When two consecutive compactions each failed to clear the threshold, the
anti-thrashing breaker blocked automatic compaction PERMANENTLY for the
life of the session: nothing decremented _ineffective_compression_count
(or _fallback_compression_streak) while blocked, so a session whose
middle region was briefly too small to compact never auto-compacted
again — it grew unbounded until the provider's hard context limit, and
only /new or /reset recovered it.
Recovery is a probation probe, not amnesty: after
_ANTI_THRASH_RECOVERY_SECONDS (300s) of continuous block the gate grants
exactly ONE attempt by dropping tripped counters to 1 strike (persisted,
so sibling agents on the same session row — gateway hygiene — unblock
too). An ineffective probe re-trips the guard on the next real-usage
verdict and the next recovery waits a full fresh window, so the worst
case in a truly incompressible session is one compaction attempt per
window — bounded, not thrash.
The recovery clock is armed lazily on the first BLOCKED evaluation and
is deliberately not durable: a restart that loads a durable tripped
counter (#69872) starts a full fresh window blocked, preserving the
restart-must-never-disarm contract (#54923).
Fixes#14694
Three durable ledgers used `with _connect() as conn:` where the sqlite3
connection context manager commits/rolls back but never closes, leaking the
db/-wal/-shm file descriptors on every call. On a long-running gateway this
exhausts RLIMIT_NOFILE and fails unrelated components with
`[Errno 24] Too many open files`. Same bug class as the cron execution ledger
(#69567 / PR #69594), which the connection helpers here are modeled on.
Fix: route every ledger operation through a `_transaction()` context manager
that guarantees `conn.close()` on exit. `_connect()` keeps its
schema-on-connect contract (several tests call it directly) and now self-closes
if schema init fails.
Adds per-module regression tests asserting every opened connection is closed,
including the no-op-update and exception-mid-transaction paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Follow-up to the cherry-picked #68633 commits, closing the final open
review point (egilewski): _relocated_replay_cache_control was applied
only inside `if replayed:`. When anthropic_content_blocks contained
only a blank cache-marked text block, `replayed` came out empty, the
function fell through to the main path's placeholder, and the cache
marker was lost; signed thinking + a blank marked text block likewise
returned with no cacheable carrier for the relocated marker.
The replay branch now appends the non-whitespace "(empty)" placeholder
when no cacheable (text/tool_use) block survives the blank filter and a
blank text block was dropped (or a marker needs a carrier) — so replay
stays schema-valid on Bedrock/strict endpoints and the breakpoint
survives on the placeholder.
Also reconciles the block-level tests from #69517 with the new
drop-then-fallback contract (blank blocks are dropped at the block
level; the message-level result is still always non-blank).
Refs #69512
Co-authored-by: ygd58 <buraysandro9@gmail.com>
Follow-up per independent review of #68633 (GPT-5.6-sol-xhigh in Codex,
reviewer egilewski) on this PR.
Two real bugs in the blank-text-block filtering added by that fix:
1. `effective = blocks or content` fell back to the RAW, unfiltered
`content` variable whenever every block was filtered out as blank --
which happens precisely when the entire message content WAS the
blank/whitespace payload the filter exists to remove (a sole blank
text block, a sole cache-marked blank block, or standalone
whitespace scalar content with no tool_calls). The fallback silently
restored the exact invalid content the filtering just stripped,
leaving the message provider-invalid.
Fixed: `effective = blocks if blocks else [{"type": "text", "text":
"(empty)"}]` -- never falls back to raw `content`. Also moved the
cache_control application (both the relocated-from-a-dropped-block
marker and the message-level marker) to run against `effective`
instead of the pre-fallback `blocks`, so a cache marker on a block
that was the ONLY content still lands on the (empty) placeholder
rather than being silently lost when `blocks` was empty at the
point it would otherwise have been applied.
2. The normal-path blank-text check used `(blk.get("text") or "").strip()`,
which is not type-safe for a truthy NON-string, non-None text value
(e.g. an int or dict from an invalid upstream payload) -- `or`
doesn't substitute for a truthy value, so `(7 or "").strip()` still
raises AttributeError. Now checks `isinstance(text, str)` first,
matching the replay path's `_sanitize_replay_block()`, which the
reviewer confirmed was already correctly type-safe.
Added regression tests for: sole blank list block, sole whitespace
scalar content, sole cache-marked blank block (marker relocation to
the placeholder), a truthy non-string (int) text value both mixed with
a surviving tool_use and as the sole content, and a dict-valued text
field. 7/7 new tests pass; 193/193 in the full
tests/agent/test_anthropic_adapter.py file; 23/23 in
tests/agent/test_prompt_caching.py (unaffected, confirmed).
Ports #63228 forward onto current main per teknium1's review.
Bedrock and strict Anthropic-compatible endpoints reject text blocks
where text is empty or whitespace-only with HTTP 400. The normal
list-content path extended blocks without filtering, and the
ordered-replay fast path (_sanitize_replay_block) returned blank text
blocks unfiltered.
Per review, fixes three gaps in the original port:
1. Type safety: the normal-path filter used blk.get('text', '').strip(),
which crashes with AttributeError when text is explicitly None (not
absent) -- .get()'s default only applies when the key is missing.
_convert_content_part_to_anthropic() can preserve None from an
invalid upstream input text block. Now uses
(blk.get('text') or '').strip() on both paths.
2. Cache marker loss: prompt_caching.py's _apply_cache_marker() sets
cache_control directly on content[-1] for list content. If that last
part happens to be blank text, dropping it without relocating
cache_control silently loses the breakpoint. Both the normal and
replay paths now capture a dropped block's cache_control and reapply
it to the new last surviving cacheable block via the existing
_apply_assistant_cache_control_to_last_cacheable_block() helper
(setdefault semantics, so it never clobbers a legitimately-placed
marker).
3. Scalar whitespace: the non-list content branch
(blocks.append({'type': 'text', 'text': str(content)})) accepted a
truthy whitespace-only string unfiltered. Now filtered the same way
as list-content blocks.
8/8 new tests pass in TestBlankTextBlockFiltering (including None-safety,
scalar-whitespace, and cache_control-relocation regressions on both
paths); 186/186 in the full tests/agent/test_anthropic_adapter.py file.
OpenAI documents GPT-5.5 / GPT-5.5 Pro as extended-cache-only: in-memory
prompt cache retention is not available for them, and only
prompt_cache_retention: "24h" is supported. Responses requests that omit
the field see near-zero cached_tokens even with a stable prompt_cache_key
and identical prefixes (observed on an OpenAI-compatible Responses relay:
0 cached across repeated identical calls before; 97% cache reads after).
Send the field for the gpt-5.5 model family (bare and namespaced ids like
openai.gpt-5.5) on OpenAI-compatible Responses routes, mirrored in the
auxiliary Codex adapter, and pass it through preflight normalization.
Skipped for xAI, GitHub/Copilot, and the chatgpt.com Codex backend, which
reject or ignore body-level cache fields.
When api_key_hint from a 401 response doesn't match any pool entry
(common with OAuth tokens where runtime_api_key rotates), the pool
rotated without marking anything exhausted and handed back a fresh
selection. Because nothing was ever marked, the pool could never reach
the "no available entries" state — the caller retried the same dead
token forever (~6 attempts/sec), starving the event loop so /stop was
never processed; only killing the gateway ended it.
Rebased onto the identity-tracking rework that landed on main
(73c4b5a045): the single-entry escape from that commit already stops
the most common OAuth case, so this fix bounds the REMAINING gap —
multi-entry pools ping-ponging A->B->A with an unmatched hint. Cap
consecutive no-mark rotations at one full lap of the available
entries, then return None so the error surfaces / fallback activates.
Deliberately does NOT mark innocent entries exhausted (the original
PR's approach): that would quarantine a healthy key for the full
cooldown TTL on a hint that provably matches nothing. No cooldown is
written by the escape, so healthy keys stay available next turn --
bounded without hammering.
The streak resets when a rotation identifies a real entry and on any
successful normal select(), so only genuinely consecutive unmatched
rotations trip the bound.
Fixes#70401
Hardening follow-up to the #69619 review fix. The previous regression
byte-pinned only the rescued pre-#69619 generation; older frozen entries
were covered solely by fragment assertions and a self-matching loop that
cannot detect a frozen entry mutating (the loop tests each entry against
itself).
- Pin all four _HISTORICAL_SUMMARY_PREFIXES generations as literals in
_FROZEN_PREFIX_GENERATIONS and assert order-sensitive tuple equality
plus detect/strip for each
- State the prepend-only contract explicitly on the tuple: never mutate
or reorder existing entries
Negative controls verified: mutating, dropping, or reordering a frozen
entry each fail the new test, while the legacy self-matching loop still
passes under mutation — confirming the closed coverage gap.
Address review on #69619: the previous commit mutated the newest frozen
entry in _HISTORICAL_SUMMARY_PREFIXES and never froze the live prefix it
retired (the generation with both the four-heading discard clause and
the tools-active clause). A summary persisted immediately before
upgrading was therefore treated as an ordinary message on
resume/re-compaction, keeping the old handoff text embedded in the body.
- Prepend the exact pre-change live prefix as a new frozen entry
(newest-first), leaving all existing frozen entries byte-identical
- Restore the Jul 2026 (#65848 class) frozen entry to its original
four-heading text
- Pin the retired generation as a literal in
test_summary_prefix_semantics.py so mutating or dropping it fails CI
- Make the #65848 tool-use regression position-agnostic (match the
pre-clause generation by content, not tuple index)
Verified byte-identity of both rescued generations against the parent
commit. 233 focused prefix/resume/compressor tests pass.
Remove three directive-heavy section headers from both the LLM
and deterministic summary templates that caused the agent to
resume stale tasks after context compression:
- Historical In-Progress State
- Historical Pending User Asks
- Historical Remaining Work
These sections read as actionable instructions even within a
REFERENCE-ONLY wrapper, hijacking the user's latest message.
The remaining sections are purely descriptive/past-tense.
Frozen prefix copies in _HISTORICAL_SUMMARY_PREFIXES updated
to match. Test 8/8 passed.
Track the selected credential by stable pool entry ID so token refreshes and shared cursor movement cannot detach failures from the entry that issued them. Stop unmatched single-entry pools from reporting a no-op rotation as successful recovery.
Co-authored-by: Maxim Esipov <maksesipov@gmail.com>
Verification follow-up for the #51226 salvage: the host call site guarded
select_context with hasattr(), but the ABC defines a default on every
engine, so the built-in ContextCompressor (and any non-implementing
engine) still paid per-request shallow copies of the conversation
history plus a hook call on every provider request. Identity-check the
bound method against ContextEngine.select_context and return the
request untouched — mirroring the existing base-method short-circuit in
_notify_context_engine_turn_complete — so the default path does zero
work, not just produces an identical result.
Adds two pins: the base no-op is never invoked (patched-to-raise base
stays silent), and ContextCompressor.__dict__ contains neither new verb.
Also registers the contributor email mapping for @chaos-xxl.
- website guide: on_turn_complete() now carries the same best-effort coverage
caveat as the ABC docstring (fires from the finalization seam; abnormal
early-return paths bypass it) — removes the doc/code inconsistency.
- test: finalization seam emits on_turn_complete with usage=None + the
interrupted flag for an interrupted finalized turn. Docstring records that
the negative early-return-bypass half is best-effort and deferred to a
shared-seam follow-up rather than pinned via a full run_conversation harness.
Addresses the hermes-sweeper review on #51226:
- _apply_context_engine_selection now passes shallow copies of the read-only
conversation_messages / incoming_message to the hook, so an engine mutating
them in place cannot corrupt persisted transcript state (enforces the
request-only contract, not just documents it). Adds a mutation-regression
test asserting persisted history + incoming message are untouched.
- on_turn_complete docstring: scope the coverage claim to the standard
finalization seam. Some abnormal early-return paths (content-policy block,
provider terminal failure) currently persist+return without finalization and
don't emit the hook; documented as best-effort with a shared-seam follow-up,
rather than over-promising a guaranteed callback for every early exit.
- _apply_context_engine_selection: reject an empty list. all([]) is True, so
a [] returned by a failing/buggy engine previously replaced a valid request
with an empty message list the downstream sanitizers can't restore; now it
falls open to the unmodified request (honors the fail-open contract).
Thanks @johnnykor82 for catching this on #41918's review.
- test: empty list keeps the original request (fail-open regression).
- docs: document select_context()/on_turn_complete() in the public
context-engine plugin guide (were still describing only the old contract).
- context_engine.py: document that select_context() runs before cache-control
and all request sanitizers, so (a) replacements still pass host validation
and (b) the no-op default keeps the request byte-stable (AGENTS.md prompt-
cache invariant). Note the hook is evaluated per provider request.
- tests: no-op path is byte-stable for cache-control; a role-unusual
replacement is passed through for the existing downstream sanitizers to
normalize (select_context does structural validation only).
The on_turn_complete() observation hook is the engine's post-turn signal,
so it should receive the completed turn's canonical token usage when the
host has it, not a hardcoded None. Per @johnnykor82's #41918 contract: the
engine uses prompt/completion + cache_read/write/reasoning buckets to judge
how large/expensive the selected context was before the next select_context().
- conversation_loop.py: stash the most recent provider response's usage_dict
(the same canonical shape fed to update_from_response) on the agent as
_last_turn_usage; reset to None at turn start so turns that never reach a
provider response (early failure / interrupt) forward None, not a stale
prior turn's usage.
- turn_finalizer.py: forward agent._last_turn_usage instead of usage=None.
- context_engine.py: document the usage param contract on the ABC hook.
- tests: cover both ends through the real finalize_turn path — completed turn
forwards the full canonical bucket set intact; no-response turn forwards None.
Co-authored-by: johnnykor82 <johnnykor82@users.noreply.github.com>
Adds the post-turn observation verb as the companion to select_context():
an optional, no-op-default on_turn_complete() called once after the
assistant/tool loop finishes, with the finalized transcript snapshot. Lets
an engine ingest/index/summarize the completed turn to inform the next
select_context(). Wired via _notify_context_engine_turn_complete() from
turn_finalizer.finalize_turn(); fail-open, base no-op short-circuited so
non-implementing engines (incl. the built-in compressor) pay nothing.
This is the request-assembly + observation pair from #41918; with this
commit the PR fully subsumes #41918's two hooks (prepare_request_messages
-> select_context, on_turn_complete) rather than only the selection half.
Co-authored-by: johnnykor82 <johnnykor82@users.noreply.github.com>
Adds an optional, no-op-default select_context() hook to the ContextEngine
ABC, called every turn after the request messages are assembled and before
provider dispatch — independent of should_compress(). Lets an engine select
or replace which context enters the prompt for a single request (retrieval,
topic routing, role/branch switching) without mutating persisted history,
removing the need to abuse should_compress()=True as a per-turn callback.
The host call site (_apply_context_engine_selection) is fail-open: a missing
hook, an exception, or an invalid return value leaves the assembled request
untouched. Additive and non-breaking: the built-in compressor and every
existing engine are unaffected.
Consolidates the per-turn request-assembly surface proposed across #41918,
Related: #36765#41918#24949#47109#50053#23837#25115#29370
Keep MoA reference display events off the machine-readable -Q stdout
surface (platform=cli with tool_progress_mode=off) while preserving them
everywhere else. Extracts the relay into module-level helpers so the
policy is testable.
Salvaged from #67334.
Adds per-reference progress events and a phase-transition marker to the
MoA display pipeline so TUI / CLI / desktop surfaces can render a status
bar like `MOA: 2/3 refs done` and surface which phase (reference vs
aggregator) is currently active.
- `moa.progress` — fired once per reference completion with
`refs_done`, `refs_total`, and the source label
- `moa.phase` — fired on phase transitions (currently the single
`phase="aggregator"` transition once the fan-out
finishes)
Plumbed through the existing `reference_callback` →
`tool_progress_callback` → gateway path; no new UI surface. The legacy
`moa.reference` / `moa.aggregating` events are unchanged for backwards
compatibility.
AI-assisted fix by https://github.com/SquabbyZ/peaks-loop
When Hermes fails over from a non-Gemini provider (xAI, Anthropic, etc.) to
Gemini mid-conversation, the existing assistant tool_calls in history carry
no Gemini ``extra_content.google.thought_signature`` (the originating provider
never emits one). The native adapter's ``_translate_tool_call_to_gemini``
omitted ``thoughtSignature`` entirely in that case, so Gemini 3 thinking
models rejected every replayed turn with::
HTTP 400 INVALID_ARGUMENT
Function call is missing a thought_signature in functionCall parts.
Additional data, function call default_api:<tool_name>, position N.
The Cloud Code Assist sibling adapter already handles this exact case by
emitting a sentinel ``"skip_thought_signature_validator"`` (see
``agent/gemini_cloudcode_adapter.py:106``, originally added in #11270 and
documented as matching ``opencode-gemini-auth``'s approach). This change
mirrors that fallback in the native adapter so the two paths behave
identically when replaying cross-provider history.
Verified live against ``generativelanguage.googleapis.com/v1beta`` with
``gemini-3-pro-preview``: synthetic 2-turn conversation with no real
``thoughtSignature`` returns 400 without the sentinel and 200 with it.
Test added: ``test_build_native_request_emits_sentinel_for_cross_provider_tool_call``.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Follow-up fixes on top of the salvaged #22566 mechanism:
- N-collector now counts only REAL actionable user turns via
_is_actionable_user_turn + _is_synthetic_compression_user_turn —
the same filter pair _find_last_user_message_idx uses post-#69291.
The contributor's bare role=='user' + _is_context_summary_content
check let blank platform echoes and continuation/todo rows consume
N slots, silently degrading the guarantee.
- Default flipped 3 -> 1 (behavior-preserving): a default of 3 was
measured to change the tail cut on transcripts whose budget covers
only the last turn. min_tail_user_messages=1 delegates to the
existing single-user anchor; N>1 is opt-in, and the call site is
gated so the default path is byte-identical to main.
- Hardened config parse in agent_init (bool rejected, fractional
floats rejected, floor 1) matching the max_attempts parser shape.
- Wired the recurring external-PR config gaps: hermes_cli/config.py
DEFAULT_CONFIG + cli-config.yaml.example (PR only had cli.py).
- Regression tests: blank echoes / synthetic rows don't count toward
N; tool-call/result pairs never split by the N-boundary (no-orphan
both directions); N-guarantee wins over tail_token_budget and the
_MAX_TAIL_MESSAGE_FLOOR (floor is a minimum, not a cap); default
parity pin; DEFAULT_CONFIG pin.
Add _ensure_last_n_user_messages_in_tail to guarantee the last N user
messages survive compression in the uncompressed tail, with surrounding
assistant/tool context preserved.
- Add min_tail_user_messages parameter (default 3) to ContextCompressor
- New _ensure_last_n_user_messages_in_tail method generalizes single-user protection
- Skip context-summary handoff banners when counting user messages
- User messages are clean boundaries — skip _align_boundary_backward
- Wire through cli.py, agent_init.py, and gateway cache busting keys
Config:
compression:
min_tail_user_messages: 3
Co-Authored-By: Claude <noreply@anthropic.com>
_collect_ghosted_skill_names() covers both ghost-skill shapes in the
compressed middle window: rows already demoted to a [SKILL_PRUNED: ...]
marker AND raw skill_view bodies (> _SKILL_VIEW_PRUNE_MIN_CHARS) that
survived Phase-1 inside an earlier protected tail and then aged into the
compression window — the summarizer paraphrases those instructions away
too. Shared threshold constant between the emit site and the scan.
Pinned by a live-probe-shaped test (real compress(), mocked aux LLM).
21 tests pinning the salvaged #44166 behavior:
- marker emit + extractor round trip (patterns adapted from PR #32375
by @LeonSGP43, with credit)
- no-duplicate re-injection when the canonical marker survived (the
original PR's presence-check defect)
- Phase-1 protection for just-loaded / user-referenced skills, and the
Pass-4 pressure override that keeps #61932 fixed
- deterministic marker survival through a REAL compress() with a mocked
aux LLM: drop → re-injected, keep → not duplicated, static-fallback
path, iterative re-compression via rehydrated handoff
- markers never classify as handoff content (classify_summary_content /
_strip_context_summary_handoff_message untouched)
- SKILLS_GUIDANCE Skill Safety Rule renders with real newlines
The Vertex AI provider (added same-day, commit c73e74386) was never added to
either of the two provider registries that agent/auxiliary_client.py and the
MoA slot-resolution chain depend on, breaking Vertex outside the main
conversation loop:
1. hermes_cli/auth.py::PROVIDER_REGISTRY had no "vertex" entry. The
plugin-auto-extend loop that normally fills gaps explicitly skips
non-api_key auth types (`if _pp.auth_type != "api_key": continue`), and
Vertex was never hand-declared like "bedrock" is. Because
resolve_provider_client() in agent/auxiliary_client.py gates everything
on `pconfig = PROVIDER_REGISTRY.get(provider)` and returns (None, None)
immediately when pconfig is None, its `elif pconfig.auth_type == "vertex"`
branch was permanently dead code — every auxiliary Vertex call (vision,
title generation, reflection, context compression, MoA reference/
aggregator slots) failed outright, not just a MoA-specific edge case.
2. hermes_cli/providers.py::HERMES_OVERLAYS also had no "vertex" entry, so
hermes_cli.providers.get_provider("vertex") returned None. This backs
_preserve_provider_with_base_url() in agent/auxiliary_client.py, which a
MoA slot's resolved (base_url, api_key) pair needs to keep its "vertex"
identity instead of silently collapsing to "custom" — losing the
identity _refresh_provider_credentials() needs to re-mint an expired
OAuth2 token (~1h lifetime) on a 401, and permanently breaking every
subsequent call in that MoA preset for the rest of the session.
Fix mirrors the existing "bedrock"/aws_sdk entries in both registries
exactly, plus adds a "vertex" branch to _refresh_provider_credentials() (it
had branches for openai-codex/nous/anthropic/xai-oauth but not vertex,
so a 401 fell through to `return False` without evicting the stale cached
client).
- hermes_cli/auth.py: hand-declared vertex ProviderConfig(auth_type="vertex")
in PROVIDER_REGISTRY, matching bedrock's shape.
- hermes_cli/providers.py: vertex HermesOverlay(auth_type="vertex") in
HERMES_OVERLAYS + "Google Vertex AI" label override.
- agent/auxiliary_client.py: vertex branch in _refresh_provider_credentials
that re-mints the token via get_vertex_config() and evicts the stale
cached client.
- 8 new regression tests across tests/hermes_cli/test_vertex_provider.py and
tests/agent/test_auxiliary_client.py: registry membership, end-to-end
resolve_provider_client("vertex", ...) building a working client (proving
the previously-dead branch is now reachable), and the 401-refresh/cache-
eviction path.
Follow-ups on top of the cherry-picked #27748 mechanism:
- move the cap constant to module level with full rationale comment
(class attribute aliases it so subclasses/tests can override)
- bound the iterative-update path too: the PREVIOUS SUMMARY block is
passed through _bound_summary_input so a pathological rehydrated
handoff cannot blow up the prompt (previous summary + new turns each
capped)
- extra regression tests: byte-identical small-input passthrough
(identity), direct bound+marker unit check, bound-after-per-message-
truncation shape (hundreds of under-_CONTENT_MAX turns), iterative
path bounded, marker vs classify_summary_content non-collision
- contributor email mapping for @robgfl45
Follow-ups on top of the cherry-picked #62644 mechanism, porting it to
current main and closing the salvage-review requirements:
- proactive_prune_min_reclaim_tokens (default 4096): a prune only COMMITS
when it reclaims a meaningful token batch, measured on the pruned output.
A committed prune rewrites already-sent history and invalidates the
provider prompt-cache prefix; this hysteresis gate keeps those breaks
episodic/amortized (like a compression boundary) instead of firing every
tool iteration. 0 disables the gate. (Design point credited to the
#62389 review cycle's prune_minimum_tokens.)
- Standard no-op caller contract: every skip path returns the INPUT list
object; the loop commits only on 'result is not messages' + non-zero count.
- Loop call is getattr+callable guarded (plugin engines predating the hook,
SimpleNamespace test doubles) and exception-swallowed at debug level.
- Config parse follows the compression.max_attempts hardened semantics:
booleans rejected, fractional floats rejected, integral floats/numeric
strings accepted; negative trigger = disabled.
- cli-config.yaml.example documented (all three keys) and gateway
_CACHE_BUSTING_CONFIG_KEYS extended so hot-reload rebuilds the agent.
- Tests: min-reclaim gate both directions, input-object no-op contract,
no-orphan tool_call_id pairing in BOTH directions (#69830 pin rule),
default-off zero-behavior-change pin, config parse seam, and behavioral
loop-wiring tests (consulted/commit/no-op/absent-method/raising).
The phase-1 tool-result prune only runs inside compress(), which fires
near 50% of the context window, so it never triggers on large-window
models; old tool outputs then ride in history and are re-sent every turn.
Add prune_tool_results_only(): the same no-LLM prune on a separate, low
proactive_prune_tokens trigger, run as an elif to the compression branch.
Opt-in (default 0), protects the recent tail by message count.
Add the method to the ContextEngine base as a no-op default so pluggable
engines inherit it safely (the post-tool-call path never AttributeErrors on
a non-built-in engine); the built-in compressor supplies the real prune.
Register both keys under the top-level compression config with defaults and
document them.
Gateway half of the #49874 salvage: pass compression_deferred through
both _run_agent_inner result dicts and guard the compression-exhausted
auto-reset block with it — a lock-contended defer keeps the session
intact (the concurrent compressor is actively shrinking it) instead of
wiping it via reset_session.
Regression tests:
- tests/run_agent/test_compression_lock_defer.py — provider-mock 413 and
400-overflow turns whose compression pass lost the lock end as
compression_deferred (failed=False, no compression_exhausted); flag
unset keeps the terminal exhaustion path byte-identical; type-pin
tests vs MagicMock agents and junk flag values; cap=1 e2e proving the
refunded pre-API defer leaves the budget for the provider-proven
413 retry.
- tests/agent/test_preflight_lock_defer.py — a lock-skipped preflight
pass stops the loop WITHOUT arming preflight_compression_blocked;
plain no-op still arms it; MagicMock junk does not defer.
- tests/gateway/test_compression_deferred_soft_result.py — AST pin that
the deferred branch guards the auto-reset chain and performs no
session mutation (mirrors test_35809_auto_reset_clean_context.py).
MoA reference_max_tokens is preset-level — one cap for all reference
models. When mixing a verbose model with a terse one, a single cap is
either too tight for the terse model or too loose for the verbose one.
Now each reference slot can optionally carry its own max_tokens:
reference_models:
- provider: openrouter
model: deepseek/deepseek-v4-pro
max_tokens: *** # per-slot cap, overrides preset-level
- provider: openai-codex
model: gpt-5.5
# no max_tokens → falls back to preset-level reference_max_tokens
_clean_slot (moa_config.py) preserves an optional max_tokens field on
the slot dict, coerced via _coerce_int_or_none. _run_reference
(moa_loop.py) reads slot-level max_tokens first, falling back to the
preset-level cap passed by the caller. Slots without the field are
unaffected — backward compatible.
Type hints on slot-handling functions updated from dict[str, str] to
dict[str, Any] to reflect the now-heterogeneous slot shape.
Requested in review: builder-level assertions that the gemini-native
branch forwards max_tokens (provider names and the native
generativelanguage.googleapis.com base_url, max_tokens=600), plus a
control showing gemini models on OpenAI-compatible endpoints — including
Gemini's own /openai compatibility endpoint — keep the existing omission
behavior (#34530).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Per review feedback from teknium1: reference_max_tokens is an advisors-only
contract. The aggregator is the acting model and must not be capped by the
reference budget. Changed _is_moa from startswith('moa_') to exact match on
'moa_reference'. Added regression test proving aggregator does NOT receive
max_tokens.
Copilot review pointed out that hardcoding kwargs['max_tokens'] would
400 on models requiring max_completion_tokens (GPT-5 family, Copilot).
The existing auxiliary_max_tokens_param() helper already selects the
correct parameter name per model — use it instead of hardcoding.
Test updated to parametrize expected_key so the Copilot gpt-5.5 case
correctly asserts max_completion_tokens instead of max_tokens.
Addresses Copilot review comments on both files.
PR #56756 added reference_max_tokens to cap MoA advisor output and cut
turn latency. The value is correctly threaded through five layers of MoA
code (moa_config → conversation_loop → aggregate_moa_context →
_run_references_parallel → _run_reference → call_llm(task='moa_reference',
max_tokens=800, ...)).
However, _build_call_kwargs() in auxiliary_client.py silently drops
max_tokens for all OpenAI-compatible providers (PR #34845, which fixed
endpoints and NVIDIA NIM keep it. This means reference_max_tokens never
reached the API for the vast majority of providers.
The bug affects every OpenAI-compatible MoA reference/aggregator slot:
Z.AI (coding plan), OpenRouter, OpenAI, GitHub Copilot, and local
providers. Only Anthropic-compat endpoints (MiniMax, /anthropic URLs)
worked — by coincidence, not MoA-aware design.
Fix: thread the 'task' parameter through all six _build_call_kwargs()
call sites. When task starts with 'moa_', max_tokens is always included
in the request kwargs regardless of provider. Non-MoA auxiliary tasks
(compression, titles, vision, etc.) keep PR #34845 behavior unchanged.
Verified end-to-end:
- Z.AI GLM-5.2 with max_tokens=50 → returned exactly 50 tokens
- Z.AI GLM-5.2 with max_tokens=20 → returned exactly 20 tokens
- Z.AI GLM-5.2 uncapped → returned 315 tokens
- 7 new regression tests covering 4 providers, Anthropic wire, non-MoA
tasks, and prefix-matching boundary
- 288 auxiliary_client tests pass (was 281, +7 new), 84 MoA tests pass
- Zero regressions
aggregate_moa_context's single max_tokens parameter was applied to
both the reference fan-out (_run_references_parallel) and the
aggregator's own synthesis call_llm. #53580 explicitly removed a
hardcoded cap from the aggregator call because it truncated long
aggregator syntheses; #56756 (reference_max_tokens, added to speed up
the advisor fan-out) reintroduced the same shared cap by passing it to
both calls, silently regressing #53580's fix.
Rename the parameter to reference_max_tokens (matching the caller's
own moa_config key) and stop forwarding it to the aggregator's
call_llm invocation, which now always runs uncapped as intended.
An assistant message with an empty or whitespace-only text content block —
produced by context compression or certain tool-call flows — is rejected by
the Anthropic Messages API with HTTP 400 "text content blocks must contain
non-whitespace text". Because the blank block is stored in session history and
replayed verbatim every turn, the session is permanently wedged behind the same
400.
The Bedrock adapter already guards this via _safe_text() (#9486); the native
Anthropic path never got the same treatment. _sanitize_replay_block() rebuilt
text blocks with the raw stored text, and the _convert_assistant_message()
guard only caught a fully empty block list, not a list still containing a
whitespace-only text block.
Add a _safe_text() helper mirroring the Bedrock one and apply it at both points:
the ordered-blocks replay path and a final in-place walk of the converted
content list. Both are self-healing — sessions that stored blank blocks recover
on the next API call. Only text blocks are coerced; thinking/tool_use/image
blocks are untouched.
Fixes#69512
Correction to the previous commit (PR #68402): the claim that api_server
never intercepts MEDIA: tags is inaccurate on current main.
_resolve_media_to_data_urls() (gateway/platforms/api_server.py) DOES
inline image MEDIA: tags (<=5MB, image extensions only) as base64 data
URLs on the four main endpoints (_handle_session_chat,
_handle_session_chat_stream, _handle_chat_completions, _handle_responses).
The real gaps elphamale's PR points at are narrower:
- the /v1/runs output path (_handle_runs) never calls the resolver;
- non-image filetypes are never resolved anywhere (_MEDIA_IMG_EXT is
image-only).
Reword the hint to teach both halves: images via MEDIA: work on the
chat/completions/responses endpoints; non-image files and anything on
the runs endpoint must fall back to plain file paths in the response
text. Update the test to pin the scoped guidance instead of a blanket
prohibition.
Every PLATFORM_HINTS entry for a messaging platform (Telegram, WhatsApp,
Discord, Slack, Signal, WebUI, desktop) teaches the model the MEDIA:/path
convention because an interception mechanism actually resolves it there
(native attachment delivery, or a validated/inlined data URL). The cli
entry, which has no such mechanism, explicitly tells the model NOT to use
it and to state the path in plain text instead.
The api_server entry had neither instruction. Its /v1/runs handler never
routes the final response through any MEDIA: resolver (confirmed against
source: none of the four call sites of the api_server module's media-tag
resolver are inside its runs-endpoint handler), so a MEDIA:/path tag there
renders as inert literal text in the API response — exposing a raw host
filesystem path to the caller with no delivery ever taking place. Nothing
platform-specific told the model not to use a convention it's correctly
taught for several sibling platforms in this same dict, so the general
cross-platform habit could surface here too, unlike cli where an explicit
prohibition already closes the gap.
Mirrors cli's prohibition, adapted for api_server's actual constraint: no
"state the path in plain text" fallback, since a typical API caller has no
filesystem access to the host at all. Points at "a registered file-delivery
tool" generically rather than naming any specific tool, since api_server
toolsets are deployment-defined.
