Follow-up to the END-MARKER reorder: moving the summary prefix after the
[PRIOR CONTEXT] wrapper meant _is_context_summary_content (prefix-at-start)
no longer recognized a merged-tail summary. That silently broke three
consumers — the last-real-user anchor (would pick the merged summary as a
real user turn, causing active-task loss), the carry-forward summary find,
and the auto-focus skip. _strip_summary_prefix would also carry the wrapper
+ stale tail content forward as the next summary body.
Extract the two delimiter strings into _MERGED_PRIOR_CONTEXT_HEADER /
_MERGED_SUMMARY_DELIMITER constants (writer + detector stay in sync), teach
_is_context_summary_content and _strip_summary_prefix to look past the
delimiter, and add a regression test. Standalone summaries unchanged.
When the compression summary is merged into the first tail message
(the alternation corner case where a standalone summary role would
collide with both head and tail), the old format was
SUMMARY + END_MARKER + OLD_TAIL_CONTENT — so the preserved tail content
appeared AFTER the end marker and the model could read it as a fresh
message to respond to.
Reorder so the END MARKER is always last: old tail content is wrapped in
[PRIOR CONTEXT ...][END OF PRIOR CONTEXT — COMPACTION SUMMARY BELOW]
delimiters, then the summary, then the END MARKER. _append_text_to_content
handles both string and multimodal-list content.
Salvaged from #56372 by @Gromykoss. Only the END-MARKER reorder half is
carried over. The PR's second change (a post-compaction pass that strips
user-role messages before the first summary marker on compression_count>=2)
was dropped: on 2nd+ compactions the protected head decays to system-only
(_effective_protect_first_n -> 0, #11996) so the targeted 'ghost head user'
does not occur, and where the strip does fire it deletes legitimate recent
tail user turns (data loss) and can leave consecutive assistant messages
(role-alternation violation).
The original cross-session contamination fix (#38788) only cleared
_previous_summary in on_session_end(), but on_session_reset() clears
14+ per-session variables. When a session ends (cron exit, gateway
expiry, session-id rotation) and the compressor instance is reused,
the surviving stale state causes:
- _ineffective_compression_count surviving → next session skips
compression prematurely (anti-thrashing guard misfires)
- _summary_failure_cooldown_until surviving → next session blocks
summary generation for an unrelated transient error
- _last_compress_aborted surviving → callers think compression is
still aborted
- _last_aux_model_failure_* surviving → stale error warnings shown
- _last_summary_dropped_count / _last_summary_fallback_used
surviving → misleading user warnings
- _context_probed / _context_probe_persistable surviving → stale
context-probe state
Also fix on_session_reset() which was missing _last_compress_aborted
clearing — a /new or /reset would inherit the aborted flag from the
prior conversation.
Add 6 targeted tests covering the leak vectors and a parity test
ensuring on_session_end and on_session_reset always clear the same
surface.
After the first compaction protect_first_n decays, so on a later compaction
the only protected head message can be the system prompt. Adapters like
Anthropic and Bedrock send the system prompt as a separate parameter, so the
summary becomes the first message in messages[] — and Anthropic rejects any
request whose first message is not role=user (HTTP 400). Pin the summary to
role=user when the head is system-only, and stop the collision-flip logic from
reverting it back to assistant.
Salvaged from #52167.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
compress() eagerly reset _last_summary_auth_failure and
_last_summary_network_failure at the top of every call. On a second
compress() during the failure cooldown, _generate_summary() returns None from
the cooldown early-return WITHOUT re-asserting those flags, so the abort guard
saw False and fell through to the destructive static-fallback that drops the
middle window — the data-loss #29559/#25585 describe. Stop resetting them
eagerly; a successful summary already clears both, so letting them persist
across calls is safe and keeps the cooldown abort protection intact.
Salvaged from #52056.
Co-authored-by: srojk34 <286497132+srojk34@users.noreply.github.com>
_sanitize_tool_pairs inserted stub role="tool" results for orphaned
tool_calls. The pre-API repair_message_sequence() tracks known call IDs by
tc.get("id") while this sanitizer keys on call_id||id; when they disagree
(Codex Responses API: id != call_id) the stubs are silently dropped by the
repair pass, re-exposing the original orphans. Strip the orphaned tool_calls
at the source instead (preserving any text content, adding a placeholder for
an otherwise-empty assistant turn) to avoid the mismatch class entirely.
Salvaged from #51225.
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
A context-compaction handoff banner is inserted with role="user" when the
protected head ends in an assistant/tool message. On a resumed or
multi-compaction session, _find_last_user_message_idx would return that
banner as the latest user turn, so _ensure_last_user_message_in_tail anchored
the tail to the summary and rolled the genuine last user message into the
next compaction — the exact active-task loss the anchor exists to prevent
(#10896/#22523).
Reuse the existing _is_context_summary_content helper to skip summary banners
when locating the last real user message.
Salvaged from #36626 by Frank Song (issue #36624). The PR's other two changes
(demoting completed tool results inside the protected tail; a preflight
compression_exhausted result) are superseded on current main by the min_tail
floor (#39170), the no-op compression counting (#40803), and the existing
413/disabled terminal-error paths.
When the last user message sits exactly at head_end (the first compressible
index), _ensure_last_user_message_in_tail's final max(last_user_idx,
head_end + 1) clamp returns head_end + 1, pushing the user into the compressed
region without its assistant reply. The summariser then records it as a
pending ask, and the next session re-executes the already-completed task
(lights off twice, file deleted twice, message re-sent).
Fix: apply Causal Coupling — a compaction boundary must never split a
(user -> assistant [-> tool results]) turn-pair. Add _find_turn_pair_end and,
when the clamp would orphan the user, push the cut forward to pair_end so the
completed pair is summarised together and marked done.
8 new tests in TestTurnPairPreservation; 133 compressor tests pass.
When the summary LLM hits a 429/transient failure, _generate_summary() sets
a cooldown and returns None; compress() inserts a static fallback marker and
returns. Tokens stay above threshold, so should_compress() kept returning
True and every subsequent agent turn re-fired _compress_context() — the CLI
appeared frozen until the cooldown expired.
Add a cooldown guard to should_compress(): return False while
_summary_failure_cooldown_until is in the future. Reuses the existing float;
no new state. Manual /compress (force=True) still clears the cooldown first.
Fixes#11529
Follow-up hardening on the salvaged #54465 backoff persistence work.
The lease refresher's loop treated ANY falsy refresh as a permanent stop
(`if not refreshed: break`), conflating two distinct cases:
- genuine lost-ownership (rowcount 0) — correct to stop, and
- a one-off transient DB error (write contention that escapes
_execute_write's retry budget) — which returned False identically.
A single transient blip therefore killed the lease for the rest of a
multi-minute compression call, silently reintroducing the exact 300s-TTL <
~361s-call expiry wedge the PR set out to fix.
Changes:
- _CompressionLockLeaseRefresher._run now tolerates a bounded run of
consecutive failures (_MAX_CONSECUTIVE_REFRESH_FAILURES = 3) before giving
up the lease; a recovered tick resets the counter. Worst-case extra hold is
cap * refresh_interval, still bounded by the acquirer's TTL.
- Replace the two remaining silent `except Exception: pass` arms in the
compression-failure-cooldown persist/clear helpers with debug logging, for
parity with their sqlite3.Error sibling arms (a non-sqlite bug was invisible).
- Document the join(timeout=1.0) quiesce bound in stop().
- Add 3 regression tests: single-blip tolerance, persistent-failure stop at the
cap, and refresh-raising tolerance.
Add compression.minimum_context_floor config key that allows users
to lower the compression threshold floor below the hardcoded 64K
default, preventing infinite tool-call loops on models whose
structured output degrades well before 64K tokens.
- agent/model_metadata.py: add get_configurable_minimum_context()
helper with 16K hard safety limit
- agent/context_compressor.py: accept minimum_context_floor param,
thread it through _compute_threshold_tokens
- agent/conversation_compression.py: use compressor's floor for
aux model context validation
- agent/agent_init.py: read compression.minimum_context_floor from
config and pass to ContextCompressor
- gateway/run.py: cache-busting includes new key
Salvaged from #31686 by @Tranquil-Flow onto current main.
Resolves conflicts with in-place compaction (#38763) and max_tokens
threshold computation (#43547) that landed after the original PR.
Closes#31600
When context compaction's summary generation fails, the compressor's default
path (abort_on_summary_failure=False) drops the middle window and inserts a
static 'summary unavailable' marker — destroying the compacted turns. #29559
reported the field impact: a Connection error at the compaction moment dropped
124->15 messages (110 lost) for a long browser-automation task; #25585 is the
same failure mode (failed summary commits a destructive compaction anyway).
compress() already has an EXCEPTION to the historical drop default: auth
failures (401/403) ALWAYS abort and preserve the session, because rotating into
a placeholder-summary child on a broken credential strands the user. A transient
network/connection error is the same situation in reverse: it WILL recover, and
retrying then is strictly better than discarding context for a momentary blip.
Extend the always-abort carve-out to terminal connection/network failures:
- new _last_summary_network_failure flag, set in _generate_summary's terminal
failure branch when _is_connection_error(e) (reached only after any main-model
fallback is exhausted), reset alongside the auth flag;
- compress() aborts when it's set (returns messages unchanged,
_last_compress_aborted=True), independent of abort_on_summary_failure;
- a network-specific operator warning (distinct from the auth + config-flag
messages).
Scoped to connection errors only: a generic 500/400 still takes the historical
fallback-drop path (test_non_auth_failure_still_uses_fallback_path stays green).
Tests: network-failure detection + abort-despite-flag-false, both mutation-checked
(removing the flag-set fails detection; removing the carve-out fails the abort).
The compaction trigger compared estimated input against context_length *
threshold, but the provider reserves max_tokens of OUTPUT out of the same
window. With a large max_tokens (e.g. 65536 on a custom provider) the usable
input budget is materially smaller than the raw window, so sessions hit a
provider 400 before compaction ever fired.
_compute_threshold_tokens now subtracts the output reservation
(context_length - max_tokens) before applying the percentage and the
small-window 85% guard. max_tokens is stored on the compressor (threaded from
agent.max_tokens at construction) and reused across update_model() switches;
None = provider default = no reservation (full-window behavior, unchanged).
Reimplemented on the current _compute_threshold_tokens surface (the inline
threshold calc the original PR targeted was since refactored for the
small-window #14690 fix); composes with that 85% guard on the effective budget.
Credit: @kyssta-exe (#43651) — original design for the output-token
reservation in the compaction threshold.
Closes#43547.
After a compaction, the post-compression path parks last_prompt_tokens=-1 and
sets awaiting_real_usage_after_compression=True, but last_real_prompt_tokens
still holds the stale pre-compression value (above threshold). should_defer_
preflight_to_real_usage() hit the 'last_real_prompt_tokens >= threshold => False'
short-circuit and let preflight fire a SECOND compaction before the provider
reported real post-compaction usage. Add an early-return on the awaiting flag so
deferral holds for exactly one turn; update_from_response() clears it.
The flag-setting half (#36718) already landed on main via the in-place
compaction path (conversation_compression.py); this adds the missing
should_defer guard that consumes it.
Credit:
- @ashishpatel26 (#38133) — diagnosis + the should_defer early-return design
- @Tranquil-Flow (#36769) — same #36718 fix, identical guard placement
Closes#36718.
The tail-protection budget walks estimated an assistant message's tokens from content + function.arguments only, dropping each tool_call's id, type and function.name (plus JSON structure). Assistant turns that fan out into parallel tool calls were undercounted by 2-15x (a 4-tool-call turn measures ~73 vs ~1,090 real tokens), so the protected tail overshot tail_token_budget and compression ran far below its intended ratio — context kept growing.
Consolidate the three duplicated budget walks (_prune_old_tool_results and the two passes in _find_tail_cut_by_tokens) into a single _estimate_msg_budget_tokens() helper that counts the full tool_call envelope via len(str(tc)), consistent with how _estimate_message_chars estimates message size elsewhere.
Tested on Windows: new tests/agent/test_compressor_tool_call_budget.py plus the existing compression suite (test_context_compressor, compressor_image_tokens, cross_session_guard, infinite_compaction_loop) — 209 passed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
When an OpenAI-compatible proxy (e.g. cmkey.cn, one-api Anthropic channels)
returns a well-formed HTTP 200 whose summary content is null or empty/
whitespace-only, _generate_summary coerced it to "" and stored a prefix-only
summary — silently replacing the compacted turns with nothing. The model then
lost all in-progress context after compression (#11978, #11914).
_validate_llm_response already guards None / empty-choices, so those never
reach the compressor; the gap was a well-formed response with empty *content*.
Now treat empty content as a summary failure: raise so it routes through the
existing main-model fallback then transient cooldown, dropping the turns
without a summary rather than wiping context with an empty one.
Also narrow the bare 'except RuntimeError' so only genuine 'No LLM provider
configured' errors take the 600s no-provider cooldown; empty/invalid-response
RuntimeErrors from a configured provider now correctly get the main-model
fallback instead of being misrouted into the long no-provider cooldown.
Reported by @Hung2124; area identified by @annguyenNous in #39590.
The compaction threshold is max(context_length * threshold_percent,
MINIMUM_CONTEXT_LENGTH=64000). The floor prevents premature compression on
large models, but degenerates at small windows: a model at exactly 64000
ctx gets max(32000, 64000) = 64000 — a threshold equal to the ENTIRE
window. should_compress() can then never fire, because the provider
rejects the request before usage reaches 100%. Auto-compression silently
never triggers for any model whose context_length <= MINIMUM /
threshold_percent (e.g. 64K-per-slot local models).
Centralize the calc in _compute_threshold_tokens(). When the floor would
meet or exceed the context window, trigger at 85% of the window
(_MIN_CTX_TRIGGER_RATIO) — high enough that a minimum-context model uses
most of its budget before compacting (compacting at the 50% percentage
would waste half the small window), but below 100% so compaction actually
fires before the provider rejects the request. This mirrors the existing
gpt-5.5/Codex 85% autoraise rationale. Large-context behavior (floor at
64000) is unchanged; both call sites (__init__ and update_model) use the
shared helper.
Co-authored-by: soynchux <soynchuux@gmail.com>
Co-authored-by: LeonSGP43 <154585401+LeonSGP43@users.noreply.github.com>
Co-authored-by: Tranquil-Flow <tranquil_flow@protonmail.com>
ContextCompressor.update_model() recomputed context_length/threshold/budgets
but kept the cross-call calibration state (last_real_prompt_tokens,
last_rough_tokens_when_real_prompt_fit, last_compression_rough_tokens,
awaiting_real_usage_after_compression, _ineffective_compression_count) from the
PREVIOUS model.
Those fields encode 'the provider proved this prompt fit' / 'preflight can be
deferred' decisions valid only for the model that produced them. Carried across
a switch to a smaller-context model, should_defer_preflight_to_real_usage() used
the old model's 'it fit' history to SKIP a preflight compression the new model
actually needed — sending an oversized prompt the provider rejects (#23767).
update_model() now clears that state; the new model's first response repopulates
it via update_from_response(). Verified E2E: after a 200K->65,536 switch, defer
no longer suppresses and should_compress fires on an over-threshold estimate.
protect_first_n keeps the first N non-system messages verbatim through
compaction so the original task framing survives. But it was applied on
EVERY compression pass: the same early user turns were re-copied into each
child session and never summarized away, so across a long, repeatedly-
compressed session those old messages became immortal and grew the
protected head unboundedly (#11996, P1).
Decay it: protect_first_n applies on the FIRST compaction only. Once the
session has been compressed at least once (compression_count >= 1, or a
handoff summary already exists), the early turns are captured in the
summary, so _effective_protect_first_n() returns 0 and only the system
prompt stays protected. The decay is read at compress_start computation
time, before compression_count/_previous_summary are mutated at the end of
compress(), so the first pass still protects correctly.
Co-authored-by: truenorth-lj <liliangjya@gmail.com>
Co-authored-by: davidvv <david.vv@icloud.com>
When LLM summarization fails, the deterministic fallback summary rendered
the latest user ask (active_task = "User asked: '<ask>'") verbatim under
THREE headings — Historical Task Snapshot, Historical In-Progress State,
and Historical Pending User Asks. Re-presenting an already-handled ask as
unresolved in-progress/pending work made the model re-answer it AND treat
the resurrected ask as the active turn, burying the genuinely-new
post-compaction user message (#49307: answer repetition + new-instruction
loss, P1).
Keep the latest ask once, under Task Snapshot, as historical context only.
The In-Progress and Pending-Asks sections now say 'Unknown / None
recoverable from deterministic fallback' (consistent with the Active
State / Key Decisions / Resolved Questions sections) and explicitly note
the ask is historical, not outstanding. The raw turn text still appears in
the verbatim 'Last Dropped Turns' transcript — that's the dropped-turn
record, not a re-labeled instruction.
Note: the separate role=assistant standalone-summary regurgitation
(#33256) is left as-is — that role choice is constrained by strict message
alternation (user collides with a user-ending head) and is already
mitigated by the summary end-marker; forcing the role would risk the
alternation invariant.
Co-authored-by: r266-tech <r2668940489@gmail.com>
Co-authored-by: kyssta-exe <kyssta-exe@users.noreply.github.com>
Context compression is atomic, but a gateway interrupt (an incoming user
message while the agent is busy) could abort the in-flight summary call.
The Codex Responses aux stream polls the thread interrupt flag and raised
InterruptedError unconditionally — so compression fell back to a degraded
static 'summary unavailable' marker, losing the real handoff (#23975).
Add a thread-local interrupt-protection flag (aux_interrupt_protection
context manager) in auxiliary_client; the Codex stream's cancellation
check honors it. The compressor wraps its summary call_llm in the context
manager. Timeouts still fire (a hung call must die) and all other aux
tasks (vision, web_extract, title_generation, …) stay interruptible.
Re-entrant, so the main-model retry recursion is safe.
Co-authored-by: konsisumer <der@konsi.org>
When the auxiliary summary call fails with an authentication/permission
error (HTTP 401/403), context compression now ABORTS and preserves the
session unchanged instead of rotating into a child session with a
placeholder summary.
Before: a 401 (invalid/blocked key, or a token pointed at the wrong
inference host) fell through every transient-error check to 'return
None', and because compression.abort_on_summary_failure defaults False,
compress() took the static-fallback path and rotated the session anyway
(messages N->N). The user landed on a fresh-but-broken session that kept
failing the same way — paying for a full-context API call each turn with
no useful compression.
After: _generate_summary classifies 401/403 as a non-recoverable auth
failure (_last_summary_auth_failure) and compress() aborts on it
regardless of abort_on_summary_failure. A distinct auxiliary summary_model
that 401s still retries once on the main model first (its dedicated creds
may be the only broken thing); the abort only sticks when the main model
itself auth-fails or the fallback also auth-fails. The existing
_last_compress_aborted handling in conversation_compression.py already
skips rotation and emits a warning, so no session rotation occurs.
Tests: TestAuthFailureAborts — 401/403 flagging, compress() aborts despite
flag=False, non-auth failures keep the historical fallback path, and
aux-model auth failure recovers on main without aborting.
Summary messages (standalone insertion and merge-into-tail) now carry a
metadata flag so frontends (CLI, Desktop, gateway, TUI) can distinguish
them from real assistant/user messages without content-prefix heuristics.
Re-applied from PR #38434 onto current main (conflicted with the
_SUMMARY_END_MARKER hoist). Key renamed from the PR's
'is_compressed_summary' to '_compressed_summary': the wire sanitizers
strip underscore-prefixed message keys, so the flag stays in-process and
can never reach strict gateways (Fireworks/Mistral/Kimi reject unknown
keys with 'Extra inputs are not permitted').
Two-pronged fix for the WebUI "context compaction block in place of
last assistant response" regression.
Agent layer (the real fix). ``_find_tail_cut_by_tokens`` already had
``_ensure_last_user_message_in_tail`` to keep the most recent user
request out of the compressed middle (#10896), but no symmetric
anchor for the assistant side. When the conversation has an
oversized recent tool result or a long stretch of tool-call/result
pairs *after* the assistant's last visible reply, the token-budget
walk can stop with the previously-visible reply on the wrong side
of ``cut_idx``. The summariser then rolls it into the single
``[CONTEXT COMPACTION — REFERENCE ONLY]`` block persisted as
``role="user"`` or ``role="assistant"``, and from the operator's
perspective the WebUI session viewer
(``web/src/pages/SessionsPage.tsx``) and the TUI chat panel both
suddenly show the opaque "Context compaction" block in the slot
where they were just reading the actual answer:
User: "i cant see the output of the last message you sent,
i did see it previously, however now see 'context
compaction'"
Added ``_ensure_last_assistant_message_in_tail`` mirror of the
user-side anchor. It looks for the most recent assistant message
with non-empty text content (skipping tool-call-only assistant
"stubs" which the UI renders as small "calling tool X" indicators
rather than a readable bubble) and walks ``cut_idx`` back through
the standard ``_align_boundary_backward`` so we don't split a
tool_call/result group that immediately precedes it. The two
anchors are chained — each only walks ``cut_idx`` backward, so
the tail can only grow.
Falls back to "most recent assistant of any kind" only when no
content-bearing reply exists in the compressible region (fresh
multi-step tool sequence with no prior reply) — in that case the
agent-side fix is effectively a no-op and the existing
user-message anchor carries the load.
WebUI layer (clarity). Added ``isCompactionMessage`` detector that
recognises the ``[CONTEXT COMPACTION — REFERENCE ONLY]`` (current)
and ``[CONTEXT SUMMARY]:`` (legacy) prefixes from
``agent/context_compressor.py``, and a new ``compaction`` entry
in ``MessageBubble``'s ``ROLE_STYLES`` map. Compaction blocks
now render as muted, italicised system-style rows labelled
``Context handoff`` — clearly metadata, not the assistant's
actual reply — so an operator scrolling back through a long
session can't mistake the summary for a real answer.
Keeping the detected prefixes inline (rather than importing them)
because the WebUI bundle has no Python interop. A guardrail comment
points readers at the source-of-truth constants in
``agent/context_compressor.py``.
Follow-up to the #33346 cherry-pick:
- the marker string was duplicated at both insertion sites (standalone +
merged-into-tail); hoist to a module constant
- _strip_summary_prefix now also strips a trailing end marker so a
rehydrated handoff body doesn't leak the boundary directive into the
iterative-update summarizer prompt (it is re-appended on insertion)
When the compression summary lands as an assistant-role message (head ends
with user), the end marker was not appended. Models may regurgitate the
summary text as their own visible output when there's no clear boundary
signal (#33256).
The end marker was already appended for user-role summaries (#11475, #14521)
but the assistant-role path was missed in the original fix. This ensures ALL
standalone summary messages carry the boundary marker, preventing summary
text from leaking into user-visible chat output.
Legibility pass on the consolidated prefix: collapse the topic-overlap rule
from three overlapping sentences into one WINS sentence + one discard/no-wrap-up
sentence (same constraints, less dilution), fix the module docstring to
describe the headings that actually shipped, and correct the #10896 comment's
heading name (Historical Pending User Asks).
The prompt consolidation above retires the carveout-era prefix. Without a
frozen copy in _HISTORICAL_SUMMARY_PREFIXES, summaries persisted by
pre-upgrade builds would lose detection (_is_context_summary_content) and
renormalization (_strip_summary_prefix) — the exact regression class the
tuple exists to prevent. Adds contract tests covering every frozen prefix.
Refs #41607#38364#42812
ContextCompressor inherited a no-op on_session_end() from ContextEngine, so
per-session iterative-summary state (_previous_summary) survived a real session
boundary on a reused compressor instance. Override it to clear the summary the
moment the owning session ends, complementing the point-of-use guard in
compress(). Closes the cross-session contamination path in #38788.
Co-authored-by: dusterbloom <32869278+dusterbloom@users.noreply.github.com>
When a cron or background session compacts, it sets _previous_summary for
iterative updates. If that session ends without /new or /reset (which calls
on_session_reset()), the stale summary survives on the ContextCompressor
instance. A subsequent live messaging session's compaction then injects it as
'PREVIOUS SUMMARY:' into the summarizer prompt — contaminating the live
session with unrelated content from the prior session.
Add an else guard in compress(): when no handoff summary is found in the
current messages but _previous_summary is non-empty, discard it so
_generate_summary() starts fresh instead of iteratively updating a stale
cross-session summary.
Fixes#38788
When summary_target_ratio is large (e.g. 0.45) and the context_length is
moderate (e.g. 96000), the soft_ceiling (token_budget * 1.5) can exceed
the total transcript size. _find_tail_cut_by_tokens walks the entire
transcript without breaking early, and the resulting compress window is
either empty (compress_start >= compress_end) or a single message whose
summary-of-one overhead saves ~0 tokens.
Both outcomes cause a no-op compression that does not increment
_ineffective_compression_count, so should_compress() returns True on
every subsequent turn and the loop repeats endlessly.
Fix (two layers):
1. _find_tail_cut_by_tokens: when the backward walk consumed the entire
transcript without breaking (cut_idx <= head_end and accumulated <=
soft_ceiling), re-walk with the raw (non-inflated) token budget to
find a meaningful cut that gives the summarizer a useful middle window.
2. compress(): when compress_start >= compress_end, increment
_ineffective_compression_count and log a warning so the existing
anti-thrashing guard in should_compress() can break the loop.
Fixes#40803
Compaction summaries now receive the current date and instruct the
summarizer to rewrite completed actions as absolute, dated, past-tense
facts (e.g. "email John about the proposal" -> "Sent the proposal email
to John on 2026-06-07"). A resumed conversation no longer re-issues work
that already happened or treats a finished action as still pending.
The date is resolved via hermes_time.now() (date-only, user-configured
timezone) inside _generate_summary. The compaction summary is a
mid-conversation message that is never part of the cached prefix, so the
date does not affect prompt-cache stability. Date resolution is
best-effort: a clock failure omits the rule rather than blocking
compaction. The rule rides the shared template, so both first-compaction
and iterative-update prompts carry it.
Inspired by Poke's summarization (temporal anchoring + semantic
preservation).
A handoff persisted under an older SUMMARY_PREFIX can be inherited into a
resumed lineage. _strip_summary_prefix only matched the current/legacy
literal, so on re-compaction the old 'resume exactly from Active Task'
directive stayed embedded in the body and kept hijacking replies to new,
unrelated user messages.
- Add _HISTORICAL_SUMMARY_PREFIXES (pre-#35344 prefix) and strip/recognize
them in _strip_summary_prefix + _is_context_summary_content so resumed
stale handoffs are re-normalized to the current latest-message-wins prefix.
- Reconcile the overlapping Active Task template edits from the salvaged
#26290 (reverse-signal cancellation) and #32787 (capture open questions /
decisions, don't write None too eagerly) — both intents kept.
- Regression coverage in tests/agent/test_resume_stale_active_task.py.
- AUTHOR_MAP entries for both salvaged contributors.
The Active Task field in compression summaries is the single most important
field for task continuity across context boundaries. The previous template
described it narrowly as a 'task assignment' or 'request', which caused the
summary LLM to write 'None' whenever the user's most recent input was a
question, a decision request, or a discussion turn rather than an
imperative command. The assistant on the other side of the compaction then
treated the conversation as resolved and gave a generic recap instead of
answering the still-open question.
Expand the template guidance to cover:
* explicit task assignments
* questions awaiting an answer
* decisions awaiting input (A vs B)
* ongoing discussions where the assistant owes the next substantive reply
Reserve 'None' for the rare case where the last exchange was fully
resolved (e.g. user said 'thanks, that's all').
Also tighten the trailing CRITICAL instruction in the summary prompt so the
LLM cannot fall back to the old 'no imperative command → None' heuristic.
No behavioural code changes — template strings only. All 83 existing
compressor tests pass.
SUMMARY_PREFIX previously contained two contradictory directives:
1. "treat it as background reference, NOT as active instructions"
"Do NOT answer questions or fulfill requests mentioned in this summary"
"Respond ONLY to the latest user message that appears AFTER this summary"
2. "Your current task is identified in the '## Active Task' section of the
summary — resume exactly from there."
When the latest user message contradicted Active Task (e.g. 'stop the
i18n refactor', 'never mind, look at grafana instead'), models tended to
follow (2) anyway because 'resume exactly' is a strong, unambiguous
directive — leading to repeated re-surfacing of already-cancelled work
across turns, even after explicit 'stop'/'don't keep bringing that up'
messages from the user.
This change:
- Removes the conflicting 'resume exactly from Active Task' clause.
- Makes the precedence explicit: latest user message is the single source
of truth; it WINS on conflict; cancelled Active Task / In Progress /
Pending User Asks / Remaining Work must be discarded entirely (no
'wrap up the old task first').
- Names canonical reverse signals (stop, undo, roll back, never mind,
just verify, topic change) so the model recognizes them as cancellation
triggers, not background context.
- Updates the summarizer template instruction so the LLM doesn't
mechanically copy a cancelled task into Active Task on the next
compaction (it's instructed to copy the reverse signal verbatim).
- Preserves: REFERENCE ONLY framing, MEMORY.md/USER.md authority, and
the 'don't repeat work already reflected in session state' clause.
Adds tests/agent/test_summary_prefix_semantics.py to pin invariants so
the conflict can't regress.
Tested:
- All compaction tests pass: tests/agent/test_context_compressor.py,
tests/agent/test_context_compressor_summary_continuity.py,
tests/run_agent/test_413_compression.py,
tests/run_agent/test_compression_persistence.py,
tests/run_agent/test_compression_boundary_hook.py,
tests/cli/test_manual_compress.py — 117/117 passing.
- Tested on macOS.
PR #28102 made the summary-failure abort path the unconditional default,
changing established behavior. Gate it behind config.yaml flag
`compression.abort_on_summary_failure` (default False = historical
fallback-placeholder behavior).
- hermes_cli/config.py: new `compression.abort_on_summary_failure` key,
default False, documented inline.
- agent/agent_init.py: read the flag from compression config and pass to
ContextCompressor.
- agent/context_compressor.py: `__init__` accepts `abort_on_summary_failure`
(default False). `compress()` failure branch gates the abort on the
flag; when False, falls through to the restored legacy fallback path
(static "summary unavailable" placeholder + drop middle window).
- tests: restore original fallback expectations as default; add new
TestAbortOnSummaryFailure class for the opt-in mode.
Gateway/CLI plumbing (force=True on /compress, hygiene/handler abort
detection, locale `gateway.compress.aborted` key) from PR #28102 stays
intact — those paths only fire when `_last_compress_aborted` is True,
which now only happens when the flag is enabled.
When auxiliary compression's summary generation returns None (aux model
errored, returned non-JSON, timed out, etc.) the compressor previously
still dropped every middle message between compress_start..compress_end
and replaced them with a static 'Summary generation was unavailable'
placeholder. The session kept going but the user silently lost N turns
of context for nothing.
New behavior: on summary failure, compress() aborts entirely — returns
the input messages unchanged and sets _last_compress_aborted=True. The
existing _summary_failure_cooldown_until gate (30-60s) keeps the aux
model from being burned on every turn. Auto-compress callers detect
the no-op (len(after) == len(before)) and stop looping. The chat is
'frozen' at its current size until the next /compress or /new.
Manual /compress (CLI + gateway) now passes force=True which clears
the cooldown so users can retry immediately after an auto-abort. If
the manual retry also fails, the user gets a visible warning telling
them nothing was dropped and how to retry.
- agent/context_compressor.py: compress() gains force= kwarg; failure
branch sets _last_compress_aborted and returns messages unchanged
instead of inserting placeholder.
- run_agent.py: _compress_context() detects abort, surfaces warning,
skips session-rotation entirely, returns messages unchanged.
- cli.py + gateway/run.py: manual /compress paths pass force=True.
- gateway/run.py: hygiene + /compress handlers detect _last_compress_aborted
and emit the new 'Compression aborted' warning (gateway.compress.aborted)
instead of the old 'N historical messages were removed' message.
- locales/*.yaml: new gateway.compress.aborted key in all 16 locales.
- tests: updated to assert the abort contract (messages preserved,
compression_count not incremented, abort flag set, no placeholder
leaked). New test_force_true_bypasses_failure_cooldown covers the
manual-retry path.
After context compression, the protected tail messages retain their
original image parts. When those include multi-MB pasted screenshots,
every subsequent API request re-ships the same base-64 blobs forever —
which can push the request past provider body-size limits and wedge the
session even though compression 'succeeded'.
Add _strip_historical_media() to agent/context_compressor.py. After the
summary is built, find the newest user message that carries an image
part and replace image parts in every earlier message with a short
text placeholder ('[Attached image — stripped after compression]').
The newest image-bearing user turn keeps its media so the model can
still analyse what the user just sent.
Handles all three multimodal shapes:
- OpenAI chat.completions image_url
- OpenAI Responses API input_image
- Anthropic native {type: image, source: ...}
Includes 27 unit tests covering the helpers and the end-to-end
compress() integration, plus a manual E2E check confirming a ~4MB
two-image conversation shrinks to ~2MB after compression.
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.
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).
Problem:
When a provider or proxy drops a streaming response mid-flight (httpcore
raises RemoteProtocolError: "incomplete chunked read", "peer closed
connection", "response ended prematurely", etc.), _generate_summary
would not classify it as a transient error. Instead of retrying on the
main model, it entered the generic 60-second cooldown, leaving context
growing unbounded until the cooldown expired. Issue #18458.
Root cause:
_is_connection_error in auxiliary_client.py did not match httpcore's
streaming premature-close error substrings. context_compressor.py's
_generate_summary except block never called _is_connection_error, so
those errors fell through to the 60-second generic cooldown rather than
triggering the retry-on-main fallback path used for timeouts.
Fix:
1. auxiliary_client.py — extend _is_connection_error keyword list with:
"incomplete chunked read", "peer closed connection",
"response ended prematurely", "unexpected eof",
"remoteprotocolerror", "localprotocolerror".
Also guard the `from openai import ...` with try/except ImportError
so the function works in environments without the openai package.
2. context_compressor.py — import _is_connection_error and call it in
_generate_summary's except block as _is_streaming_closed. Include
_is_streaming_closed in the fallback-to-main condition (alongside
_is_model_not_found, _is_timeout, _is_json_decode) and use the
shorter 30s transient cooldown for streaming-closed errors.
Tests:
4 new regression tests in TestStreamingClosedFallback:
- test_incomplete_chunked_read_falls_back_to_main
- test_peer_closed_connection_falls_back_to_main
- test_streaming_closed_on_main_uses_short_cooldown (stash-verified)
- test_non_streaming_unknown_error_still_uses_long_cooldown
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
When an auxiliary LLM provider (or an upstream proxy) returns a non-JSON
body with `Content-Type: application/json` — e.g. an HTML 502 page from a
misconfigured gateway — the OpenAI SDK's `response.json()` raises a raw
`json.JSONDecodeError` (or wraps it in `APIResponseValidationError` whose
message contains "expecting value"). Previously this fell through to the
unknown-error branch and entered a 60s cooldown without retrying on the
main model, dropping the middle conversation turns instead.
This change folds JSON-decode detection into the existing fast-path
fallback chain: detect by `isinstance(e, JSONDecodeError)` OR substring
match for "expecting value", retry once on the main model, and use a
shorter 30s cooldown when already on main (the body shape tends to flip
back to valid quickly when the upstream proxy recovers).
The three duplicated fallback bodies (model-not-found, unknown-error,
JSON-decode) are consolidated into a single `_fallback_to_main_for_compression`
helper that handles the shared bookkeeping (record aux-model failure for
`/usage`-style callers, clear summary_model, clear cooldown).
Also adds three unit tests covering: raw `JSONDecodeError` retries on main,
substring-match for wrapped exceptions, and the 30s cooldown when already
on main.
Salvage of #22248 by @0xharryriddle. Closes#22244.
Co-authored-by: Harry Riddle <ntconguit@gmail.com>
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.
- Fix /compact → /compress in context-overflow tips (closes#20020)
- Evict cached agent after session hygiene and /compress so system
prompt refreshes with current SOUL.md, memory, and skills
- Restore memory authority across compaction: change 'informational
background data' to 'authoritative reference data' in memory block
and SUMMARY_PREFIX, with backward-compatible regex
Based on:
- PR #20027 by @LeonSGP43
- PR #18767 by @MacroAnarchy
- PR #17380 by @vominh1919
PR #17121 boundary marker fix already merged to main (2eef395e1).
PR #9262 user-message anchoring already on main via _ensure_last_user_message_in_tail().
When the head ends with assistant/tool and the tail starts with assistant,
the summary is inserted as a standalone role="user" message. The body's
verbatim "## Active Task" quote then gets read as fresh user input by
weak/local models (#11475, #14521).
The merge-into-tail path already appends an explicit end-of-summary marker
for this reason. Mirror it on the standalone path so both insertion routes
give the model the same "summary above, not new input" signal.
Commit 408dd8aa added a non-string guard for Pass 1 (dedup), but the same
pattern exists in Pass 2 (summarization/pruning) where content.startswith()
and len() are called on potentially non-string tool content.
When a provider returns tool results with non-string content (e.g. dict or
int from llama.cpp or similar), the pruning pass crashes with AttributeError.
Add the same isinstance(content, str) guard to Pass 2 for consistency.
Previously only HTTP 404/503 and specific error strings triggered a fallback
to the main model when the summary model was unavailable. Timeout errors
(HTTP 408/429/502/504, or error strings containing 'timeout') entered a
short cooldown instead, leaving context to grow unbounded for the rest of
the session.
Add _is_timeout detection alongside _is_model_not_found so that transient
timeout errors on the summary model also trigger immediate fallback to the
main model, preventing compression failure from cascading.
Closes#15935
on_session_reset() cleared _previous_summary, _last_summary_error, and
_ineffective_compression_count but left _summary_failure_cooldown_until
intact. When a transient summary error sets a 60 s cooldown (or 600 s
for a missing-provider RuntimeError) and the user immediately runs /reset
or /new, the cooldown carries into the new session. If the new session
reaches the compression threshold before the cooldown expires,
_generate_summary() returns None early, middle turns are silently dropped
without a summary, and the agent continues with no indication that
compaction was skipped.
Fix: set _summary_failure_cooldown_until = 0.0 in on_session_reset(),
matching the value assigned in __init__ and symmetric with the other
per-session fields already cleared there.
Fixes#15547
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.
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.
The existing retry-on-main path in _generate_summary only fires for errors that match the _is_model_not_found heuristic (404/503, 'model_not_found', 'does not exist', 'no available channel'). Other misconfiguration errors — 400s from aggregators, provider-specific 'no route' strings, opaque rejections — fall straight through to the transient-cooldown branch, which drops N turns of context and inserts a static placeholder.
Losing context is almost always worse than one extra summary attempt. Add a best-effort retry-on-main for the unknown-error branch, guarded by the same invariants as the existing fast-path retry: only when summary_model differs from main, and only once per compressor (_summary_model_fallen_back).
Tests cover: 404 fast-path fallback still works, unknown 400 now falls back, same-model aux skips retry (no infinite loop), and a double-failure (aux + main) stops at 2 calls.
The per-call reset block at the top of compress() cleared
_last_summary_dropped_count and _last_summary_fallback_used but
not _last_summary_error. Functionally this didn't break the
gateway warning path (callers gate on _last_summary_fallback_used
first, and _last_summary_error is overwritten on the next failure),
but it left the three tracking fields inconsistent — anyone
reading _last_summary_error standalone after a successful compress
would see a stale value from a previous failed compress.
Reset all three together so the per-call contract is uniform.
The fallback placeholder said "N conversation turns were removed" while the
gateway warning said "N historical message(s) were removed". Use "messages"
in both so users don't wonder if the two counters refer to different things.
When auxiliary compression's summary LLM call fails (e.g. model 404,
auxiliary model misconfigured), the compressor still drops the selected
turns and inserts a static fallback placeholder — the dropped context
is unrecoverable.
Previously the only signal of this was a WARNING in agent.log. Gateway
users (Telegram/Discord/etc.) had no way to know context was lost
because the existing _emit_warning path requires a status_callback,
and the gateway hygiene path uses a temporary _hyg_agent with
quiet_mode=True and no callback wired up.
Changes:
- ContextCompressor: track _last_summary_fallback_used and
_last_summary_dropped_count on each compress() call. Cleared at the
start of compress() and on session reset.
- gateway/run.py hygiene: after auto-compress, inspect the temp
agent's compressor; if fallback was used, send a visible ⚠️ warning
to the user via the platform adapter (TG/Discord/etc.) including
dropped count and the underlying error.
- gateway/run.py /compress: append the same warning to the manual
compress reply so users running /compress see the failure too.
Acceptance:
- Summary success: no user-visible warning (unchanged).
- Summary failure on gateway hygiene: user receives a TG/Discord
message with dropped count + error + remediation hint.
- Summary failure on /compress: warning appended to the command reply.
- CLI status_callback / _emit_warning path is untouched.
- Test coverage: two new tests verify the tracking fields are set on
failure and cleared on subsequent success.
* 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.
The bare-string isinstance guard added in 80ae2621 covered _find_tail_cut_by_tokens
(line 1084) but missed the identical pattern in _calculate_protect_tail_boundary
(line 487, the protect-tail scan loop). Both loops call .get("text", "") on every
list item in message["content"]; both crash with AttributeError when that list
contains a bare string.
Apply the same dict/str/fallback isinstance guard to the protect-tail path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
raw_content from message["content"] can be a list that contains bare
strings, not only dicts. The previous `p.get("text", "")` call raised
AttributeError on string items, crashing context compression for any
session that had a message with mixed content.
Guard with isinstance checks: dict → .get("text"), str → len(p),
fallback → len(str(p)). Adds a regression test covering the bare-string
case that would have AttributeError'd on the pre-fix code.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
_find_tail_cut_by_tokens called len(content) to estimate message tokens.
When content is a list of blocks (multimodal: text + image_url), len()
returns block count (e.g. 2) rather than character count, so a message
with 500 chars of text was counted as ~10 tokens instead of ~135.
This caused the backward walk to exhaust all messages before hitting the
budget ceiling; the head_end safeguard then forced cut = n - min_tail,
shrinking the protected tail to the bare minimum and preventing effective
compression of long multimodal conversations.
Fix mirrors the existing pattern in _prune_old_tool_results (line 487):
sum(len(p.get("text", "")) for p in raw_content)
if isinstance(raw_content, list) else len(raw_content)
Tests: 3 new cases in TestTokenBudgetTailProtection — regression guard
(confirms the test fails with the bug), plain-string regression guard,
and image-only block edge case.
Fixes#16087.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
update_model() recalculated threshold_tokens but left tail_token_budget
and max_summary_tokens at their __init__ values. When switching from a
200K model to 32K, the tail budget stayed at ~20K tokens (62% of 32K)
instead of the intended ~10%.
Adds budget recalculation in update_model() and 2 regression tests.
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).
_generate_summary() takes (turns_to_summarize, focus_topic) but the
summary model fallback path passed (messages, summary_budget) — where
'messages' is not even in scope, causing a NameError.
Fix the recursive call to pass the correct variables so the fallback
to the main model actually works when the summary model is unavailable.
Fixes: #10721
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.
Three-layer defense against secrets leaking into compaction summaries:
1. Input redaction: redact_sensitive_text() on message content and tool
call arguments in _serialize_for_summary() before sending to summarizer
2. Prompt instructions: NEVER include API keys/tokens/passwords in the
summarizer preamble, template Critical Context section, and focus topic
3. Output redaction: redact_sensitive_text() on the summary output and
_previous_summary for iterative updates
Reuses existing agent/redact.py patterns (sk-*, ghp_*, key=value, etc).
Cherry-picked from PR #9200 by @entropidelic.
Context compaction summaries were always produced in English regardless
of the conversation language, which injected English context into
non-English conversations and muddied the continuation experience.
Adds a one-sentence instruction to the shared `_summarizer_preamble`
used by both the initial-compaction and iterative-update prompt paths.
Placing it in the preamble (rather than adding it separately to each
prompt) means both code paths stay in sync with one edit.
Ported from anomalyco/opencode#20581. The original PR (#4670) landed
before main's prompt templates were refactored to share the
`_summarizer_preamble` and `_template_sections` blocks, so the
cherry-pick conflicted on the now-obsolete inline sections; re-applied
the essential one-line change on top of the current structure.
Verified: 48/48 existing compressor tests pass.
Pass 3 of `_prune_old_tool_results` previously shrunk long `function.arguments`
blobs by slicing the raw JSON string at byte 200 and appending the literal
text `...[truncated]`. That routinely produced payloads like::
{"path": "/foo.md", "content": "# Long markdown
...[truncated]
— an unterminated string with no closing brace. Strict providers (observed
on MiniMax) reject this as `invalid function arguments json string` with a
non-retryable 400. Because the broken call survives in the session history,
every subsequent turn re-sends the same malformed payload and gets the same
400, locking the session into a re-send loop until the call falls out of
the window.
Fix: parse the arguments first, shrink long string leaves inside the parsed
structure, and re-serialise. Non-string values (paths, ints, booleans, lists)
pass through intact. Arguments that are not valid JSON to begin with (rare,
some backends use non-JSON tool args) are returned unchanged rather than
replaced with something neither we nor the provider can parse.
Observed in the wild: a `write_file` with ~800 chars of markdown `content`
triggered this on a real session against MiniMax-M2.7; every turn after
compression got rejected until the session was manually reset.
Tests:
- 7 direct tests of `_truncate_tool_call_args_json` covering valid-JSON
output, non-JSON pass-through, nested structures, non-string leaves,
scalar JSON, and Unicode preservation
- 1 end-to-end test through `_prune_old_tool_results` Pass 3 that
reproduces the exact failure payload shape from the incident
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ensure _align_boundary_backward never pushes the last user message
into the compressed region. Without this, compression could delete
the user active task instruction mid-session.
Cherry-picked from #10969 by @sontianye. Fixes#10896.
Four 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. The generated ExecStart pointed
to a Python binary without site-packages.
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 sometimes return empty after tool calls. The agent
silently abandoned the remaining work.
Fix: append assistant('(empty)') + user nudge message and retry
once. Resets after each successful tool round.
4. Compression model fallback on permanent errors (#8620 sub-issue 4)
When the default summary model (gemini-3-flash) returns 503
'model_not_found' on custom proxies, the compressor entered a
600s cooldown, leaving context growing unbounded.
Fix: detect permanent model-not-found errors (503, 404,
'model_not_found', 'no available channel') and fall back to
using the main model for compression instead of entering
cooldown. One-time fallback with immediate retry.
Test plan: 40 compressor tests + 97 gateway/CLI tests + 9 venv tests pass
After compression, models (especially Kimi 2.5) would sometimes respond
to questions from the summary instead of the latest user message. This
happened ~30% of the time on Telegram.
Root cause: the summary's 'Next Steps' section read as active instructions,
and the SUMMARY_PREFIX didn't explicitly tell the model to ignore questions
in the summary. When the summary merged into the first tail message, there
was no clear separator between historical context and the actual user message.
Changes inspired by competitor analysis (Claude Code, OpenCode, Codex):
1. SUMMARY_PREFIX rewritten with explicit 'Do NOT answer questions from
this summary — respond ONLY to the latest user message AFTER it'
2. Summarizer preamble (shared by both prompts) adds:
- 'Do NOT respond to any questions' (from OpenCode's approach)
- 'Different assistant' framing (from Codex) to create psychological
distance between summary content and active conversation
3. New summary sections:
- '## Resolved Questions' — tracks already-answered questions with
their answers, preventing re-answering (from Claude Code's
'Pending user asks' pattern)
- '## Pending User Asks' — explicitly marks unanswered questions
- '## Remaining Work' replaces '## Next Steps' — passive framing
avoids reading as active instructions
4. merge-summary-into-tail path now inserts a clear separator:
'--- END OF CONTEXT SUMMARY — respond to the message below ---'
5. Iterative update prompt now instructs: 'Move answered questions to
Resolved Questions' to maintain the resolved/pending distinction
across multiple compactions.
Adds an optional focus topic to /compress: `/compress database schema`
guides the summariser to preserve information related to the focus topic
(60-70% of summary budget) while compressing everything else more aggressively.
Inspired by Claude Code's /compact <focus>.
Changes:
- context_compressor.py: focus_topic parameter on _generate_summary() and
compress(); appends FOCUS TOPIC guidance block to the LLM prompt
- run_agent.py: focus_topic parameter on _compress_context(), passed through
to the compressor
- cli.py: _manual_compress() extracts focus topic from command string,
preserves existing manual_compression_feedback integration (no regression)
- gateway/run.py: _handle_compress_command() extracts focus from event args
and passes through — full gateway parity
- commands.py: args_hint="[focus topic]" on /compress CommandDef
Salvaged from PR #7459 (CLI /compress focus only — /context command deferred).
15 new tests across CLI, compressor, and gateway.
Three root causes of the 'agent stops mid-task' gateway bug:
1. Compression threshold floor (64K tokens minimum)
- The 50% threshold on a 100K-context model fired at 50K tokens,
causing premature compression that made models lose track of
multi-step plans. Now threshold_tokens = max(50% * context, 64K).
- Models with <64K context are rejected at startup with a clear error.
2. Budget warning removal — grace call instead
- Removed the 70%/90% iteration budget warnings entirely. These
injected '[BUDGET WARNING: Provide your final response NOW]' into
tool results, causing models to abandon complex tasks prematurely.
- Now: no warnings during normal execution. When the budget is
actually exhausted (90/90), inject a user message asking the model
to summarise, allow one grace API call, and only then fall back
to _handle_max_iterations.
3. Activity touches during long terminal execution
- _wait_for_process polls every 0.2s but never reported activity.
The gateway's inactivity timeout (default 1800s) would fire during
long-running commands that appeared 'idle.'
- Now: thread-local activity callback fires every 10s during the
poll loop, keeping the gateway's activity tracker alive.
- Agent wires _touch_activity into the callback before each tool call.
Also: docs update noting 64K minimum context requirement.
Closes#7915 (root cause was agent-loop termination, not Weixin delivery limits).
Follow-up fixes for the context engine plugin slot (PR #5700):
- Enhance ContextEngine ABC: add threshold_percent, protect_first_n,
protect_last_n as class attributes; complete update_model() default
with threshold recalculation; clarify on_session_end() lifecycle docs
- Add ContextCompressor.update_model() override for model/provider/
base_url/api_key updates
- Replace all direct compressor internal access in run_agent.py with
ABC interface: switch_model(), fallback restore, context probing
all use update_model() now; _context_probed guarded with getattr/
hasattr for plugin engine compatibility
- Create plugins/context_engine/ directory with discovery module
(mirrors plugins/memory/ pattern) — discover_context_engines(),
load_context_engine()
- Add context.engine config key to DEFAULT_CONFIG (default: compressor)
- Config-driven engine selection in run_agent.__init__: checks config,
then plugins/context_engine/<name>/, then general plugin system,
falls back to built-in ContextCompressor
- Wire on_session_end() in shutdown_memory_provider() at real session
boundaries (CLI exit, /reset, gateway expiry)
- PluginContext.register_context_engine() lets plugins replace the
built-in ContextCompressor with a custom ContextEngine implementation
- PluginManager stores the registered engine; only one allowed
- run_agent.py checks for a plugin engine at init before falling back
to the default ContextCompressor
- reset_session_state() now calls engine.on_session_reset() instead of
poking internal attributes directly
- ContextCompressor.on_session_reset() handles its own internals
(_context_probed, _previous_summary, etc.)
- 19 new tests covering ABC contract, defaults, plugin slot registration,
rejection of duplicates/non-engines, and compressor reset behavior
- All 34 existing compressor tests pass unchanged
Introduces agent/context_engine.py — an abstract base class that defines
the pluggable context engine interface. ContextCompressor now inherits
from ContextEngine as the default implementation.
No behavior change. All 34 existing compressor tests pass.
This is the foundation for a context engine plugin slot, enabling
third-party engines like LCM (Lossless Context Management) to replace
the built-in compressor via the plugin system.
Automated dead code audit using vulture + coverage.py + ast-grep intersection,
confirmed by Opus deep verification pass. Every symbol verified to have zero
production callers (test imports excluded from reachability analysis).
Removes ~1,534 lines of dead production code across 46 files and ~1,382 lines
of stale test code. 3 entire files deleted (agent/builtin_memory_provider.py,
hermes_cli/checklist.py, tests/hermes_cli/test_setup_model_selection.py).
Co-authored-by: alt-glitch <balyan.sid@gmail.com>
When _generate_summary() failed (no provider, timeout, model error),
the compressor silently dropped all middle turns with just a debug
log. The agent would then see head + tail with no explanation of the
gap, causing total context amnesia (generic greetings instead of
continuing the conversation).
Now generates a static fallback marker that tells the model context
was lost and to continue from the recent tail messages. The fallback
flows through the same role-alternation logic as a real summary so
message structure stays valid.
Tail protection was effectively message-count based despite having a
token budget, because protect_last_n=20 acted as a hard floor. A single
50K-token tool output would cause all 20 recent messages to be
preserved regardless of budget, leaving little room for summarization.
Changes:
- _find_tail_cut_by_tokens: min_tail reduced from protect_last_n (20)
to 3; token budget is now the primary criterion
- Soft ceiling at 1.5x budget to avoid cutting mid-oversized-message
- _prune_old_tool_results: accepts optional protect_tail_tokens so
pruning also respects the token budget instead of a fixed count
- compress() minimum message check relaxed from protect_first_n +
protect_last_n + 1 to protect_first_n + 3 + 1
- Tool group alignment (no splitting tool_call/result) preserved
Three targeted improvements to the compression system:
1. Replace hardcoded truncation limits with named class constants
(_CONTENT_MAX=6000, _CONTENT_HEAD=4000, _CONTENT_TAIL=1500,
_TOOL_ARGS_MAX=1500, _TOOL_ARGS_HEAD=1200). Previous limits
(3000/500) heavily truncated the summarizer's input — a 200-line
edit got cut to 3000 chars before the summarizer ever saw it.
2. Add '## Tools & Patterns' section to both compression prompt
templates (first-pass and iterative). Preserves working tool
invocations, preferred flags, and tool-specific discoveries
across compaction boundaries.
3. Warn users on 2nd+ compression: 'Session compressed N times —
accuracy may degrade. Consider /new to start fresh.'
Ref #499
Salvage of PR #3533 (binhnt92). Follow-up to #3480 — applies min(100, ...) to 5 remaining unclamped percentage display sites in context_compressor, cli /stats, gateway /stats, and memory tool. Defensive clamps now that the root cause (estimation heuristic) was already removed in #3480.
Co-Authored-By: binhnt92 <binhnt92@users.noreply.github.com>
Add per-task timeout settings under auxiliary.{task}.timeout in config.yaml
instead of hardcoded values. Users with slow local models (Ollama, llama.cpp)
can now increase timeouts for compression, vision, session search, etc.
Defaults:
- auxiliary.compression.timeout: 120s (was hardcoded 45s)
- auxiliary.vision.timeout: 30s (unchanged)
- all other aux tasks: 30s (was hardcoded 30s)
- title_generator: 30s (was hardcoded 15s)
call_llm/async_call_llm now auto-resolve timeout from config when not
explicitly passed. Callers can still override with an explicit timeout arg.
Based on PR #3406 by alanfwilliams. Converted from env vars to config.yaml
per project conventions.
Co-authored-by: alanfwilliams <alanfwilliams@users.noreply.github.com>
- threshold: 0.80 → 0.50 (compress at 50%, not 80%)
- target_ratio: 0.40 → 0.20, now relative to threshold not total context
(20% of 50% = 10% of context as tail budget)
- summary ceiling: 32K → 12K (Gemini can't output more than ~12K)
- Updated DEFAULT_CONFIG, config display, example config, and tests
The summary_target_tokens parameter was accepted in the constructor,
stored on the instance, and never used — the summary budget was always
computed from hardcoded module constants (_SUMMARY_RATIO=0.20,
_MAX_SUMMARY_TOKENS=8000). This caused two compounding problems:
1. The config value was silently ignored, giving users no control
over post-compression size.
2. Fixed budgets (20K tail, 8K summary cap) didn't scale with
context window size. Switching from a 1M-context model to a
200K model would trigger compression that nuked 350K tokens
of conversation history down to ~30K.
Changes:
- Replace summary_target_tokens with summary_target_ratio (default 0.40)
which sets the post-compression target as a fraction of context_length.
Tail token budget and summary cap now scale proportionally:
MiniMax 200K → ~80K post-compression
GPT-5 1M → ~400K post-compression
- Change threshold_percent default: 0.50 → 0.80 (don't fire until
80% of context is consumed)
- Change protect_last_n default: 4 → 20 (preserve ~10 full turns)
- Summary token cap scales to 5% of context (was fixed 8K), capped
at 32K ceiling
- Read target_ratio and protect_last_n from config.yaml compression
section (both are now configurable)
- Remove hardcoded summary_target_tokens=500 from run_agent.py
- Add 5 new tests for ratio scaling, clamping, and new defaults
Two fixes for local model context detection:
1. Hardcoded DEFAULT_CONTEXT_LENGTHS matching was case-sensitive.
'qwen' didn't match 'Qwen3.5-9B-Q4_K_M.gguf' because of the
capital Q. Now uses model.lower() for comparison.
2. Added compressor initialization logging showing the detected
context_length, threshold, model, provider, and base_url.
This makes turn-1 compression bugs diagnosable from logs —
previously there was no log of what context length was detected.
Six improvements to reduce information loss during context compression,
informed by analysis of Cline, OpenCode, Pi-mono, Codex, and ClawdBot:
1. Structured summary template — sections for Goal, Progress (Done/
In Progress/Blocked), Key Decisions, Relevant Files, Next Steps,
and Critical Context. Forces the summarizer to preserve each
category instead of writing a vague paragraph.
2. Iterative summary updates — on re-compression, the prompt says
'PRESERVE existing info, ADD new progress, UPDATE done/in-progress
status.' Previous summary is stored and fed back to the summarizer
so accumulated context survives across multiple compactions.
3. Token-budget tail protection — instead of fixed protect_last_n=4,
walks backward keeping ~20K tokens of recent context. Adapts to
message density: sessions with big tool results protect fewer
messages, short exchanges protect more. Falls back to protect_last_n
for small conversations.
4. Tool output pruning (pre-pass) — before the expensive LLM summary,
replaces old tool result contents with a placeholder. This is free
(no LLM call) and can save 30%+ of context by itself.
5. Scaled summary budget — instead of fixed 2500 tokens, allocates 20%
of compressed content tokens (clamped to 2000-8000). A 50-turn
conversation gets more summary space than a 10-turn one.
6. Richer summarizer input — tool calls now include arguments (up to
500 chars) and tool results keep up to 3000 chars (was 1500).
The summarizer sees 'terminal(git status) → M src/config.py'
instead of just '[Tool calls: terminal]'.
Replace the fragile hardcoded context length system with a multi-source
resolution chain that correctly identifies context windows per provider.
Key changes:
- New agent/models_dev.py: Fetches and caches the models.dev registry
(3800+ models across 100+ providers with per-provider context windows).
In-memory cache (1hr TTL) + disk cache for cold starts.
- Rewritten get_model_context_length() resolution chain:
0. Config override (model.context_length)
1. Custom providers per-model context_length
2. Persistent disk cache
3. Endpoint /models (local servers)
4. Anthropic /v1/models API (max_input_tokens, API-key only)
5. OpenRouter live API (existing, unchanged)
6. Nous suffix-match via OpenRouter (dot/dash normalization)
7. models.dev registry lookup (provider-aware)
8. Thin hardcoded defaults (broad family patterns)
9. 128K fallback (was 2M)
- Provider-aware context: same model now correctly resolves to different
context windows per provider (e.g. claude-opus-4.6: 1M on Anthropic,
128K on GitHub Copilot). Provider name flows through ContextCompressor.
- DEFAULT_CONTEXT_LENGTHS shrunk from 80+ entries to ~16 broad patterns.
models.dev replaces the per-model hardcoding.
- CONTEXT_PROBE_TIERS changed from [2M, 1M, 512K, 200K, 128K, 64K, 32K]
to [128K, 64K, 32K, 16K, 8K]. Unknown models no longer start at 2M.
- hermes model: prompts for context_length when configuring custom
endpoints. Supports shorthand (32k, 128K). Saved to custom_providers
per-model config.
- custom_providers schema extended with optional models dict for
per-model context_length (backward compatible).
- Nous Portal: suffix-matches bare IDs (claude-opus-4-6) against
OpenRouter's prefixed IDs (anthropic/claude-opus-4.6) with dot/dash
normalization. Handles all 15 current Nous models.
- Anthropic direct: queries /v1/models for max_input_tokens. Only works
with regular API keys (sk-ant-api*), not OAuth tokens. Falls through
to models.dev for OAuth users.
Tests: 5574 passed (18 new tests for models_dev + updated probe tiers)
Docs: Updated configuration.md context length section, AGENTS.md
Co-authored-by: Test <test@test.com>
* fix: detect context length for custom model endpoints via fuzzy matching + config override
Custom model endpoints (non-OpenRouter, non-known-provider) were silently
falling back to 2M tokens when the model name didn't exactly match what the
endpoint's /v1/models reported. This happened because:
1. Endpoint metadata lookup used exact match only — model name mismatches
(e.g. 'qwen3.5:9b' vs 'Qwen3.5-9B-Q4_K_M.gguf') caused a miss
2. Single-model servers (common for local inference) required exact name
match even though only one model was loaded
3. No user escape hatch to manually set context length
Changes:
- Add fuzzy matching for endpoint model metadata: single-model servers
use the only available model regardless of name; multi-model servers
try substring matching in both directions
- Add model.context_length config override (highest priority) so users
can explicitly set their model's context length in config.yaml
- Log an informative message when falling back to 2M probe, telling
users about the config override option
- Thread config_context_length through ContextCompressor and AIAgent init
Tests: 6 new tests covering fuzzy match, single-model fallback, config
override (including zero/None edge cases).
* fix: auto-detect local model name and context length for local servers
Cherry-picked from PR #2043 by sudoingX.
- Auto-detect model name from local server's /v1/models when only one
model is loaded (no manual model name config needed)
- Add n_ctx_train and n_ctx to context length detection keys for llama.cpp
- Query llama.cpp /props endpoint for actual allocated context (not just
training context from GGUF metadata)
- Strip .gguf suffix from display in banner and status bar
- _auto_detect_local_model() in runtime_provider.py for CLI init
Co-authored-by: sudo <sudoingx@users.noreply.github.com>
* fix: revert accidental summary_target_tokens change + add docs for context_length config
- Revert summary_target_tokens from 2500 back to 500 (accidental change
during patching)
- Add 'Context Length Detection' section to Custom & Self-Hosted docs
explaining model.context_length config override
---------
Co-authored-by: Test <test@test.com>
Co-authored-by: sudo <sudoingx@users.noreply.github.com>
_align_boundary_backward only checked messages[idx-1] to decide if
the compress-end boundary splits a tool_call/result group. When an
assistant issues 3+ parallel tool calls, their results span multiple
consecutive messages. If the boundary fell in the middle of that group,
the parent assistant was summarized away and orphaned tool results were
silently deleted by _sanitize_tool_pairs.
Now walks backward through all consecutive tool results to find the
parent assistant, then pulls the boundary before the entire group.
6 regression tests added in tests/test_compression_boundary.py.
Co-authored-by: Guts <Gutslabs@users.noreply.github.com>
* perf: cache base_url.lower() via property, consolidate triple load_config(), hoist set constant
run_agent.py:
- Add base_url property that auto-caches _base_url_lower on every
assignment, eliminating 12+ redundant .lower() calls per API cycle
across __init__, _build_api_kwargs, _supports_reasoning_extra_body,
and the main conversation loop
- Consolidate three separate load_config() disk reads in __init__
(memory, skills, compression) into a single call, reusing the
result dict for all three config sections
model_tools.py:
- Hoist _READ_SEARCH_TOOLS set to module level (was rebuilt inside
handle_function_call on every tool invocation)
* Use endpoint metadata for custom model context and pricing
---------
Co-authored-by: kshitij <82637225+kshitijk4poor@users.noreply.github.com>
Replaces all remaining print() calls in compress() with logger.info()
and logger.warning() for consistency with the rest of the module.
Inspired by PR #1822.
compress() checks both the head and tail neighbors when choosing the
summary message role. When only the tail collides, the role is flipped.
When BOTH roles would create consecutive same-role messages (e.g.
head=assistant, tail=user), the summary is merged into the first tail
message instead of inserting a standalone message that breaks role
alternation and causes API 400 errors.
The previous code handled head-side collision but left the tail-side
uncovered — long conversations would crash mid-reply with no useful
error, forcing the user to /reset and lose session history.
Based on PR #1186 by @alireza78a, with improved double-collision
handling (merge into tail instead of unconditional 'user' fallback).
Co-authored-by: alireza78a <alireza78.crypto@gmail.com>
The summary message role was determined only by the last head message,
ignoring the first tail message. This could create consecutive user
messages (rejected by Anthropic) when the tail started with 'user'.
Now checks both neighbors. Priority: avoid colliding with the head
(already committed). If the chosen role also collides with the tail,
flip it — but only if flipping wouldn't re-collide with the head.
Adapt PR #916 onto current main by replacing the old context summary marker
with a clearer handoff wrapper, updating the summarization prompt for
resume-oriented summaries, and preserving the current call_llm-based
compression path.
* fix: Home Assistant event filtering now closed by default
Previously, when no watch_domains or watch_entities were configured,
ALL state_changed events passed through to the agent, causing users
to be flooded with notifications for every HA entity change.
Now events are dropped by default unless the user explicitly configures:
- watch_domains: list of domains to monitor (e.g. climate, light)
- watch_entities: list of specific entity IDs to monitor
- watch_all: true (new option — opt-in to receive all events)
A warning is logged at connect time if no filters are configured,
guiding users to set up their HA platform config.
All 49 gateway HA tests + 52 HA tool tests pass.
* docs: update Home Assistant integration documentation
- homeassistant.md: Fix event filtering docs to reflect closed-by-default
behavior. Add watch_all option. Replace Python dict config example with
YAML. Fix defaults table (was incorrectly showing 'all'). Add required
configuration warning admonition.
- environment-variables.md: Add HASS_TOKEN and HASS_URL to Messaging section.
- messaging/index.md: Add Home Assistant to description, architecture
diagram, platform toolsets table, and Next Steps links.
* fix(terminal): strip provider env vars from background and PTY subprocesses
Extends the env var blocklist from #1157 to also cover the two remaining
leaky paths in process_registry.py:
- spawn_local() PTY path (line 156)
- spawn_local() background Popen path (line 197)
Both were still using raw os.environ, leaking provider vars to background
processes and interactive PTY sessions. Now uses the same dynamic
_HERMES_PROVIDER_ENV_BLOCKLIST from local.py.
Explicit env_vars passed to spawn_local() still override the blocklist,
matching the existing behavior for callers that intentionally need these.
Gap identified by PR #1004 (@PeterFile).
* feat(delegate): add observability metadata to subagent results
Enrich delegate_task results with metadata from the child AIAgent:
- model: which model the child used
- exit_reason: completed | interrupted | max_iterations
- tokens.input / tokens.output: token counts
- tool_trace: per-tool-call trace with byte sizes and ok/error status
Tool trace uses tool_call_id matching to correctly pair parallel tool
calls with their results, with a fallback for messages without IDs.
Cherry-picked from PR #872 by @omerkaz, with fixes:
- Fixed parallel tool call trace pairing (was always updating last entry)
- Removed redundant 'iterations' field (identical to existing 'api_calls')
- Added test for parallel tool call trace correctness
Co-authored-by: omerkaz <omerkaz@users.noreply.github.com>
* feat(stt): add free local whisper transcription via faster-whisper
Replace OpenAI-only STT with a dual-provider system mirroring the TTS
architecture (Edge TTS free / ElevenLabs paid):
STT: faster-whisper local (free, default) / OpenAI Whisper API (paid)
Changes:
- tools/transcription_tools.py: Full rewrite with provider dispatch,
config loading, local faster-whisper backend, and OpenAI API backend.
Auto-downloads model (~150MB for 'base') on first voice message.
Singleton model instance reused across calls.
- pyproject.toml: Add faster-whisper>=1.0.0 as core dependency
- hermes_cli/config.py: Expand stt config to match TTS pattern with
provider selection and per-provider model settings
- agent/context_compressor.py: Fix .strip() crash when LLM returns
non-string content (dict from llama.cpp, None). Fixes#1100 partially.
- tests/: 23 new tests for STT providers + 2 for compressor fix
- docs/: Updated Voice & TTS page with STT provider table, model sizes,
config examples, and fallback behavior
Fallback behavior:
- Local not installed → OpenAI API (if key set)
- OpenAI key not set → local whisper (if installed)
- Neither → graceful error message to user
Co-authored-by: Jah-yee <Jah-yee@users.noreply.github.com>
---------
Co-authored-by: omerkaz <omerkaz@users.noreply.github.com>
Co-authored-by: Jah-yee <Jah-yee@users.noreply.github.com>
* fix: ClawHub skill install — use /download ZIP endpoint
The ClawHub API v1 version endpoint only returns file metadata
(path, size, sha256, contentType) without inline content or download
URLs. Our code was looking for inline content in the metadata, which
never existed, causing all ClawHub installs to fail with:
'no inline/raw file content was available'
Fix: Use the /api/v1/download endpoint (same as the official clawhub
CLI) to download skills as ZIP bundles and extract files in-memory.
Changes:
- Add _download_zip() method that downloads and extracts ZIP bundles
- Retry on 429 rate limiting with Retry-After header support
- Path sanitization and binary file filtering for security
- Keep _extract_files() as a fallback for inline/raw content
- Also fix nested file lookup (version_data.version.files)
* chore: lower default compression threshold from 85% to 50%
Triggers context compression earlier — at 50% of the model's context
window instead of 85%. Updated in all four places where the default
is defined: context_compressor.py, cli.py, run_agent.py, config.py,
and gateway/run.py.
Add centralized call_llm() and async_call_llm() functions that own the
full LLM request lifecycle:
1. Resolve provider + model from task config or explicit args
2. Get or create a cached client for that provider
3. Format request args (max_tokens handling, provider extra_body)
4. Make the API call with max_tokens/max_completion_tokens retry
5. Return the response
Config: expanded auxiliary section with provider:model slots for all
tasks (compression, vision, web_extract, session_search, skills_hub,
mcp, flush_memories). Config version bumped to 7.
Migrated all auxiliary consumers:
- context_compressor.py: uses call_llm(task='compression')
- vision_tools.py: uses async_call_llm(task='vision')
- web_tools.py: uses async_call_llm(task='web_extract')
- session_search_tool.py: uses async_call_llm(task='session_search')
- browser_tool.py: uses call_llm(task='vision'/'web_extract')
- mcp_tool.py: uses call_llm(task='mcp')
- skills_guard.py: uses call_llm(provider='openrouter')
- run_agent.py flush_memories: uses call_llm(task='flush_memories')
Tests updated for context_compressor and MCP tool. Some test mocks
still need updating (15 remaining failures from mock pattern changes,
2 pre-existing).
Three interconnected fixes for auxiliary client infrastructure:
1. CENTRALIZED PROVIDER ROUTER (auxiliary_client.py)
Add resolve_provider_client(provider, model, async_mode) — a single
entry point for creating properly configured clients. Given a provider
name and optional model, it handles auth lookup (env vars, OAuth
tokens, auth.json), base URL resolution, provider-specific headers,
and API format differences (Chat Completions vs Responses API for
Codex). All auxiliary consumers should route through this instead of
ad-hoc env var lookups.
Refactored get_text_auxiliary_client, get_async_text_auxiliary_client,
and get_vision_auxiliary_client to use the router internally.
2. FIX CODEX VISION BYPASS (vision_tools.py)
vision_tools.py was constructing a raw AsyncOpenAI client from the
sync vision client's api_key/base_url, completely bypassing the Codex
Responses API adapter. When the vision provider resolved to Codex,
the raw client would hit chatgpt.com/backend-api/codex with
chat.completions.create() which only supports the Responses API.
Fix: Added get_async_vision_auxiliary_client() which properly wraps
Codex into AsyncCodexAuxiliaryClient. vision_tools.py now uses this
instead of manual client construction.
3. FIX COMPRESSION FALLBACK + VISION ERROR HANDLING
- context_compressor.py: Removed _get_fallback_client() which blindly
looked for OPENAI_API_KEY + OPENAI_BASE_URL (fails for Codex OAuth,
API-key providers, users without OPENAI_BASE_URL set). Replaced
with fallback loop through resolve_provider_client() for each
known provider, with same-provider dedup.
- vision_tools.py: Added error detection for vision capability
failures. Returns clear message to the model when the configured
model doesn't support vision, instead of a generic error.
Addresses #886
The summary message was always injected as 'user' role, which causes
consecutive user messages when the last preserved head message is also
'user'. Some APIs reject this (400 error), and it produces malformed
training data.
Fix: check the role of the last head message and pick the opposite role
for the summary — 'user' after assistant/tool, 'assistant' after user.
Based on PR #328 by johnh4098. Closes#328.
- Added support for auxiliary model overrides in the configuration, allowing users to specify providers and models for vision and web extraction tasks.
- Updated the CLI configuration example to include new auxiliary model settings.
- Enhanced the environment variable mapping in the CLI to accommodate auxiliary model configurations.
- Improved the resolution logic for auxiliary clients to support task-specific provider overrides.
- Updated relevant documentation and comments for clarity on the new features and their usage.
Updated the _generate_summary method to attempt summary generation using the auxiliary model first, with a fallback to the main model. If both attempts fail, the method now returns None instead of a placeholder, allowing the caller to handle missing summaries appropriately. This change enhances the robustness of context compression and improves logging for failure scenarios.
Enhance message compression by adding a method to clean up orphaned tool-call and tool-result pairs. This ensures that the API receives well-formed messages, preventing errors related to mismatched IDs. The new functionality includes removing orphaned results and adding stub results for missing calls, improving overall message integrity during compression.
Replaces the unsafe 128K fallback for unknown models with a descending
probe strategy (2M → 1M → 512K → 200K → 128K → 64K → 32K). When a
context-length error occurs, the agent steps down tiers and retries.
The discovered limit is cached per model+provider combo in
~/.hermes/context_length_cache.yaml so subsequent sessions skip probing.
Also parses API error messages to extract the actual context limit
(e.g. 'maximum context length is 32768 tokens') for instant resolution.
The CLI banner now displays the context window size next to the model
name (e.g. 'claude-opus-4 · 200K context · Nous Research').
Changes:
- agent/model_metadata.py: CONTEXT_PROBE_TIERS, persistent cache
(save/load/get), parse_context_limit_from_error(), get_next_probe_tier()
- agent/context_compressor.py: accepts base_url, passes to metadata
- run_agent.py: step-down logic in context error handler, caches on success
- cli.py + hermes_cli/banner.py: context length in welcome banner
- tests: 22 new tests for probing, parsing, and caching
Addresses #132. PR #319's approach (8K default) rejected — too conservative.
When the auxiliary client (used for context compression summaries) fails
— e.g. due to a stale OpenRouter API key after switching to a local LLM
— fall back to the user's active endpoint (OPENAI_BASE_URL) instead of
returning a useless static summary string.
This handles the common scenario where a user switches providers via
'hermes model' but the old provider's API key remains in .env. The
auxiliary client picks up the stale key, fails (402/auth error), and
previously compression would produce garbage. Now it gracefully retries
with the working endpoint.
On successful fallback, the working client is cached for future
compressions in the same session so the fallback cost is paid only once.
Ref: #348
The OpenAI API returns content: null on assistant messages that only
contain tool calls. msg.get('content', '') returns None (not '') when
the key exists with value None, causing TypeError on len() and string
concatenation in _generate_summary and compress.
Fix: msg.get('content') or '' — handles both missing keys and None.
Tests from PR #216 (@Farukest). Fix also in PR #215 (@cutepawss).
Both PRs had stale branches and couldn't be merged directly.
Closes#211
- Enhanced Codex model discovery by fetching available models from the API, with fallback to local cache and defaults.
- Updated the context compressor's summary target tokens to 2500 for improved performance.
- Added external credential detection for Codex CLI to streamline authentication.
- Refactored various components to ensure consistent handling of authentication and model selection across the application.
- Added _max_tokens_param method in AIAgent to return appropriate max tokens parameter based on the provider (OpenAI vs. others).
- Updated API calls in AIAgent to utilize the new max tokens handling.
- Introduced auxiliary_max_tokens_param function in auxiliary_client for consistent max tokens management across auxiliary clients.
- Refactored multiple tools to use auxiliary_max_tokens_param for improved compatibility with different models and providers.
kshitijk4poor
Gromykoss
Ben Barclay
liuhao1024
srojk34
Frank Song
H2KFORGIVEN
ygd58
Vladimir Smirnov
Rod Boev
Tranquil-Flow
Basil Al Shukaili
Teknium
kyssta-exe
konsisumer
xxxigm
dusterbloom
Basil Al Shukaili
islam666
Mathijs van den Hurk
Zhipeng Li
helix4u
hinotoi-agent
0z1-ghb
glennc
Stephen Schoettler
snav
Wesley Simplicio
Teknium
LeonSGP43
wmagev
revaraver
JasonOA888
swithek
pander
nftpoetrist
sprmn24
iamagenius00
iamagenius00
briandevans
vominh1919
sicnuyudidi
Yukipukii1
alt-glitch
entropidelic
Honghua Yang
sontianye
kshitijk4poor
Harish Kukreja
BongSuCHOI
SHL0MS
KangYu
Test