). Remote http(s) URLs and data:image/*
URLs are supported.
Uploaded files (file, input_file, file_id) and non-image data: URLs are
rejected with 400 unsupported_content_type.
Changes:
- gateway/platforms/api_server.py
- _normalize_multimodal_content(): validates + normalizes both Chat and
Responses content shapes. Returns a plain string for text-only content
(preserves prompt-cache behavior on existing callers) or a canonical
[{type:text|image_url,...}] list when images are present.
- _content_has_visible_payload(): replaces the bare truthy check so a
user turn with only an image no longer rejects as 'No user message'.
- _handle_chat_completions and _handle_responses both call the new helper
for user/assistant content; system messages continue to flatten to text.
- Codex conversation_history, input[], and inline history paths all share
the same validator. No duplicated normalizers.
- run_agent.py
- _summarize_user_message_for_log(): produces a short string summary
('[1 image] describe this') from list content for logging, spinner
previews, and trajectory writes. Fixes AttributeError when list
user_message hit user_message[:80] + '...' / .replace().
- _chat_content_to_responses_parts(): module-level helper that converts
chat-style multimodal content to Responses 'input_text'/'input_image'
parts. Used in _chat_messages_to_responses_input for Codex routing.
- _preflight_codex_input_items() now validates and passes through list
content parts for user/assistant messages instead of stringifying.
- tests/gateway/test_api_server_multimodal.py (new, 38 tests)
- Unit coverage for _normalize_multimodal_content, including both part
formats, data URL gating, and all reject paths.
- Real aiohttp HTTP integration on /v1/chat/completions and /v1/responses
verifying multimodal payloads reach _run_agent intact.
- 400 coverage for file / input_file / non-image data URL.
- tests/run_agent/test_run_agent_multimodal_prologue.py (new)
- Regression coverage for the prologue no-crash contract.
- _chat_content_to_responses_parts round-trip coverage.
- website/docs/user-guide/features/api-server.md
- Inline image examples for both endpoints.
- Updated Limitations: files still unsupported, images now supported.
Validated live against openrouter/anthropic/claude-opus-4.6:
POST /v1/chat/completions → 200, vision-accurate description
POST /v1/responses → 200, same image, clean output_text
POST /v1/chat/completions [file] → 400 unsupported_content_type
POST /v1/responses [input_file] → 400 unsupported_content_type
POST /v1/responses [non-image data URL] → 400 unsupported_content_type
Closes#5621, #8253, #4046, #6632.
Co-authored-by: Paul Bergeron <paul@gamma.app>
Co-authored-by: zhangxicen <zhangxicen@example.com>
Co-authored-by: Manuel Schipper <manuelschipper@users.noreply.github.com>
Co-authored-by: pradeep7127 <pradeep7127@users.noreply.github.com>
Previously, /steer text was only injected after an entire tool batch
completed (_execute_tool_calls_sequential/concurrent returned). If the
batch had a long-running tool (delegate_task, terminal build), the
steer waited for ALL tools to finish before landing — functionally
identical to /queue from the user's perspective.
Now _apply_pending_steer_to_tool_results() is called after EACH
individual tool result is appended to messages, in both the sequential
and concurrent paths. A steer arriving during Tool 1 lands in Tool 1's
result before Tool 2 starts executing.
Also handles leftover steers in the gateway: if a steer arrives during
the final API call (no tool batch to drain into), it's now delivered as
the next user turn instead of being silently dropped.
Fixes user report from Utku.
Context compression silently failed when the auxiliary compression model's
context window was smaller than the main model's compression threshold
(e.g. GLM-4.5-air at 131k paired with a 150k threshold). The feasibility
check warned but the session kept running and compression attempts errored
out mid-conversation.
Two changes in _check_compression_model_feasibility():
1. Hard floor: if detected aux context < MINIMUM_CONTEXT_LENGTH (64k),
raise ValueError so the session refuses to start. Mirrors the existing
main-model rejection at AIAgent.__init__ line 1600. A compression model
below 64k cannot summarise a full threshold-sized window.
2. Auto-correct: when aux context is >= 64k but below the computed
threshold, lower the live compressor's threshold_tokens to aux_context
(and update threshold_percent to match so later update_model() calls
stay in sync). Warning reworded to say what was done and how to
persist the fix in config.yaml.
Only ValueError re-raises; other exceptions in the check remain swallowed
as non-fatal.
Third-party gateways that speak the native Anthropic protocol (MiniMax,
Zhipu GLM, Alibaba DashScope, Kimi, LiteLLM proxies) now work end-to-end
with the same feature set as direct api.anthropic.com callers. Synthesizes
eight stale community PRs into one consolidated change.
Five fixes:
- URL detection: consolidate three inline `endswith("/anthropic")`
checks in runtime_provider.py into the shared _detect_api_mode_for_url
helper. Third-party /anthropic endpoints now auto-resolve to
api_mode=anthropic_messages via one code path instead of three.
- OAuth leak-guard: all five sites that assign `_is_anthropic_oauth`
(__init__, switch_model, _try_refresh_anthropic_client_credentials,
_swap_credential, _try_activate_fallback) now gate on
`provider == "anthropic"` so a stale ANTHROPIC_TOKEN never trips
Claude-Code identity injection on third-party endpoints. Previously
only 2 of 5 sites were guarded.
- Prompt caching: new method `_anthropic_prompt_cache_policy()` returns
`(should_cache, use_native_layout)` per endpoint. Replaces three
inline conditions and the `native_anthropic=(api_mode=='anthropic_messages')`
call-site flag. Native Anthropic and third-party Anthropic gateways
both get the native cache_control layout; OpenRouter gets envelope
layout. Layout is persisted in `_primary_runtime` so fallback
restoration preserves the per-endpoint choice.
- Auxiliary client: `_try_custom_endpoint` honors
`api_mode=anthropic_messages` and builds `AnthropicAuxiliaryClient`
instead of silently downgrading to an OpenAI-wire client. Degrades
gracefully to OpenAI-wire when the anthropic SDK isn't installed.
- Config hygiene: `_update_config_for_provider` (hermes_cli/auth.py)
clears stale `api_key`/`api_mode` when switching to a built-in
provider, so a previous MiniMax custom endpoint's credentials can't
leak into a later OpenRouter session.
- Truncation continuation: length-continuation and tool-call-truncation
retry now cover `anthropic_messages` in addition to `chat_completions`
and `bedrock_converse`. Reuses the existing `_build_assistant_message`
path via `normalize_anthropic_response()` so the interim message
shape is byte-identical to the non-truncated path.
Tests: 6 new files, 42 test cases. Targeted run + tests/run_agent,
tests/agent, tests/hermes_cli all pass (4554 passed).
Synthesized from (credits preserved via Co-authored-by trailers):
#7410 @nocoo — URL detection helper
#7393 @keyuyuan — OAuth 5-site guard
#7367 @n-WN — OAuth guard (narrower cousin, kept comment)
#8636 @sgaofen — caching helper + native-vs-proxy layout split
#10954 @Only-Code-A — caching on anthropic_messages+Claude
#7648 @zhongyueming1121 — aux client anthropic_messages branch
#6096 @hansnow — /model switch clears stale api_mode
#9691 @TroyMitchell911 — anthropic_messages truncation continuation
Closes: #7366, #8294 (third-party Anthropic identity + caching).
Supersedes: #7410, #7367, #7393, #8636, #10954, #7648, #6096, #9691.
Rejects: #9621 (OpenAI-wire caching with incomplete blocklist — risky),
#7242 (superseded by #9691, stale branch),
#8321 (targets smart_model_routing which was removed in #12732).
Co-authored-by: nocoo <nocoo@users.noreply.github.com>
Co-authored-by: Keyu Yuan <leoyuan0099@gmail.com>
Co-authored-by: Zoee <30841158+n-WN@users.noreply.github.com>
Co-authored-by: sgaofen <135070653+sgaofen@users.noreply.github.com>
Co-authored-by: Only-Code-A <bxzt2006@163.com>
Co-authored-by: zhongyueming <mygamez@163.com>
Co-authored-by: Xiaohan Li <hansnow@users.noreply.github.com>
Co-authored-by: Troy Mitchell <i@troy-y.org>
Bedrock rejects ``global-anthropic-claude-opus-4-7`` with ``HTTP 400:
The provided model identifier is invalid`` because its inference
profile IDs embed structural dots
(``global.anthropic.claude-opus-4-7``) that ``normalize_model_name``
was converting to hyphens. ``AIAgent._anthropic_preserve_dots`` did
not include ``bedrock`` in its provider allowlist, so every Claude-on-
Bedrock request through the AnthropicBedrock SDK path shipped with
the mangled model ID and failed.
Root cause
----------
``run_agent.py:_anthropic_preserve_dots`` (previously line 6589)
controls whether ``agent.anthropic_adapter.normalize_model_name``
converts dots to hyphens. The function listed Alibaba, MiniMax,
OpenCode Go/Zen and ZAI but not Bedrock, so when a user set
``provider: bedrock`` with a dotted inference-profile model the flag
returned False and ``normalize_model_name`` mangled every dot in the
ID. All four call sites in run_agent.py
(``build_anthropic_kwargs`` + three fallback / review / summary paths
at lines 6707, 7343, 8408, 8440) read from this same helper.
The bug shape matches #5211 for opencode-go, which was fixed in commit
f77be22c by extending this same allowlist.
Fix
---
* Add ``"bedrock"`` to the provider allowlist.
* Add ``"bedrock-runtime."`` to the base-URL heuristic as
defense-in-depth, so a custom-provider-shaped config with
``base_url: https://bedrock-runtime.<region>.amazonaws.com`` also
takes the preserve-dots path even if ``provider`` isn't explicitly
set to ``"bedrock"``. This mirrors how the code downstream at
run_agent.py:759 already treats either signal as "this is Bedrock".
Bedrock model ID shapes covered
-------------------------------
| Shape | Preserved |
| --- | --- |
| ``global.anthropic.claude-opus-4-7`` (reporter's exact ID) | ✓ |
| ``us.anthropic.claude-sonnet-4-5-20250929-v1:0`` | ✓ |
| ``apac.anthropic.claude-haiku-4-5`` | ✓ |
| ``anthropic.claude-3-5-sonnet-20241022-v2:0`` (foundation) | ✓ |
| ``eu.anthropic.claude-3-5-sonnet`` (regional inference profile) | ✓ |
Non-Claude Bedrock models (Nova, Llama, DeepSeek) take the
``bedrock_converse`` / boto3 path which does not call
``normalize_model_name``, so they were never affected by this bug
and remain unaffected by the fix.
Narrow scope — explicitly not changed
-------------------------------------
* ``bedrock_converse`` path (non-Claude Bedrock models) — already
correct; no ``normalize_model_name`` in that pipeline.
* Provider aliases (``aws``, ``aws-bedrock``, ``amazon``,
``amazon-bedrock``) — if a user bypasses the alias-normalization
pipeline and passes ``provider="aws"`` directly, the base-URL
heuristic still catches it because Bedrock always uses a
``bedrock-runtime.`` endpoint. Adding the aliases themselves to the
provider set is cheap but would be scope creep for this fix.
* No other places in ``agent/anthropic_adapter.py`` mangle dots, so
the fix is confined to ``_anthropic_preserve_dots``.
Regression coverage
-------------------
``tests/agent/test_bedrock_integration.py`` gains three new classes:
* ``TestBedrockPreserveDotsFlag`` (5 tests): flag returns True for
``provider="bedrock"`` and for Bedrock runtime URLs (us-east-1 and
ap-northeast-2 — the reporter's region); returns False for non-
Bedrock AWS URLs like ``s3.us-east-1.amazonaws.com``; canary that
Anthropic-native still returns False.
* ``TestBedrockModelNameNormalization`` (5 tests): every documented
Bedrock model-ID shape survives ``normalize_model_name`` with the
flag on; inverse canary pins that ``preserve_dots=False`` still
mangles (so a future refactor can't decouple the flag from its
effect).
* ``TestBedrockBuildAnthropicKwargsEndToEnd`` (2 tests): integration
through ``build_anthropic_kwargs`` shows the reporter's exact model
ID ends up unmangled in the outgoing kwargs.
Three of the new flag tests fail on unpatched ``origin/main`` with
``assert False is True`` (preserve-dots returning False for Bedrock),
confirming the regression is caught.
Validation
----------
``source venv/bin/activate && python -m pytest
tests/agent/test_bedrock_integration.py tests/agent/test_minimax_provider.py
-q`` -> 84 passed (40 new bedrock tests + 44 pre-existing, including
the minimax canaries that pin the pattern this fix mirrors).
CI-aligned broad suite: 12827 passed, 39 skipped, 19 pre-existing
baseline failures (all reproduce on clean ``origin/main``; none in
the touched code path).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Follow up salvaged PR #12668 by threading base_url through the
remaining direct-call sites so kimi-k2.5 uses temperature=1.0 on
api.moonshot.ai and keeps 0.6 on api.kimi.com/coding. Add focused
regression tests for run_agent, trajectory_compressor, and
mini_swe_runner.
- only use the native adapter for the canonical Gemini native endpoint
- keep custom and /openai base URLs on the OpenAI-compatible path
- preserve Hermes keepalive transport injection for native Gemini clients
- stabilize streaming tool-call replay across repeated SSE events
- add follow-up tests for base_url precedence, async streaming, and duplicate tool-call chunks
- add a native Gemini adapter over generateContent/streamGenerateContent
- switch the built-in gemini provider off the OpenAI-compatible endpoint
- preserve thought signatures and native functionResponse replay
- route auxiliary Gemini clients through the same adapter
- add focused unit coverage plus native-provider integration checks
The cherry-picked salvage (admin28980's commit) added codex headers only on the
primary chat client path, with two inaccuracies:
- originator was 'hermes-agent' — Cloudflare whitelists codex_cli_rs,
codex_vscode, codex_sdk_ts, and Codex* prefixes. 'hermes-agent' isn't on
the list, so the header had no mitigating effect on the 403 (the
account-id header alone may have been carrying the fix).
- account-id header was 'ChatGPT-Account-Id' — upstream codex-rs auth.rs
uses canonical 'ChatGPT-Account-ID' (PascalCase, trailing -ID).
Also, the auxiliary client (_try_codex + resolve_provider_client raw_codex
branch) constructs OpenAI clients against the same chatgpt.com endpoint with
no default headers at all — so compression, title generation, vision, session
search, and web_extract all still 403 from VPS IPs.
Consolidate the header set into _codex_cloudflare_headers() in
agent/auxiliary_client.py (natural home next to _read_codex_access_token and
the existing JWT decode logic) and call it from all four insertion points:
- run_agent.py: AIAgent.__init__ (initial construction)
- run_agent.py: _apply_client_headers_for_base_url (credential rotation)
- agent/auxiliary_client.py: _try_codex (aux client)
- agent/auxiliary_client.py: resolve_provider_client raw_codex branch
Net: -36/+55 lines, -25 lines of duplicated inline JWT decode replaced by a
single helper. User-Agent switched to 'codex_cli_rs/0.0.0 (Hermes Agent)' to
match the codex-rs shape while keeping product attribution.
Tests in tests/agent/test_codex_cloudflare_headers.py cover:
- originator value, User-Agent shape, canonical header casing
- account-ID extraction from a real JWT fixture
- graceful handling of malformed / non-string / claim-missing tokens
- wiring at all four insertion points (primary init, rotation, both aux paths)
- non-chatgpt base URLs (openrouter) do NOT get codex headers
- switching away from chatgpt.com drops the headers
Add ChatGPT-Account-Id and originator headers when using chatgpt.com
backend-api endpoint. Matches official codex-rs CLI behavior to prevent
Cloudflare JavaScript challenges on non-residential IPs (VPS, Mac Mini,
always-on servers).
Applied in AIAgent.__init__ and _update_base_url_headers to cover both
initial setup and credential rotation paths.
When creating httpx.Client with a custom transport for TCP keepalive,
proxy environment variables (HTTP_PROXY, HTTPS_PROXY) were ignored because
httpx only auto-reads them when transport=None.
Add _get_proxy_from_env() to explicitly read proxy settings and pass them
to httpx.Client, ensuring providers like kimi-coding-cn work correctly
when behind a proxy.
Fixes connection errors when HTTP_PROXY/HTTPS_PROXY are set.
Live test with timeout_seconds: 0.5 on claude-sonnet-4.6 proved the
initial wiring was insufficient: run_agent.py was overriding the
client-level timeout on every call via hardcoded per-request kwargs.
Root cause: run_agent.py had two sites that pass an explicit timeout=
kwarg into chat.completions.create() — api_kwargs['timeout'] at line
7075 (HERMES_API_TIMEOUT=1800s default) and the streaming path's
_httpx.Timeout(..., read=HERMES_STREAM_READ_TIMEOUT=120s, ...) at line
5760. Both override the per-provider config value the client was
constructed with, so a 0.5s config timeout would silently not enforce.
This commit:
- Adds AIAgent._resolved_api_call_timeout() — config > HERMES_API_TIMEOUT env > 1800s default.
- Uses it for the non-streaming api_kwargs['timeout'] field.
- Uses it for the streaming path's httpx.Timeout(connect, read, write, pool)
so both connect and read respect the configured value when set.
Local-provider auto-bump (Ollama/vLLM cold-start) only applies when
no explicit config value is set.
- New test: test_resolved_api_call_timeout_priority covers all three
precedence cases (config, env, default).
Live verified: 0.5s config on claude-sonnet-4.6 now triggers
APITimeoutError at ~3s per retry, exhausts 3 retries in ~15s total
(was: 29-47s success with timeout ignored). Positive case (60s config
+ gpt-4o-mini) still succeeds at 1.3s.
Follow-up on top of mvanhorn's cherry-picked commit. Original PR only
wired request_timeout_seconds into the explicit-creds OpenAI branch at
run_agent.py init; router-based implicit auth, native Anthropic, and the
fallback chain were still hardcoded to SDK defaults.
- agent/anthropic_adapter.py: build_anthropic_client() accepts an optional
timeout kwarg (default 900s preserved when unset/invalid).
- run_agent.py: resolve per-provider/per-model timeout once at init; apply
to Anthropic native init + post-refresh rebuild + stale/interrupt
rebuilds + switch_model + _restore_primary_runtime + the OpenAI
implicit-auth path + _try_activate_fallback (with immediate client
rebuild so the first fallback request carries the configured timeout).
- tests: cover anthropic adapter kwarg honoring; widen mock signatures
to accept the new timeout kwarg.
- docs/example: clarify that the knob now applies to every transport,
the fallback chain, and rebuilds after credential rotation.
Adds optional providers.<id>.request_timeout_seconds and
providers.<id>.models.<model>.timeout_seconds config, resolved via a new
hermes_cli/timeouts.py helper and applied where client_kwargs is built
in run_agent.py. Zero default behavior change: when both keys are unset,
the openai SDK default takes over.
Mirrors the existing _get_task_timeout pattern in agent/auxiliary_client.py
for auxiliary tasks - the primary turn path just never got the equivalent
knob.
Cross-project demand: openclaw/openclaw#43946 (17 reactions) asks for
exactly this config - specifically calls out Ollama cold-start hanging
the client.
Commit 4a9c3565 added a reference to `self.config` in
`_check_compression_model_feasibility()` to pass the user-configured
`auxiliary.compression.context_length` to `get_model_context_length()`.
However, `AIAgent` never stores the loaded config dict as an instance
attribute — the config is loaded into a local variable `_agent_cfg` in
`__init__()` and discarded after init.
This causes an `AttributeError: 'AIAgent' object has no attribute
'config'` on every session start when compression is enabled, caught by
the try/except and logged as a non-fatal DEBUG message.
Fix: store the loaded config as `self._config` in `__init__()` and
update the reference in the feasibility check to use `self._config`.
Follow-up for the helix4u easy-fix salvage batch:
- route remaining context-engine quiet-mode output through
_should_emit_quiet_tool_messages() so non-CLI/library callers stay
silent consistently
- drop the extra senderAliases computation from WhatsApp allowlist-drop
logging and remove the now-unused import
This keeps the batch scoped to the intended fixes while avoiding
leaked quiet-mode output and unnecessary duplicate work in the bridge.
Inline reasoning tags in an assistant message's content field leak to every downstream consumer: messaging platforms (#8878, #9568), API replay of prior turns, session transcript, CLI recap, generated session titles, and context compression. _extract_reasoning() already captures the reasoning text into msg['reasoning'] separately, so the raw tags in content are redundant.
Stripping once at the storage boundary in _build_assistant_message() cleans the content for every downstream path in one place — no per-platform or per-path stripper needed. Measured impact on a real MiniMax M2.7-highspeed session (per @luoyejiaoe-source, #9306): 55% of assistant messages started with <think> blocks, 51/100 session titles were polluted, 16% content-size reduction.
3 new regression tests in TestBuildAssistantMessage: closed-pair strip with reasoning capture, no-think-tag passthrough, and unterminated-block strip.
Resolves#8878 and #9568.
Originally proposed as PR #9250.
Providers served via NIM (MiniMax M2.7, some Moonshot/DeepSeek proxies) sometimes drop the closing </think> tag, leaving raw reasoning in the assistant's content field. _strip_think_blocks()'s closed-pair regex is non-greedy so it only matches complete blocks — any orphan <think>...EOF survived the stripper and leaked to users (#8878, #9568, #10408).
Adds an unterminated-tag pass that fires when an open reasoning tag sits at a block boundary (start of text or after a newline) with no matching close. Everything from that tag to end of string is stripped. The block-boundary check mirrors gateway/stream_consumer.py's filter so models that mention <think> in prose are not over-stripped.
Also makes the closed-pair regexes consistently case-insensitive so <THINK>...</THINK> and <Thinking>...</Thinking> are handled uniformly — previously the mixed-case open tag would bypass the closed-pair pass and be caught by the unterminated-tag pass, taking trailing visible content with it.
6 new regression tests in TestStripThinkBlocks covering: unterminated <think>, unterminated <thought>, multi-line unterminated, line-start orphan with preserved prefix, prose-mention non-regression, mixed-case closed pairs.
The implementation is inspired by @luinbytes's PR #10408 report of the NIM/MiniMax symptom. This commit does not include the 💭/🧠 emoji regexes from that PR — those glyphs are Hermes CLI display decorations, not model content markers.
Anthropic migrated their developer console from console.anthropic.com
to platform.claude.com. Two user-facing display URLs were still pointing
to the old domain:
- hermes_cli/main.py — API key prompt in the Anthropic model flow
- run_agent.py — 401 troubleshooting output
The OAuth token refresh endpoint was already migrated in PR #3246
(with fallback).
Spotted by @LucidPaths in PR #3237.
(Salvage of #3758 — dropped the setup.py hunk since that section was
refactored away and no longer contains the stale URL.)
Based on #12152 by @LVT382009.
Two fixes to run_agent.py:
1. _ephemeral_max_output_tokens consumption in chat_completions path:
The error-recovery ephemeral override was only consumed in the
anthropic_messages branch of _build_api_kwargs. All chat_completions
providers (OpenRouter, NVIDIA NIM, Qwen, Alibaba, custom, etc.)
silently ignored it. Now consumed at highest priority, matching the
anthropic pattern.
2. NVIDIA NIM max_tokens default (16384):
NVIDIA NIM falls back to a very low internal default when max_tokens
is omitted, causing models like GLM-4.7 to truncate immediately
(thinking tokens exhaust the budget before the response starts).
3. Progressive length-continuation boost:
When finish_reason='length' triggers a continuation retry, the output
budget now grows progressively (2x base on retry 1, 3x on retry 2,
capped at 32768) via _ephemeral_max_output_tokens. Previously the
retry loop just re-sent the same token limit on all 3 attempts.
Based on #11984 by @maxchernin. Fixes#8259.
Some providers (MiniMax M2.7 via NVIDIA NIM) resend the full function
name in every streaming chunk instead of only the first. The old
accumulator used += which concatenated them into 'read_fileread_file'.
Changed to simple assignment (=), matching the OpenAI Node SDK, LiteLLM,
and Vercel AI SDK patterns. Function names are atomic identifiers
delivered complete — no provider splits them across chunks, so
concatenation was never correct semantics.
* feat(steer): /steer <prompt> injects a mid-run note after the next tool call
Adds a new slash command that sits between /queue (turn boundary) and
interrupt. /steer <text> stashes the message on the running agent and
the agent loop appends it to the LAST tool result's content once the
current tool batch finishes. The model sees it as part of the tool
output on its next iteration.
No interrupt is fired, no new user turn is inserted, and no prompt
cache invalidation happens beyond the normal per-turn tool-result
churn. Message-role alternation is preserved — we only modify an
existing role:"tool" message's content.
Wiring
------
- hermes_cli/commands.py: register /steer + add to ACTIVE_SESSION_BYPASS_COMMANDS.
- run_agent.py: add _pending_steer state, AIAgent.steer(), _drain_pending_steer(),
_apply_pending_steer_to_tool_results(); drain at end of both parallel and
sequential tool executors; clear on interrupt; return leftover as
result['pending_steer'] if the agent exits before another tool batch.
- cli.py: /steer handler — route to agent.steer() when running, fall back to
the regular queue otherwise; deliver result['pending_steer'] as next turn.
- gateway/run.py: running-agent intercept calls running_agent.steer(); idle-agent
path strips the prefix and forwards as a regular user message.
- tui_gateway/server.py: new session.steer JSON-RPC method.
- ui-tui: SessionSteerResponse type + local /steer slash command that calls
session.steer when ui.busy, otherwise enqueues for the next turn.
Fallbacks
---------
- Agent exits mid-steer → surfaces in run_conversation result as pending_steer
so CLI/gateway deliver it as the next user turn instead of silently dropping it.
- All tools skipped after interrupt → re-stashes pending_steer for the caller.
- No active agent → /steer reduces to sending the text as a normal message.
Tests
-----
- tests/run_agent/test_steer.py — accept/reject, concatenation, drain,
last-tool-result injection, multimodal list content, thread safety,
cleared-on-interrupt, registry membership, bypass-set membership.
- tests/gateway/test_steer_command.py — running agent, pending sentinel,
missing steer() method, rejected payload, empty payload.
- tests/gateway/test_command_bypass_active_session.py — /steer bypasses
the Level-1 base adapter guard.
- tests/test_tui_gateway_server.py — session.steer RPC paths.
72/72 targeted tests pass under scripts/run_tests.sh.
* feat(steer): register /steer in Discord's native slash tree
Discord's app_commands tree is a curated subset of slash commands (not
derived from COMMAND_REGISTRY like Telegram/Slack). /steer already
works there as plain text (routes through handle_message → base
adapter bypass → runner), but registering it here adds Discord's
native autocomplete + argument hint UI so users can discover and
type it like any other first-class command.
When streaming died after text was already delivered to the user but
before a tool-call's arguments finished streaming, the partial-stream
stub at the end of _interruptible_streaming_api_call silently set
`tool_calls=None` on the returned message and kept `finish_reason=stop`.
The agent treated the turn as complete, the session exited cleanly with
code 0, and the attempted action was lost with zero user-facing signal.
Live-observed Apr 2026 with MiniMax M2.7 on a ~6-minute audit task:
agent streamed 'Let me write the audit:', started emitting a write_file
tool call, MiniMax stalled for 240s mid-arguments, the stale-stream
detector killed the connection, the stub fired, session ended, no file
written, no error shown.
Fix: the streaming accumulator now records each tool-call's name into
`result['partial_tool_names']` as soon as the name is known. When the
stub builder fires after a partial delivery and finds any recorded tool
names, it appends a human-visible warning to the stub's content — and
also fires it as a live stream delta so the user sees it immediately,
not only in the persisted transcript. The next turn's model also sees
the warning in conversation history and can retry on its own. Text-only
partial streams keep the original bare-recovery behaviour (no warning).
Validation:
| Scenario | Before | After |
|---------------------------------------------|---------------------------|---------------------------------------------|
| Stream dies mid tool-call, text already sent | Silent exit, no indication | User sees ⚠ warning naming the dropped tool |
| Text-only partial stream | Bare recovered text | Unchanged |
| tests/run_agent/test_streaming.py | 24 passed | 26 passed (2 new) |
* fix(interrupt): propagate to concurrent-tool workers + opt-in debug trace
interrupt() previously only flagged the agent's _execution_thread_id.
Tools running inside _execute_tool_calls_concurrent execute on
ThreadPoolExecutor worker threads whose tids are distinct from the
agent's, so is_interrupted() inside those tools returned False no matter
how many times the gateway called .interrupt() — hung ssh / curl / long
make-builds ran to their own timeout.
Changes:
- run_agent.py: track concurrent-tool worker tids in a per-agent set,
fan interrupt()/clear_interrupt() out to them, and handle the
register-after-interrupt race at _run_tool entry. getattr fallback
for the tracker so test stubs built via object.__new__ keep working.
- tools/environments/base.py: opt-in _wait_for_process trace (ENTER,
per-30s HEARTBEAT with interrupt+activity-cb state, INTERRUPT
DETECTED, TIMEOUT, EXIT) behind HERMES_DEBUG_INTERRUPT=1.
- tools/interrupt.py: opt-in set_interrupt() trace (caller tid, target
tid, set snapshot) behind the same env flag.
- tests: new regression test runs a polling tool on a concurrent worker
and asserts is_interrupted() flips to True within ~1s of interrupt().
Second new test guards clear_interrupt() clearing tracked worker bits.
Validation: tests/run_agent/ all 762 pass; tests/tools/ interrupt+env
subset 216 pass.
* fix(interrupt-debug): bypass quiet_mode logger filter so trace reaches agent.log
AIAgent.__init__ sets logging.getLogger('tools').setLevel(ERROR) when
quiet_mode=True (the CLI default). This would silently swallow every
INFO-level trace line from the HERMES_DEBUG_INTERRUPT=1 instrumentation
added in the parent commit — confirmed by running hermes chat -q with
the flag and finding zero trace lines in agent.log even though
_wait_for_process was clearly executing (subprocess pid existed).
Fix: when HERMES_DEBUG_INTERRUPT=1, each traced module explicitly sets
its own logger level to INFO at import time, overriding the 'tools'
parent-level filter. Scoped to the opt-in case only, so production
(quiet_mode default) logs stay quiet as designed.
Validation: hermes chat -q with HERMES_DEBUG_INTERRUPT=1 now writes
'_wait_for_process ENTER/EXIT' lines to agent.log as expected.
* fix(cli): SIGTERM/SIGHUP no longer orphans tool subprocesses
Tool subprocesses spawned by the local environment backend use
os.setsid so they run in their own process group. Before this fix,
SIGTERM/SIGHUP to the hermes CLI killed the main thread via
KeyboardInterrupt but the worker thread running _wait_for_process
never got a chance to call _kill_process — Python exited, the child
was reparented to init (PPID=1), and the subprocess ran to its
natural end (confirmed live: sleep 300 survived 4+ min after SIGTERM
to the agent until manual cleanup).
Changes:
- cli.py _signal_handler (interactive) + _signal_handler_q (-q mode):
route SIGTERM/SIGHUP through agent.interrupt() so the worker's poll
loop sees the per-thread interrupt flag and calls _kill_process
(os.killpg) on the subprocess group. HERMES_SIGTERM_GRACE (default
1.5s) gives the worker time to complete its SIGTERM+SIGKILL
escalation before KeyboardInterrupt unwinds main.
- tools/environments/base.py _wait_for_process: wrap the poll loop in
try/except (KeyboardInterrupt, SystemExit) so the cleanup fires
even on paths the signal handlers don't cover (direct sys.exit,
unhandled KI from nested code, etc.). Emits EXCEPTION_EXIT trace
line when HERMES_DEBUG_INTERRUPT=1.
- New regression test: injects KeyboardInterrupt into a running
_wait_for_process via PyThreadState_SetAsyncExc, verifies the
subprocess process group is dead within 3s of the exception and
that KeyboardInterrupt re-raises cleanly afterward.
Validation:
| Before | After |
|---------------------------------------------------------|--------------------|
| sleep 300 survives 4+ min as PPID=1 orphan after SIGTERM | dies within 2 s |
| No INTERRUPT DETECTED in trace | INTERRUPT DETECTED fires + killing process group |
| tests/tools/test_local_interrupt_cleanup | 1/1 pass |
| tests/run_agent/test_concurrent_interrupt | 4/4 pass |
Replace the hardcoded 'kimi-for-coding' string check with the helper
from auxiliary_client so there is one source of truth for the list of
models with fixed-temperature contracts. Adding a new entry to
_FIXED_TEMPERATURE_MODELS now automatically covers flush_memories too.
Byte-level reasoning models (xiaomi/mimo-v2-pro, kimi, glm) can emit lone
surrogates in reasoning output. The proactive sanitizer walked content/
name/tool_calls but not extra fields like reasoning or the nested
reasoning_details array. Surrogates in those fields survived the
proactive pass, crashed json.dumps() in the OpenAI SDK, and the recovery
block's _sanitize_messages_surrogates(messages) call also didn't check
those fields — so 'found' was False, no retry happened, and after 3
attempts the user saw:
API call failed after 3 retries. 'utf-8' codec can't encode characters
in position N-M: surrogates not allowed
Changes:
- _sanitize_messages_surrogates: walk any extra string fields (reasoning,
reasoning_content, etc.) and recurse into nested dict/list values
(reasoning_details). Mirrors _sanitize_messages_non_ascii coverage
added in PR #10537.
- _sanitize_structure_surrogates: new recursive walker, mirror of
_sanitize_structure_non_ascii but for surrogate recovery.
- UnicodeEncodeError recovery block: also sanitize api_messages,
api_kwargs, and prefill_messages (not just the canonical messages
list — the API-copy carries reasoning_content transformed from
reasoning and that's what the SDK actually serializes). Always
retry on detected surrogate errors, not only when we found
something to strip — gate on error type per PR #10537's pattern.
Tests: extended tests/cli/test_surrogate_sanitization.py with
coverage for reasoning, reasoning_content, reasoning_details (flat
and deeply nested), structure walker, and an integration case that
reproduces the exact api_messages shape that was crashing.
The 'Thinking Budget Exhausted' user-facing error message advised users to
'set model.max_tokens in config.yaml'. That config key is documented but
intentionally not wired through to the API call in CLI/gateway paths — we
omit max_tokens by default so the inference server uses its full output
budget (llama-server -1=infinity, vLLM max_model_len-prompt_len, etc.).
Users followed the suggestion, saw no change, and kept filing bugs (see
closed#4404, #10917, #6955 and PRs #5001/#6080/#6446/#6707/#7075/#8804/
#10924/#11173/#11268 — all reporting the same misdirection).
Replace the misleading suggestion with an actionable one: switch models
via /model. Lowering reasoning effort remains the primary remediation.
* fix(gateway): bound _agent_cache with LRU cap + idle TTL eviction
The per-session AIAgent cache was unbounded. Each cached AIAgent holds
LLM clients, tool schemas, memory providers, and a conversation buffer.
In a long-lived gateway serving many chats/threads, cached agents
accumulated indefinitely — entries were only evicted on /new, /model,
or session reset.
Changes:
- Cache is now an OrderedDict so we can pop least-recently-used entries.
- _enforce_agent_cache_cap() pops entries beyond _AGENT_CACHE_MAX_SIZE=64
when a new agent is inserted. LRU order is refreshed via move_to_end()
on cache hits.
- _sweep_idle_cached_agents() evicts entries whose AIAgent has been idle
longer than _AGENT_CACHE_IDLE_TTL_SECS=3600s. Runs from the existing
_session_expiry_watcher so no new background task is created.
- The expiry watcher now also pops the cache entry after calling
_cleanup_agent_resources on a flushed session — previously the agent
was shut down but its reference stayed in the cache dict.
- Evicted agents have _cleanup_agent_resources() called on a daemon
thread so the cache lock isn't held during slow teardown.
Both tuning constants live at module scope so tests can monkeypatch
them without touching class state.
Tests: 7 new cases in test_agent_cache.py covering LRU eviction,
move_to_end refresh, cleanup thread dispatch, idle TTL sweep,
defensive handling of agents without _last_activity_ts, and plain-dict
test fixture tolerance.
* tweak: bump _AGENT_CACHE_MAX_SIZE 64 -> 128
* fix(gateway): never evict mid-turn agents; live spillover tests
The prior commit could tear down an active agent if its session_key
happened to be LRU when the cap was exceeded. AIAgent.close() kills
process_registry entries for the task, tears down the terminal
sandbox, closes the OpenAI client (sets self.client = None), and
cascades .close() into any active child subagents — all fatal if
the agent is still processing a turn.
Changes:
- _enforce_agent_cache_cap and _sweep_idle_cached_agents now look at
GatewayRunner._running_agents and skip any entry whose AIAgent
instance is present (identity via id(), so MagicMock doesn't
confuse lookup in tests). _AGENT_PENDING_SENTINEL is treated
as 'not active' since no real agent exists yet.
- Eviction only considers the LRU-excess window (first size-cap
entries). If an excess slot is held by a mid-turn agent, we skip
it WITHOUT compensating by evicting a newer entry. A freshly
inserted session (zero cache history) shouldn't be punished to
protect a long-lived one that happens to be busy.
- Cache may therefore stay transiently over cap when load spikes;
a WARNING is logged so operators can see it, and the next insert
re-runs the check after some turns have finished.
New tests (TestAgentCacheActiveSafety + TestAgentCacheSpilloverLive):
- Active LRU entry is skipped; no newer entry compensated
- Mixed active/idle excess window: only idle slots go
- All-active cache: no eviction, WARNING logged, all clients intact
- _AGENT_PENDING_SENTINEL doesn't block other evictions
- Idle-TTL sweep skips active agents
- End-to-end: active agent's .client survives eviction attempt
- Live fill-to-cap with real AIAgents, then spillover
- Live: CAP=4 all active + 1 newcomer — cache grows to 5, no teardown
- Live: 8 threads racing 160 inserts into CAP=16 — settles at 16
- Live: evicted session's next turn gets a fresh agent that works
30 tests pass (13 pre-existing + 17 new). Related gateway suites
(model switch, session reset, proxy, etc.) all green.
* fix(gateway): cache eviction preserves per-task state for session resume
The prior commits called AIAgent.close() on cache-evicted agents, which
tears down process_registry entries, terminal sandbox, and browser
daemon for that task_id — permanently. Fine for session-expiry (session
ended), wrong for cache eviction (session may resume).
Real-world scenario: a user leaves a Telegram session open for 2+ hours,
idle TTL evicts the cached AIAgent, user returns and sends a message.
Conversation history is preserved via SessionStore, but their terminal
sandbox (cwd, env vars, bg shells) and browser state were destroyed.
Fix: split the two cleanup modes.
close() Full teardown — session ended. Kills bg procs,
tears down terminal sandbox + browser daemon,
closes LLM client. Used by session-expiry,
/new, /reset (unchanged).
release_clients() Soft cleanup — session may resume. Closes
LLM client only. Leaves process_registry,
terminal sandbox, browser daemon intact
for the resuming agent to inherit via
shared task_id.
Gateway cache eviction (_enforce_agent_cache_cap, _sweep_idle_cached_agents)
now dispatches _release_evicted_agent_soft on the daemon thread instead
of _cleanup_agent_resources. All session-expiry call sites of
_cleanup_agent_resources are unchanged.
Tests (TestAgentCacheIdleResume, 5 new cases):
- release_clients does NOT call process_registry.kill_all
- release_clients does NOT call cleanup_vm / cleanup_browser
- release_clients DOES close the LLM client (agent.client is None after)
- close() vs release_clients() — semantic contract pinned
- Idle-evicted session's rebuild with same session_id gets same task_id
Updated test_cap_triggers_cleanup_thread to assert the soft path fires
and the hard path does NOT.
35 tests pass in test_agent_cache.py; 67 related tests green.
Re-land of #10933, now guarded by the tests in #11266.
When a provider drops a TCP connection mid-stream, the socket can enter
CLOSE-WAIT and ''epoll_wait'' may never fire — no data or error signal
arrives, so the httpx read timeout never triggers and the agent hangs
indefinitely. The other defenses (''_force_close_tcp_sockets'', stale
stream detector) all ride on the socket layer reporting the dead
connection, which it never does without probes.
Inject ''SO_KEEPALIVE'' + ''TCP_KEEPIDLE''/''KEEPINTVL''/''KEEPCNT''
into the httpx transport. Kernel probes after 30s idle, retries every
10s, gives up after 3 → dead peer detected within ~60s instead of
hanging forever. Platform-aware: ''TCP_KEEPIDLE'' on Linux,
''TCP_KEEPALIVE'' on macOS. Silent no-op on Windows or anywhere
the socket options aren't available.
The original land (#10933) mutated ''client_kwargs'' in place when it
injected the ''httpx.Client''. Since callers pass ''self._client_kwargs''
by reference, the injected client leaked into the instance state. After
the first request, the OpenAI SDK closed its ''http_client'' — including
the injected one. The next ''_create_openai_client'' call re-read the
now-closed ''httpx.Client'' from ''self._client_kwargs'' and every
subsequent chat raised ''APIConnectionError'' with cause ''RuntimeError:
Cannot send a request, as the client has been closed'' (AlexKucera's
Discord report, 2026-04-16).
The defensive ''client_kwargs = dict(client_kwargs)'' copy already on
main (taeuk178's #10978) means this injection only lands in the
per-call local copy. Each ''_create_openai_client'' invocation gets
its OWN fresh ''httpx.Client'' whose lifetime is tied to the paired
''OpenAI'' client. When that ''OpenAI'' client is closed (rebuild,
teardown, credential rotation), its ''httpx.Client'' closes with it
and the next call constructs a fresh one — no stale closed transport
can be reused.
Full 4-test matrix all green (unit + live with real OpenRouter round
trips, HERMES_LIVE_TESTS=1):
tests/run_agent/test_create_openai_client_kwargs_isolation.py PASS
tests/run_agent/test_create_openai_client_reuse.py PASS (2)
tests/run_agent/test_sequential_chats_live.py PASS
Socket options verified on the live httpx transport:
_socket_options: [(1, 9, 1), (6, 4, 30), (6, 5, 10), (6, 6, 3)]
= (SO_KEEPALIVE=1, TCP_KEEPIDLE=30s, TCP_KEEPINTVL=10s, TCP_KEEPCNT=3)
Sequential-chat reproduction of the #10933 failure was explicitly
run against this patch — the defensive copy on main prevents the
closed transport from leaking back into ''self._client_kwargs'', so
every rebuild constructs a fresh transport.
Closes#10324
PR #4918 fixed the double-/v1 bug at fresh agent init by stripping the
trailing /v1 from OpenCode base URLs when api_mode is anthropic_messages
(so the Anthropic SDK's own /v1/messages doesn't land on /v1/v1/messages).
The same logic was missing from the /model mid-session switch path.
Repro: start a session on opencode-go with GLM-5 (or any chat_completions
model), then `/model minimax-m2.7`. switch_model() correctly sets
api_mode=anthropic_messages via opencode_model_api_mode(), but base_url
passes through as https://opencode.ai/zen/go/v1. The Anthropic SDK then
POSTs to https://opencode.ai/zen/go/v1/v1/messages, which returns the
OpenCode website 404 HTML page (title 'Not Found | opencode').
Same bug affects `/model claude-sonnet-4-6` on opencode-zen.
Verified upstream: POST /v1/messages returns clean JSON 401 with x-api-key
auth (route works), while POST /v1/v1/messages returns the exact HTML 404
users reported.
Fix mirrors runtime_provider.resolve_runtime_provider:
- hermes_cli/model_switch.py::switch_model() strips /v1 after the OpenCode
api_mode override when the resolved mode is anthropic_messages.
- run_agent.py::AIAgent.switch_model() applies the same strip as
defense-in-depth, so any direct caller can't reintroduce the double-/v1.
Tests: 9 new regression tests in tests/hermes_cli/test_model_switch_opencode_anthropic.py
covering minimax on opencode-go, claude on opencode-zen, chat_completions
(GLM/Kimi/Gemini) keeping /v1 intact, codex_responses (GPT) keeping /v1
intact, trailing-slash handling, and the agent-level defense-in-depth.
All 61 TUI-related tests green across 3 consecutive xdist runs.
tests/tui_gateway/test_protocol.py:
- rename `get_messages` → `get_messages_as_conversation` on mock DB (method
was renamed in the real backend, test was still stubbing the old name)
- update tool-message shape expectation: `{role, name, context}` matches
current `_history_to_messages` output, not the legacy `{role, text}`
tests/hermes_cli/test_tui_resume_flow.py:
- `cmd_chat` grew a first-run provider-gate that bailed to "Run: hermes
setup" before `_launch_tui` was ever reached; 3 tests stubbed
`_resolve_last_session` + `_launch_tui` but not the gate
- factored a `main_mod` fixture that stubs `_has_any_provider_configured`,
reused by all three tests
tests/test_tui_gateway_server.py:
- `test_config_set_personality_resets_history_and_returns_info` was flaky
under xdist because the real `_write_config_key` touches
`~/.hermes/config.yaml`, racing with any other worker that writes
config. Stub it in the test.
Claude Opus 4.7 introduced several breaking API changes that the current
codebase partially handled but not completely. This patch finishes the
migration per the official migration guide at
https://platform.claude.com/docs/en/about-claude/models/migration-guideFixesNousResearch/hermes-agent#11137
Breaking-change coverage:
1. Adaptive thinking + output_config.effort — 4.7 is now recognized by
_supports_adaptive_thinking() (extends previous 4.6-only gate).
2. Sampling parameter stripping — 4.7 returns 400 for any non-default
temperature / top_p / top_k. build_anthropic_kwargs drops them as a
safety net; the OpenAI-protocol auxiliary path (_build_call_kwargs)
and AnthropicCompletionsAdapter.create() both early-exit before
setting temperature for 4.7+ models. This keeps flush_memories and
structured-JSON aux paths that hardcode temperature from 400ing
when the aux model is flipped to 4.7.
3. thinking.display = "summarized" — 4.7 defaults display to "omitted",
which silently hides reasoning text from Hermes's CLI activity feed
during long tool runs. Restoring "summarized" preserves 4.6 UX.
4. Effort level mapping — xhigh now maps to xhigh (was xhigh→max, which
silently over-efforted every coding/agentic request). max is now a
distinct ceiling per Anthropic's 5-level effort model.
5. New stop_reason values — refusal and model_context_window_exceeded
were silently collapsed to "stop" (end_turn) by the adapter's
stop_reason_map. Now mapped to "content_filter" and "length"
respectively, matching upstream finish-reason handling already in
bedrock_adapter.
6. Model catalogs — claude-opus-4-7 added to the Anthropic provider
list, anthropic/claude-opus-4.7 added at top of OpenRouter fallback
catalog (recommended), claude-opus-4-7 added to model_metadata
DEFAULT_CONTEXT_LENGTHS (1M, matching 4.6 per migration guide).
7. Prefill docstrings — run_agent.AIAgent and BatchRunner now document
that Anthropic Sonnet/Opus 4.6+ reject a trailing assistant-role
prefill (400).
8. Tests — 4 new tests in test_anthropic_adapter covering display
default, xhigh preservation, max on 4.7, refusal / context-overflow
stop_reason mapping, plus the sampling-param predicate. test_model_metadata
accepts 4.7 at 1M context.
Tested on macOS 15.5 (darwin). 119 tests pass in
tests/agent/test_anthropic_adapter.py, 1320 pass in tests/agent/.
Shallow-copy client_kwargs at the top of _create_openai_client() to
prevent in-place mutation from leaking back into self._client_kwargs.
Defensive fix that locks the contract for future httpx/transport work.
Cherry-picked from #10978 by @taeuk178.
The gateway compression notifications were already removed in commit cc63b2d1
(PR #4139), but the agent-level context pressure warnings (85%/95% tiered
alerts via _emit_context_pressure) were still firing on both CLI and gateway.
Removed:
- _emit_context_pressure method and all call sites in run_conversation()
- Class-level dedup state (_context_pressure_last_warned, _CONTEXT_PRESSURE_COOLDOWN)
- Instance attribute _context_pressure_warned_at
- Pressure reset logic in _compress_context
- format_context_pressure and format_context_pressure_gateway from agent/display.py
- Orphaned ANSI constants that only served these functions
- tests/run_agent/test_context_pressure.py (all 361 lines)
Compression itself continues to run silently in the background.
Closes#3784
When a model returns an empty response after tool calls with no new
tool_calls in the follow-up turn, the code enters the "nudge" recovery
path which referenced `assistant_msg` before it was assigned. This
variable is only set in the tool-calls branch (line 10098), but the
nudge code lives in the no-tool-calls branch (line 10263+).
The fix builds a fresh assistant message dict via `_build_assistant_message()`
instead of reusing the unbound variable, consistent with the exhausted-
retries path at line 10457.
Three targeted fixes for the 'agent stuck on terminal command' report:
1. **Concurrent tool wait loop now checks interrupts** (run_agent.py)
The sequential path checked _interrupt_requested before each tool call,
but the concurrent path's wait loop just blocked with 30s timeouts.
Now polls every 5s and cancels pending futures on interrupt, giving
already-running tools 3s to notice the per-thread interrupt signal.
2. **Cancelled concurrent tools get proper interrupt messages** (run_agent.py)
When a concurrent tool is cancelled or didn't return a result due to
interrupt, the tool result message says 'skipped due to user interrupt'
instead of a generic error.
3. **Typing indicator fires before follow-up turn** (gateway/run.py)
After an interrupt is acknowledged and the pending message dequeued,
the gateway now sends a typing indicator before starting the recursive
_run_agent call. This gives the user immediate visual feedback that
the system is processing their new message (closing the perceived
'dead air' gap between the interrupt ack and the response).
Reported by @_SushantSays.
When a custom provider drops a connection mid-stream, the TCP socket
can enter CLOSE-WAIT and the httpx read timeout may never fire —
epoll_wait blocks indefinitely because no data or error signal arrives.
The agent hangs until manually killed.
The existing defenses (httpx read timeout, stale stream detector,
_force_close_tcp_sockets) are all time-based and work correctly once
triggered, but they rely on the socket layer reporting the dead
connection. Without TCP keepalives, the kernel has no reason to probe
a silent connection.
Fix: inject SO_KEEPALIVE + TCP_KEEPIDLE/KEEPINTVL/KEEPCNT into the
httpx transport via socket_options. The kernel probes idle connections
after 30s, retries every 10s, gives up after 3 failures — dead peer
detected within ~60s instead of hanging forever.
Platform-aware: uses TCP_KEEPIDLE on Linux, TCP_KEEPALIVE on macOS.
Falls back silently if socket options aren't available (Windows, etc.).
Closes#10324
Skins define waiting_faces, thinking_faces, and thinking_verbs in their
spinner config, but all 7 call sites in run_agent.py used hardcoded class
constants. Add three classmethods on KawaiiSpinner that query the active
skin first and fall back to the class constants, matching the existing
pattern used for wings/tool_prefix/tool_emojis.
Co-authored-by: nosleepcassette <nosleepcassette@users.noreply.github.com>
When a custom/Ollama provider is used and reasoning_effort is set to 'none'
(or enabled: false), inject 'think': false into the request extra_body.
Ollama does not recognise the OpenRouter-style 'reasoning' extra_body field,
so thinking-capable models (Qwen3, etc.) generate <think> blocks regardless
of the reasoning_effort setting. This produces empty-response errors that
corrupt session state.
The fix adds a provider-specific block in _build_api_kwargs() that sets
think=false in extra_body whenever self.provider == 'custom' and reasoning
is explicitly disabled.
Closes#3191
When no provider was set in config.yaml and auto-detection found no
credentials, the agent silently fell back to bare OPENROUTER_API_KEY
from the environment and sent the configured model name to OpenRouter.
This produced undefined behavior -- wrong provider, wrong model routing,
and auxiliary tasks (compression, vision) hitting the wrong endpoint.
Fix: replace the silent fallback with a hard RuntimeError telling
the user to run hermes model or hermes setup. The provider must
be explicitly configured -- env vars are for secrets, not config.
* fix: show correct env var name in provider API key error (#9506)
The error message for missing provider API keys dynamically built
the env var name as PROVIDER_API_KEY (e.g. ALIBABA_API_KEY), but
some providers use different names (alibaba uses DASHSCOPE_API_KEY).
Users following the error message set the wrong variable.
Fix: look up the actual env var from PROVIDER_REGISTRY before
building the error. Falls back to the dynamic name if the registry
lookup fails.
Closes#9506
* fix: five HERMES_HOME profile-isolation leaks (#5947)
Bug A: Thread session_title from session_db to memory provider init kwargs
so honcho can derive chat-scoped session keys instead of falling back to
cwd-based naming that merges all gateway users into one session.
Bug B: Replace 14 hardcoded ~/.hermes/skills/ paths across 10 skill files
with HERMES_HOME-aware alternatives (${HERMES_HOME:-$HOME/.hermes} in
shell, os.environ.get('HERMES_HOME', ...) in Python).
Bug C: install.sh now respects HERMES_HOME env var and adds --hermes-home
flag. Previously --dir only set INSTALL_DIR while HERMES_HOME was always
hardcoded to $HOME/.hermes.
Bug D: Remove hardcoded ~/.hermes/honcho.json fallback in resolve_config_path().
Non-default profiles no longer silently inherit the default profile's honcho
config. Falls through to ~/.honcho/config.json (global) instead.
Bug E: Guard _edit_skill, _patch_skill, _delete_skill, _write_file, and
_remove_file against writing to skills found in external_dirs. Skills
outside the local SKILLS_DIR are now read-only from the agent's perspective.
Closes#5947
When Nous returns a 429, the retry amplification chain burns up to 9
API requests per conversation turn (3 SDK retries × 3 Hermes retries),
each counting against RPH and deepening the rate limit. With multiple
concurrent sessions (cron + gateway + auxiliary), this creates a spiral
where retries keep the limit tapped indefinitely.
New module: agent/nous_rate_guard.py
- Shared file-based rate limit state (~/.hermes/rate_limits/nous.json)
- Parses reset time from x-ratelimit-reset-requests-1h, x-ratelimit-
reset-requests, retry-after headers, or error context
- Falls back to 5-minute default cooldown if no header data
- Atomic writes (tempfile + rename) for cross-process safety
- Auto-cleanup of expired state files
run_agent.py changes:
- Top-of-retry-loop guard: when another session already recorded Nous
as rate-limited, skip the API call entirely. Try fallback provider
first, then return a clear message with the reset time.
- On 429 from Nous: record rate limit state and skip further retries
(sets retry_count = max_retries to trigger fallback path)
- On success from Nous: clear the rate limit state so other sessions
know they can resume
auxiliary_client.py changes:
- _try_nous() checks rate guard before attempting Nous in the auxiliary
fallback chain. When rate-limited, returns (None, None) so the chain
skips to the next provider instead of piling more requests onto Nous.
This eliminates three sources of amplification:
1. Hermes-level retries (saves 6 of 9 calls per turn)
2. Cross-session retries (cron + gateway all skip Nous)
3. Auxiliary fallback to Nous (compression/session_search skip too)
Includes 24 tests covering the rate guard module, header parsing,
state lifecycle, and auxiliary client integration.
The error message for missing provider API keys dynamically built
the env var name as PROVIDER_API_KEY (e.g. ALIBABA_API_KEY), but
some providers use different names (alibaba uses DASHSCOPE_API_KEY).
Users following the error message set the wrong variable.
Fix: look up the actual env var from PROVIDER_REGISTRY before
building the error. Falls back to the dynamic name if the registry
lookup fails.
Closes#9506
The GPT-5 auto-upgrade logic unconditionally overrode api_mode to
codex_responses for any model starting with gpt-5, even when the
user explicitly set api_mode=chat_completions. Custom proxies that
serve GPT-5 via /chat/completions became unusable.
Fix: check api_mode is None before the override fires. If the caller
passed any explicit api_mode, it is final -- no auto-upgrade.
Closes#10473
When proxy env vars (HTTP_PROXY, HTTPS_PROXY, ALL_PROXY) contain
malformed URLs — e.g. 'http://127.0.0.1:6153export' from a broken
shell config — the OpenAI/httpx client throws a cryptic 'Invalid port'
error that doesn't identify the offending variable.
Add _validate_proxy_env_urls() and _validate_base_url() in
auxiliary_client.py, called from resolve_provider_client() and
_create_openai_client() to fail fast with a clear, actionable error
message naming the broken env var or URL.
Closes#6360
Co-authored-by: MestreY0d4-Uninter <MestreY0d4-Uninter@users.noreply.github.com>
The recovery block previously only retried (continue) when one of the
per-component sanitization checks (messages, tools, system prompt,
headers, credentials) found and stripped non-ASCII content. When the
non-ASCII lived only in api_messages' reasoning_content field (which
is built from messages['reasoning'] and not checked by the original
_sanitize_messages_non_ascii), all checks returned False and the
recovery fell through to the normal error path — burning a retry
attempt despite _force_ascii_payload being set.
Now the recovery always continues (retries) when _is_ascii_codec is
detected. The _force_ascii_payload flag guarantees the next iteration
runs _sanitize_structure_non_ascii(api_kwargs) on the full API payload,
catching any remaining non-ASCII regardless of where it lives.
Also adds test for the 'reasoning' field on canonical messages.
Fixes#6843
The ASCII-locale recovery path in run_agent.py sanitized the canonical
'messages' list but left 'api_messages' untouched. api_messages is a
separate API-copy built before the retry loop and may carry extra fields
(reasoning_content, extra_body entries) that are not present in
'messages'. This caused the retry to still raise UnicodeEncodeError even
after the 'System encoding is ASCII — stripped...' log line appeared.
Two changes:
- _sanitize_messages_non_ascii now walks all extra top-level string fields
in each message dict (any key not in {content, name, tool_calls, role})
so reasoning_content and future extras are cleaned in both 'messages'
and 'api_messages'.
- The ASCII-codec recovery block now also calls sanitize on api_messages
and api_kwargs so no non-ASCII survives into the next retry attempt.
Adds regression tests covering:
- reasoning_content with non-ASCII in api_messages
- extra_body with non-ASCII in api_kwargs
- canonical messages clean but api_messages dirty
Fixes#6843
Memory provider plugins (e.g. Mnemosyne) can register tools via two paths:
1. Plugin system (ctx.register_tool) → tool registry → get_tool_definitions()
2. Memory manager → get_all_tool_schemas() → direct append in AIAgent.__init__
Path 2 blindly appended without checking if path 1 already added the same
tool names. This created duplicate function names in the tools array sent
to the API. Most providers silently handle duplicates, but Xiaomi MiMo
(via Nous Portal) strictly rejects them with a 400 Bad Request.
Fix: build a set of existing tool names before memory manager injection
and skip any tool whose name is already present.
Confirmed via live testing against Nous Portal:
- Unique tool names → 200 OK
- Duplicate tool names → 400 'Provider returned error'
The on_memory_write bridge that notifies external memory providers
(ClawMem, retaindb, supermemory, etc.) of built-in memory writes was
only present in the concurrent tool execution path (_invoke_tool).
The sequential path (_execute_tool_calls_sequential) — which handles
all single tool calls, the common case — was missing it entirely.
This meant external memory providers silently missed every single-call
memory write, which is the vast majority of memory operations.
Fix: add the identical bridge block to the sequential path, right
after the memory_tool call returns.
Closes#10174
Multiple gaps in activity tracking could cause the gateway's inactivity
timeout to fire while the agent is actively working:
1. Streaming wait loop had no periodic heartbeat — the outer thread only
touched activity when the stale-stream detector fired (180-300s), and
for local providers (Ollama) the stale timeout was infinity, meaning
zero heartbeats. Now touches activity every 30s.
2. Concurrent tool execution never set the activity callback on worker
threads (threading.local invisible across threads) and never set
_current_tool. Workers now set the callback, and the concurrent wait
uses a polling loop with 30s heartbeats.
3. Modal backend's execute() override had its own polling loop without
any activity callback. Now matches _wait_for_process cadence (10s).
The _last_content_with_tools fallback was firing indiscriminately for ALL
content+tool turns, including mid-task narration alongside substantive
tools (terminal, search_files, etc.). This caused the agent to exit
the loop with 'I'll scan the directory...' as the final answer instead
of nudging the model to continue processing tool results.
The fix restricts the fallback to housekeeping-only turns (memory, todo,
skill_manage, session_search) where the content genuinely IS the final
answer. When substantive tools are present, the existing post-tool
nudge mechanism now fires instead, prompting the model to continue.
Affected models: xiaomi/mimo-v2-pro, GLM-5, and other weaker models
that intermittently return empty after tool results.
Reported by user Renaissance on Discord.
OV transparently handles message history across /new and /compress: old
messages stay in the same session and extraction is idempotent, so there's
no need to rebind providers to a new session_id. The only thing the
session boundary actually needs is to trigger extraction.
- MemoryProvider / MemoryManager: remove on_session_reset hook
- OpenViking: remove on_session_reset override (nothing to do)
- AIAgent: replace rotate_memory_session with commit_memory_session
(just calls on_session_end, no rebind)
- cli.py / run_agent.py: single commit_memory_session call at the
session boundary before session_id rotates
- tests: replace on_session_reset coverage with routing tests for
MemoryManager.on_session_end
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace hasattr-forked OpenViking-specific paths with a proper base-class
hook. Collapse the two agent wrappers into a single rotate_memory_session
so callers don't orchestrate commit + rebind themselves.
- MemoryProvider: add on_session_reset(new_session_id) as a default no-op
- MemoryManager: on_session_reset fans out unconditionally (no hasattr,
no builtin skip — base no-op covers it)
- OpenViking: rename reset_session -> on_session_reset; drop the explicit
POST /api/v1/sessions (OV auto-creates on first message) and the two
debug raise_for_status wrappers
- AIAgent: collapse commit_memory_session + reinitialize_memory_session
into rotate_memory_session(new_sid, messages)
- cli.py / run_agent.py: replace hasattr blocks and the split calls with
a single unconditional rotate_memory_session call; compression path
now passes the real messages list instead of []
- tests: align with on_session_reset, assert reset does NOT POST /sessions
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
The OpenViking memory provider extracts memories when its session is
committed (POST /api/v1/sessions/{id}/commit). Before this fix, the
CLI had two code paths that changed the active session_id without ever
committing the outgoing OpenViking session:
1. /new (new_session() in cli.py) — called flush_memories() to write
MEMORY.md, then immediately discarded the old session_id. The
accumulated OpenViking session was never committed, so all context
from that session was lost before extraction could run.
2. /compress and auto-compress (_compress_context() in run_agent.py) —
split the SQLite session (new session_id) but left the OpenViking
provider pointing at the old session_id with no commit, meaning all
messages synced to OpenViking were silently orphaned.
The gateway already handles session commit on /new and /reset via
shutdown_memory_provider() on the cached agent; the CLI path did not.
Fix: introduce a lightweight session-transition lifecycle alongside
the existing full shutdown path:
- OpenVikingMemoryProvider.reset_session(new_session_id): waits for
in-flight background threads, resets per-session counters, and
creates the new OV session via POST /api/v1/sessions — without
tearing down the HTTP client (avoids connection overhead on /new).
- MemoryManager.restart_session(new_session_id): calls reset_session()
on providers that implement it; falls back to initialize() for
providers that do not. Skips the builtin provider (no per-session
state).
- AIAgent.commit_memory_session(messages): wraps
memory_manager.on_session_end() without shutdown — commits OV session
for extraction but leaves the provider alive for the next session.
- AIAgent.reinitialize_memory_session(new_session_id): wraps
memory_manager.restart_session() — transitions all external providers
to the new session after session_id has been assigned.
Call sites:
- cli.py new_session(): commit BEFORE session_id changes, reinitialize
AFTER — ensuring OV extraction runs on the correct session and the
new session is immediately ready for the next turn.
- run_agent._compress_context(): same pattern, inside the
if self._session_db: block where the session_id split happens.
/compress and auto-compress are functionally identical at this layer:
both call _compress_context(), so both are fixed by the same change.
Tests added to tests/agent/test_memory_provider.py:
- TestMemoryManagerRestartSession: reset_session() routing, builtin
skip, initialize() fallback, failure tolerance, empty-manager noop.
- TestOpenVikingResetSession: session_id update, per-session state
clear, POST /api/v1/sessions call, API failure tolerance, no-client
noop.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
With store=False (our default for the Responses API), the API does not
persist response items. When reasoning items with 'id' fields were
replayed on subsequent turns, the API attempted a server-side lookup
for those IDs and returned 404:
Item with id 'rs_...' not found. Items are not persisted when store
is set to false.
The encrypted_content blob is self-contained for reasoning chain
continuity — the id field is unnecessary and triggers the failed lookup.
Fix: strip 'id' from reasoning items in both _chat_messages_to_responses_input
(message conversion) and _preflight_codex_input_items (normalization layer).
The id is still used for local deduplication but never sent to the API.
Reported by @zuogl448 on GPT-5.4.
The existing recovery block sanitized self.api_key and
self._client_kwargs['api_key'] but did not update self.client.api_key.
The OpenAI SDK stores its own copy of api_key and reads it dynamically
via the auth_headers property on every request. Without this fix, the
retry after sanitization would still send the corrupted key in the
Authorization header, causing the same UnicodeEncodeError.
The bug manifests when an API key contains Unicode lookalike characters
(e.g. ʋ U+028B instead of v) from copy-pasting out of PDFs, rich-text
editors, or web pages with decorative fonts. httpx hard-encodes all
HTTP headers as ASCII, so the non-ASCII char in the Authorization
header triggers the error.
Adds TestApiKeyClientSync with two tests verifying:
- All three key locations are synced after sanitization
- Recovery handles client=None (pre-init) without crashing
Three independent fixes:
1. Reset activity timestamp on cached agent reuse (#9051)
When the gateway reuses a cached AIAgent for a new turn, the
_last_activity_ts from the previous turn (possibly hours ago)
carried over. The inactivity timeout handler immediately saw
the agent as idle for hours and killed it.
Fix: reset _last_activity_ts, _last_activity_desc, and
_api_call_count when retrieving an agent from the cache.
2. Detect uv-managed virtual environments (#8620 sub-issue 1)
The systemd unit generator fell back to sys.executable (uv's
standalone Python) when running under 'uv run', because
sys.prefix == sys.base_prefix (uv doesn't set up traditional
venv activation). The generated ExecStart pointed to a Python
binary without site-packages, crashing the service on startup.
Fix: check VIRTUAL_ENV env var before falling back to
sys.executable. uv sets VIRTUAL_ENV even when sys.prefix
doesn't reflect the venv.
3. Nudge model to continue after empty post-tool response (#9400)
Weaker models (GLM-5, mimo-v2-pro) sometimes return empty
responses after tool calls instead of continuing to the next
step. The agent silently abandoned the remaining work with
'(empty)' or used prior-turn fallback text.
Fix: when the model returns empty after tool calls AND there's
no prior-turn content to fall back on, inject a one-time user
nudge message telling the model to process the tool results and
continue. The flag resets after each successful tool round so it
can fire again on later rounds.
Test plan: 97 gateway + CLI tests pass, 9 venv detection tests pass
Previously, non-integer context_length values (e.g. '256K') in
config.yaml were silently ignored, causing the agent to fall back
to 128K auto-detection with no user feedback. This was confusing
for users with custom LiteLLM endpoints expecting larger context.
Now prints a clear stderr warning and logs at WARNING level when
model.context_length or custom_providers[].models.<model>.context_length
cannot be parsed as an integer, telling users to use plain integers
(e.g. 256000 instead of '256K').
Reported by community user ChFarhan via Discord.
When compression fails after max attempts, the agent returns
{completed: False, partial: True} but was missing the 'failed' flag.
The gateway's agent_failed_early guard checked for 'failed' AND
'not final_response', but _run_agent_blocking always converts errors
to final_response — making the guard dead code. This caused the
oversized session to persist, creating an infinite fail loop where
every subsequent message hits the same compression failure.
Changes:
- run_agent.py: add 'failed: True' and 'compression_exhausted: True'
to all 5 compression-exhaustion return paths
- gateway/run.py (_run_agent_blocking): forward 'failed' and
'compression_exhausted' flags through to the caller
- gateway/run.py (_handle_message_with_agent): fix agent_failed_early
to check bool(failed) without the broken 'not final_response' clause;
auto-reset the session when compression is exhausted so the next
message starts fresh
- Update tests to match new guard logic and add
TestCompressionExhaustedFlag test class
Closes#9893
Three bugfixes in the agent loop:
1. Reset retry counters after context compression. Without this,
pre-compression retry counts carry over, causing the model to
hit empty-response recovery immediately after a compression-
induced context loss, wasting API calls on a now-valid context.
2. Unmute output in the final-response (no-tool-call) branch.
_mute_post_response could be left True from a prior housekeeping
turn, silently suppressing empty-response warnings and recovery
status that the user should see.
3. Stop injecting 'Calling the X tools...' into assistant message
content when falling back to prior-turn content. This mutated
conversation history with synthetic text that the model never
produced, poisoning subsequent turns.
API keys containing Unicode lookalike characters (e.g. ʋ U+028B instead
of v) cause UnicodeEncodeError when httpx encodes the Authorization
header as ASCII. This commonly happens when users copy-paste keys from
PDFs, rich-text editors, or web pages with decorative fonts.
Three layers of defense:
1. **Save-time validation** (hermes_cli/config.py):
_check_non_ascii_credential() strips non-ASCII from credential values
when saving to .env, with a clear warning explaining the issue.
2. **Load-time sanitization** (hermes_cli/env_loader.py):
_sanitize_loaded_credentials() strips non-ASCII from credential env
vars (those ending in _API_KEY, _TOKEN, _SECRET, _KEY) after dotenv
loads them, so the rest of the codebase never sees non-ASCII keys.
3. **Runtime recovery** (run_agent.py):
The UnicodeEncodeError recovery block now also sanitizes self.api_key
and self._client_kwargs['api_key'], fixing the gap where message/tool
sanitization succeeded but the API key still caused httpx to fail on
the Authorization header.
Also: hermes_logging.py RotatingFileHandler now explicitly sets
encoding='utf-8' instead of relying on locale default (defensive
hardening for ASCII-locale systems).
* feat(skills): add fitness-nutrition skill to optional-skills
Cherry-picked from PR #9177 by @haileymarshall.
Adds a fitness and nutrition skill for gym-goers and health-conscious users:
- Exercise search via wger API (690+ exercises, free, no auth)
- Nutrition lookup via USDA FoodData Central (380K+ foods, DEMO_KEY fallback)
- Offline body composition calculators (BMI, TDEE, 1RM, macros, body fat %)
- Pure stdlib Python, no pip dependencies
Changes from original PR:
- Moved from skills/ to optional-skills/health/ (correct location)
- Fixed BMR formula in FORMULAS.md (removed confusing -5+10, now just +5)
- Fixed author attribution to match PR submitter
- Marked USDA_API_KEY as optional (DEMO_KEY works without signup)
Also adds optional env var support to the skill readiness checker:
- New 'optional: true' field in required_environment_variables entries
- Optional vars are preserved in metadata but don't block skill readiness
- Optional vars skip the CLI capture prompt flow
- Skills with only optional missing vars show as 'available' not 'setup_needed'
* fix: increase CLI response text padding to 4-space tab indent
Increases horizontal padding on all response display paths:
- Rich Panel responses (main, background, /btw): padding (1,2) -> (1,4)
- Streaming text: add 4-space indent prefix to each line
- Streaming TTS: add 4-space indent prefix to sentences
Gives response text proper breathing room with a tab-width indent.
Rich Panel word wrapping automatically adjusts for the wider padding.
Requested by AriesTheCoder.
* fix: word-wrap verbose tool call args and results to terminal width
Verbose mode (tool_progress: verbose) printed tool args and results as
single unwrapped lines that could be thousands of characters long.
Adds _wrap_verbose() helper that:
- Pretty-prints JSON args with indent=2 instead of one-line dumps
- Splits text on existing newlines (preserves JSON/structured output)
- Wraps lines exceeding terminal width with 5-char continuation indent
- Uses break_long_words=True for URLs and paths without spaces
Applied to all 4 verbose print sites:
- Concurrent tool call args
- Concurrent tool results
- Sequential tool call args
- Sequential tool results
---------
Co-authored-by: haileymarshall <haileymarshall@users.noreply.github.com>
GPT-5.4 supports none/low/medium/high/xhigh but not 'minimal'.
Users may configure 'minimal' via OpenRouter conventions, which would
cause a 400 on native OpenAI. Clamp to 'low' in the codex_responses
path before sending.
Plugins can now return {"action": "block", "message": "reason"} from
their pre_tool_call hook to prevent a tool from executing. The error
message is returned to the model as a tool result so it can adjust.
Covers both execution paths: handle_function_call (model_tools.py) and
agent-level tools (run_agent.py _invoke_tool + sequential/concurrent).
Blocked tools skip all side effects (counter resets, checkpoints,
callbacks, read-loop tracker).
Adds skip_pre_tool_call_hook flag to avoid double-firing the hook when
run_agent.py already checked and then calls handle_function_call.
Salvaged from PR #5385 (gianfrancopiana) and PR #4610 (oredsecurity).
- Use isinstance() with try/except import for CopilotACPClient check
in _to_async_client instead of fragile __class__.__name__ string check
- Restore accurate comment: GPT-5.x models *require* (not 'often require')
the Responses API on OpenAI/OpenRouter; ACP is the exception, not a
softening of the requirement
- Add inline comment explaining the ACP exclusion rationale
Plugin context engines loaded via load_context_engine() were never
given context_length, causing the CLI status bar to show "ctx --"
with an empty progress bar. Call update_model() immediately after
loading the plugin engine, mirroring what switch_model() already does.
FixesNousResearch/hermes-agent#9071
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previously, long-running streamed responses could be incorrectly treated
as idle by the gateway/cron inactivity timeout even while tokens were
actively arriving. The _touch_activity() call (which feeds
get_activity_summary() polled by the external timeout) was either called
only on the first chunk (chat completions) or not at all (Anthropic,
Codex, Codex fallback).
Add _touch_activity() on every chunk/event in all four streaming paths
so the inactivity monitor knows data is still flowing.
Fixes#8760
The v11→v12 migration converts custom_providers (list) into providers
(dict), then deletes the list. But all runtime resolvers read from
custom_providers — after migration, named custom endpoints silently stop
resolving and fallback chains fail with AuthError.
Add get_compatible_custom_providers() that reads from both config schemas
(legacy custom_providers list + v12+ providers dict), normalizes entries,
deduplicates, and returns a unified list. Update ALL consumers:
- hermes_cli/runtime_provider.py: _get_named_custom_provider() + key_env
- hermes_cli/auth_commands.py: credential pool provider names
- hermes_cli/main.py: model picker + _model_flow_named_custom()
- agent/auxiliary_client.py: key_env + custom_entry model fallback
- agent/credential_pool.py: _iter_custom_providers()
- cli.py + gateway/run.py: /model switch custom_providers passthrough
- run_agent.py + gateway/run.py: per-model context_length lookup
Also: use config.pop() instead of del for safer migration, fix stale
_config_version assertions in tests, add pool mock to codex test.
Co-authored-by: 墨綠BG <s5460703@gmail.com>
Closes#8776, salvaged from PR #8814
The existing ASCII codec handler only sanitized conversation messages,
leaving tool schemas, system prompts, ephemeral prompts, prefill messages,
and HTTP headers as unhandled sources of non-ASCII content. On systems
with LANG=C or non-UTF-8 locale, Unicode symbols in tool descriptions
(e.g. arrows, em-dashes from prompt_builder) and system prompt content
would cause UnicodeEncodeError that fell through to the error path.
Changes:
- Add _sanitize_structure_non_ascii() generic recursive walker for
nested dict/list payloads
- Add _sanitize_tools_non_ascii() thin wrapper for tool schemas
- Add _force_ascii_payload flag: once ASCII locale is detected, all
subsequent API calls get proactively sanitized (prevents recurring
failures from new tool results bringing fresh Unicode each turn)
- Extend the ASCII codec error handler to sanitize: prefill_messages,
tool schemas (self.tools), system prompt, ephemeral system prompt,
and default HTTP headers
- Update stale comment that acknowledged the gap
Cherry-picked from PR #8834 (credential pool changes dropped as
separate concern).
Remove the backward-compat code paths that read compression provider/model
settings from legacy config keys and env vars, which caused silent failures
when auto-detection resolved to incompatible backends.
What changed:
- Remove compression.summary_model, summary_provider, summary_base_url from
DEFAULT_CONFIG and cli.py defaults
- Remove backward-compat block in _resolve_task_provider_model() that read
from the legacy compression section
- Remove _get_auxiliary_provider() and _get_auxiliary_env_override() helper
functions (AUXILIARY_*/CONTEXT_* env var readers)
- Remove env var fallback chain for per-task overrides
- Update hermes config show to read from auxiliary.compression
- Add config migration (v16→17) that moves non-empty legacy values to
auxiliary.compression and strips the old keys
- Update example config and openclaw migration script
- Remove/update tests for deleted code paths
Compression model/provider is now configured exclusively via:
auxiliary.compression.provider / auxiliary.compression.model
Closes#8923
Add <thought>(.*?)</thought> to inline_patterns so Gemma 4
reasoning content is captured for /reasoning display, not just
stripped from visible output.
Closes#8891
Co-authored-by: RhushabhVaghela <rhushabhvaghela@users.noreply.github.com>
Three targeted changes to close the gaps between retry layers that
caused users to experience 'No response from provider for 580s' and
'No activity for 15 minutes' despite having 5 layers of retry:
1. Remove non-streaming fallback from streaming path
Previously, when all 3 stream retries exhausted, the code fell back
to _interruptible_api_call() which had no stale detection and no
activity tracking — a black hole that could hang for up to 1800s.
Now errors propagate to the main retry loop which has richer recovery
(credential rotation, provider fallback, backoff).
For 'stream not supported' errors, sets _disable_streaming flag so
the main retry loop automatically switches to non-streaming on the
next attempt.
2. Add _touch_activity to recovery dead zones
The gateway inactivity monitor relies on _touch_activity() to know
the agent is alive, but activity was never touched during:
- Stale stream detection/kill cycles (180-300s gaps)
- Stream retry connection rebuilds
- Main retry backoff sleeps (up to 120s)
- Error recovery classification
Now all these paths touch activity every ~30s, keeping the gateway
informed during recovery cycles.
3. Add stale-call detector to non-streaming path
_interruptible_api_call() now has the same stale detection pattern
as the streaming path: kills hung connections after 300s (default,
configurable via HERMES_API_CALL_STALE_TIMEOUT), scaled for large
contexts (450s for 50K+ tokens, 600s for 100K+ tokens), disabled
for local providers.
Also touches activity every ~30s during the wait so the gateway
monitor stays informed.
Env vars:
- HERMES_API_CALL_STALE_TIMEOUT: non-streaming stale timeout (default 300s)
- HERMES_STREAM_STALE_TIMEOUT: unchanged (default 180s)
Before: worst case ~2+ hours of sequential retries with no feedback
After: worst case bounded by gateway inactivity timeout (default 1800s)
with continuous activity reporting
The post-loop grace call mechanism was broken: it injected a user
message and set _budget_grace_call=True, but could never re-enter the
while loop (already exited). Worse, the flag blocked the fallback
_handle_max_iterations from running, so final_response stayed None.
Users saw empty/no response when the agent hit max iterations.
Fix: remove the dead grace block and let _handle_max_iterations handle
it directly — it already injects a summary request and makes one extra
toolless API call.
When streaming fails after partial content delivery (e.g. OpenRouter
timeout kills connection mid-response), the stub response now carries
the accumulated streamed text instead of content=None.
Two fixes:
1. The partial-stream stub response includes recovered content from
_current_streamed_assistant_text — the text that was already
delivered to the user via stream callbacks before the connection
died.
2. The empty response recovery chain now checks for partial stream
content BEFORE falling back to _last_content_with_tools (prior
turn content) or wasting API calls on retries. This prevents:
- Showing wrong content from a prior turn
- Burning 3+ unnecessary retry API calls
- Falling through to '(empty)' when the user already saw content
The root cause: OpenRouter has a ~125s inactivity timeout. When
Anthropic's SSE stream goes silent during extended reasoning, the
proxy kills the connection. The model's text was already partially
streamed but the stub discarded it, triggering the empty recovery
chain which would show stale prior-turn content or waste retries.
OpenCode Zen was in _DOT_TO_HYPHEN_PROVIDERS, causing all dotted model
names (minimax-m2.5-free, gpt-5.4, glm-5.1) to be mangled. The fix:
Layer 1 (model_normalize.py): Remove opencode-zen from the blanket
dot-to-hyphen set. Add an explicit block that preserves dots for
non-Claude models while keeping Claude hyphenated (Zen's Claude
endpoint uses anthropic_messages mode which expects hyphens).
Layer 2 (run_agent.py _anthropic_preserve_dots): Add opencode-zen and
zai to the provider allowlist. Broaden URL check from opencode.ai/zen/go
to opencode.ai/zen/ to cover both Go and Zen endpoints. Add bigmodel.cn
for ZAI URL detection.
Also adds glm-5.1 to ZAI model lists in models.py and setup.py.
Closes#7710
Salvaged from contributions by:
- konsisumer (PR #7739, #7719)
- DomGrieco (PR #8708)
- Esashiero (PR #7296)
- sharziki (PR #7497)
- XiaoYingGee (PR #8750)
- APTX4869-maker (PR #8752)
- kagura-agent (PR #7157)
_check_compression_model_feasibility() called get_model_context_length()
without passing config_context_length, so custom endpoints that do not
support /models API queries always fell through to the 128K default,
ignoring auxiliary.compression.context_length in config.yaml.
Fix: read auxiliary.compression.context_length from config and pass it
as config_context_length (highest-priority hint) so the user-configured
value is always respected regardless of API availability.
Fixes#8499
Three fixes for the (empty) response bug affecting open reasoning models:
1. Allow retries after prefill exhaustion — models like mimo-v2-pro always
populate reasoning fields via OpenRouter, so the old 'not _has_structured'
guard on the retry path blocked retries for EVERY reasoning model after
the 2 prefill attempts. Now: 2 prefills + 3 retries = 6 total attempts
before (empty).
2. Reset prefill/retry counters on tool-call recovery — the counters
accumulated across the entire conversation, never resetting during
tool-calling turns. A model cycling empty→prefill→tools→empty burned
both prefill attempts and the third empty got zero recovery. Now
counters reset when prefill succeeds with tool calls.
3. Strip think blocks before _truly_empty check — inline <think> content
made the string non-empty, skipping both retry paths.
Reported by users on Telegram with xiaomi/mimo-v2-pro and qwen3.5 models.
Reproduced: qwen3.5-9b emits tool calls as XML in reasoning field instead
of proper function calls, causing content=None + tool_calls=None + reasoning
with embedded <tool_call> XML. Prefill recovery works but counter
accumulation caused permanent (empty) in long sessions.
teknium1
ethernet
helix4u
Stephen Schoettler
ayushere
PaTTeeL
raymaylee
ephron-ren
snav
WorldWriter
kshitijk4poor
GodsBoy
AgentArcLab
ryptotalent
Siddharth Balyan
liuhao1024
Gutslabs
Wesley Simplicio
HenkDz
Teknium
ddupont
Austin Pickett
Blake Johnson
LeonSGP43
stormhierta
wxst
BarnacleBoy
rob-maron
Justin Kausel
rxdxxxx
Leon
happy5318
Chris Danis
Ricardo-M-L
0xsir0000
LLing486
thchen
QifengKuang
molvikar
ai-ag2026
luyao618
IMHaoyan
Mind-Dragon
lsdsjy
Sanjays2402
Rob Moen
Vlad Ra
Nanako0129
Rugved Somwanshi
nfb0408
simonweng
Erosika
dontcallmejames
hermes-agent-dhabibi
Andrew Ho
Tosko4
MRHwick
FocusFlow Dev
akhater
pein892
HangGlidersRule
nerijusas
brooklyn!
codez
chen1749144759
Brian D. Evans
AJ
luyao618
bsgdigital
Andre Kurait
vlwkaos
LeonSGP43
CruxExperts
Bartok9
Prasad Subrahmanya
YueLich
l0hde
Devzo
WildCat Eng Manager
shamork
Reginaldas
maelrx
kshitijk4poor
Abner
王强
pefontana
unlinearity
Kian Meng
alt-glitch
Aslaaen
jerilynzheng
Severin Bretscher
admin28980
zrc
Matt Van Horn
Tranquil-Flow
LVT382009
jarvischer
AviArora02-commits
trevthefoolish
taeuk178
Ko
nosleepcassette
ygd58
Mibayy
JiaDe WU
MestreY0d4-Uninter
MestreY0d4-Uninter
zhiheng.liu
阿泥豆
Gianfranco Piana
yongtenglei
墨綠BG
kimsr96
ygd58