Local faster-whisper called model.transcribe with bare {'beam_size': 5}:
no VAD, cross-window conditioning on, no confidence filtering. Pure
silence produced hallucinated tokens (E2E: 5s anullsrc WAV -> 'You',
no_speech_prob=0.705) and noisy clips could produce runs of junk, often
in other languages.
Three-layer class fix, one shared owner for every local-whisper call
site (build_local_transcribe_kwargs):
1. Silero VAD filter (bundled with faster-whisper) on by default —
silence never reaches the model. stt.local.vad: false restores the
raw behavior for music/ambient transcription.
stt.local.vad_min_silence_ms tunes chunk splitting (default 500).
2. condition_on_previous_text=False — one hallucinated token can no
longer seed a self-reinforcing run; negligible cost for
voice-note-length audio.
3. Segment confidence gate (_join_confident_segments): drop a segment
only when no_speech_prob > 0.6 AND avg_logprob < -1.0 (openai-whisper's
own heuristic shape; both must hit so quiet-but-real speech survives).
Config: stt.local.no_speech_prob_threshold / logprob_threshold.
The WHISPER_HALLUCINATIONS blocklist in voice_mode.py stays as
last-resort defense but should now almost never fire.
E2E (real faster-whisper 'base', CPU int8):
silence.wav before 'You' -> after ''
noise.wav before '' -> after ''
speech.wav before/after 'Hello World, this is a test of the
transcription system.' (unchanged)
Docs (EN + zh-Hans), DEFAULT_CONFIG, cli-config.yaml.example updated;
19 unit tests (kwargs contract, off-switch, confidence gate incl.
quiet-speech survival, _transcribe_local wiring), sabotage-verified.
Adds gpt-transcribe (OpenAI's new file-transcription model, $0.0045/min)
to the OpenAI STT provider:
- OPENAI_MODELS set: gpt-transcribe is recognized so provider
auto-correction keeps it on OpenAI and rejects it on Groq
- Language hint wiring: gpt-transcribe replaces the singular
'language' field with a 'languages' list; the API rejects the legacy
field, so the hint is sent via extra_body {languages: [..]}
- Config comment (DEFAULT_CONFIG), cli-config.yaml.example, desktop
settings enum, and docs (en + zh-Hans) updated
- Tests: model pass-through, languages-list hint shape, legacy singular
hint preserved for gpt-4o-transcribe, Groq auto-correction
gpt-live-transcribe (realtime WebSocket, $0.017/min) is NOT wired here:
the file-based STT pipeline has no realtime session path; it belongs in
a future realtime/voice-mode integration.
Local/self-updated macOS builds were finished with a plain
'codesign --force --deep --sign -', leaving a cdhash-only Designated
Requirement and stripping electron-builder's entitlements. Every rebuild
changes the cdhash, so TCC treats the new bundle as different code and
forgets Full Disk Access, Desktop/Downloads/Documents, Accessibility,
Automation, and microphone grants — users re-approve everything after
every update.
Rework the relaunch fixup to sign inside-out (standalone Mach-O
binaries, nested frameworks/helpers, then the main bundle), preserving
the repo's entitlement plists, and pin an identifier-based Designated
Requirement when signing ad-hoc so TCC has a stable identity to persist.
Opt-in desktop.macos_signing_identity names a persistent keychain cert
(self-signed Code Signing cert works — no Apple Developer account) for a
certificate-anchored DR, the strongest form. An intact Developer ID
signature is detected and never clobbered, callers can pass the
publisher-signing decision explicitly so a later dotenv load can't flip
it, and the legacy deep ad-hoc sign remains the last-resort fallback.
Co-authored-by: lewis4x4 <lewis4x4@users.noreply.github.com>
Co-authored-by: natebransc <natebransc@users.noreply.github.com>
Co-authored-by: caseyanthony <caseyanthony@users.noreply.github.com>
Co-authored-by: gvago <gvago@users.noreply.github.com>
Co-authored-by: twe-cloud <twe-cloud@users.noreply.github.com>
* Revert "fix(credits): remove the 'Grant spent · $X top-up left' notice"
This reverts commit 5dc6a14c14.
* fix(credits): reintroduce grant_spent behind an in-session crossing gate
The removed notice nagged because its condition is a steady state for
accounts living on top-up, the latch is per-session, and the cold-start seed
runs the policy at every session open — every session re-announced
'Grant spent · $X top-up left'. Reintroduce it gated so only a session that
WATCHES the grant run out announces:
- seen_grant_unspent crossing gate, mirroring seen_below_90: opens when the
session observes the grant meaningfully unspent (>= GRANT_UNSPENT_MIN_MICROS,
1 cent — portal-seeded states derive micros from float dollars and can carry
sub-cent residue where headers report exactly spent). Seeds never prime it.
- The gate guards only the show branch and is consumed by the announcement —
one announcement per crossing. Header flicker (used_fraction None and back)
clears the sticky line but cannot re-announce; a renewal re-opens the gate.
- new_credits_latch() centralizes the latch shape (agent build, lazy re-init,
and the policy test helper all build through it).
- Tests: steady-state-open stays silent (seed + policy seams), live crossing
announces once, flicker/renewal/residual cases locked; the rendering test
primes the gate and asserts its leg count so a gate regression cannot
silently shrink coverage.
Closes#55908. The CLI voice-mode beep amplitude is hardcoded at 0.3 inside
tools.voice_mode:play_beep(), which makes the record start/stop cues too
quiet on low-volume systems and headphones. Users couldn't adjust it
without editing source.
Move the literal into a configurable voice.beep_volume setting (clamped to
0.0-1.0, default 0.3 to preserve prior behaviour). The new
_get_beep_volume() helper reads via the same load_config() pattern used by
cli.py's _voice_beeps_enabled() and hermes_cli/voice.py's _beeps_enabled(),
keeps bools / out-of-range / non-numeric / NaN values safely on the default,
and falls back silently if config can't load so the audio cue never breaks
the voice loop on a degenerate config.yaml.
Covered by tests/tools/test_voice_mode.py:
- TestGetBeepVolume (12 cases: missing key, custom value, boundary 0.0/1.0,
out-of-range clamp, type coercion, bool guard, NaN guard, exception
guard, dict-typed voice section)
- TestPlayBeepVolumeWiring (guards against re-introducing a hardcoded 0.3
literal in play_beep)
Docs: website/docs/user-guide/configuration.md mentions the new key.
Other locale translations (zh-Hans etc.) intentionally untouched —
handled by the regular i18n sync pipeline as a separate change.
No change in default behaviour: existing users hear exactly the same beep.
Per-job cron inference pins are now user-owned: the agent-facing cronjob
tool schema no longer exposes model/provider/base_url, and the registered
handler ignores them even if a model hallucinates the old parameters.
Users set pins via the dashboard, hermes cron create/edit --model/--provider,
or jobs.json directly — and once set, a pin sticks until the user changes it.
Existing agent-era pins are grandfathered untouched.
New cron.model / cron.model_provider config keys give the cron fleet its
own default model, independent of the chat model. Fire-time resolution:
per-job pin > cron.model > HERMES_MODEL > model.default. An axis covered
by the explicit cron-fleet default is deliberate routing, not drift, so
the #44585 fail-closed guard skips it — switching your chat model with
/model or hermes model no longer breaks unpinned cron fleets.
- tools/cronjob_tools.py: drop model param from agent schema + handler;
remove now-dead _resolve_model_override
- cron/scheduler.py: cron.model/model_provider resolution + per-axis
drift-guard skip
- cron/jobs.py: snapshot resolution mirrors the new precedence
- hermes_cli/subcommands/cron.py + hermes_cli/cron.py: --model/--provider
on hermes cron create/edit
- hermes_cli/config.py: cron.model / cron.model_provider defaults
- docs: cron.md model-resolution tip rewritten
The grant_spent notice fired for every subscription user with top-up
funds the moment their cap was reached and camped in the CLI/TUI status
bar and desktop toasts with no action to take — the account keeps
working off top-up. Remove it everywhere:
- agent/credits_tracker.py: drop grant_cond + the emit/clear block;
the dev fixture state now (correctly) produces no notice
- TUI: keep the turn-start clear of credits.grant_spent as back-compat
for older backends that still emit the key
- Desktop: drop the demo step and stale comment references
- Docs/config comments: remove grant-spent from credits_notices text
- Tests updated: policy/cold-start now assert the key never fires
Usage bands, depleted, and restored notices are unchanged; /usage still
reports the full balance breakdown.
'hey hor' triggered the wake word: the default sensitivity was 0.5, which
for openWakeWord IS the raw per-frame score threshold — openWakeWord's own
permissive baseline that near-misses clear. Raised the default to 0.6 so
phonetic near-misses fall short while real 'hey hermes' (typically 0.9+)
still fires easily.
Also fixed a real cross-engine inconsistency found while checking: the
sensitivity knob is documented 'higher = stricter' and behaves that way
for openWakeWord (threshold = sensitivity) and sherpa (0.05 + 0.4*s), but
Porcupine's own 'sensitivities' param runs the opposite way (higher = MORE
false alarms, per Picovoice). Turning sensitivity up made Porcupine looser
— backwards. Now inverted (1 - sensitivity) so 'higher = stricter' holds
for every engine.
- tools/wake_word.py: default 0.6; _sensitivity fallback uses _DEFAULTS;
Porcupine sensitivity inverted with rationale
- hermes_cli/config.py + docs: default + consistent-direction note
- tests: Porcupine inversion, default>=0.6 regression, fallback-to-default
openWakeWord scores one ~80ms frame at a time and the detector fired the
instant a SINGLE frame crossed threshold — so a stray phoneme in background
conversation could trigger the wake word unintentionally (reported in
testing). A real utterance of the phrase holds a high score across several
consecutive frames; an ambient blip spikes just one.
_OpenWakeWordEngine now requires N consecutive over-threshold frames
(wake_word.confirmation_frames, default 3) before firing. The streak resets
on any sub-threshold frame and on engine reset() (pause/resume), so a
pre-pause frame can't count toward a post-resume fire. confirmation_frames=1
restores the old single-frame behavior; clamped 1..10.
Only openWakeWord is affected — sherpa (streaming transducer) and porcupine
decode the whole phrase internally and already reject single-frame spikes.
- tools/wake_word.py: _confirmation_frames() accessor, streak logic in
process()/reset(), config default
- hermes_cli/config.py: wake_word.confirmation_frames documented default
- tests: 5 new (spike rejected, sustained fires once, =1 legacy behavior,
reset clears streak, config clamp) — 58 wake tests green
- docs: 'Reducing false triggers on ambient speech' section
openWakeWord's ONNX backend returns near-zero scores on Apple Silicon
(dscripka/openWakeWord#336), so "Hey Hermes" never crossed the 0.5
threshold: the listener armed, the microphone worked, and nothing fired.
Bisecting the pipeline puts the fault in exactly one stage — feeding the
same audio through both backends, the melspectrogram front-end is
bit-identical (maxdiff 0.00000) and the wake classifier agrees on
identical features, while the shared embedding model diverges by 45.44.
Cross-feeding confirms it: tflite features scored through the *onnx*
classifier give 0.9948 vs 0.000009 for onnx features. A telling
secondary symptom is that scores fall as input gets louder (0.5x ->
0.00031, 8x -> 0.000066), which is garbage inference rather than a weak
detection.
Selecting tflite in config alone does not fix it. openWakeWord hardcodes
`import tflite_runtime.interpreter` but declares tflite-runtime for
`platform_system == "Linux"` only; on macOS the equivalent wheel is
ai-edge-litert, so that import always fails and model.py silently
downgrades back to onnx. The result is a detector that reports itself
listening and can never fire.
- default the backend per platform (tflite on macOS ARM64, onnx
elsewhere) instead of hardcoding onnx, and pick the matching bundled
model artifact
- bridge tflite_runtime -> ai_edge_litert through sys.modules, in-process,
with no writes to site-packages
- refuse the silent onnx downgrade on macOS ARM64 and report the missing
runtime through check_wake_word_requirements() so the GUI surfaces an
actionable hint rather than arming a dead ear
- lazy-install ai-edge-litert via its own feature key, because lazy-dep
specs cannot carry PEP 508 markers (_spec_is_safe rejects ";")
An explicit `inference_framework` in config still wins, so anyone pinning
a backend keeps it.
Verified on macOS 26.5.2 / M-series: "hey hermes" scores 0.0005 on onnx
and 0.9423 on tflite from the same clip, with cross-phrase controls at
0.0003. Live over-the-air through the real microphone fires 4/4
utterances (peak 0.9532).
One sherpa listener now enrolls every wake-enabled profile's phrase
(defaulting to "hey <profile name>") and reports WHICH phrase fired.
wake.detected gains a profile field; the desktop live-switches to the
matching profile (same path as the profile rail), opens a fresh session
there, and starts hands-free voice. The single-profile CLI/TUI print the
hermes -p switch command for foreign-profile phrases instead of
answering as the wrong profile. Opt out per listener with
wake_word.profile_routing: false.
New "sherpa" wake_word provider: the configured phrase is BPE-tokenized
at runtime against a small streaming zipformer KWS model (~13 MB English,
one-time download cached under HERMES_HOME), so ANY typed phrase works
with zero training — including per-profile phrases like "hey coder".
wake.sherpa lazy-dep group + [wake] extra grow sherpa-onnx/sentencepiece;
requirements probe routes per provider; sensitivity maps onto sherpa
keywords_threshold. E2E-verified on real audio (target phrase fires,
foreign phrase stays silent, reset drops buffered state).
From #53378: ships hey_hermes.onnx/.tflite (openWakeWord pipeline,
Apache-2.0) under tools/wakewords/, resolves the default (and hey_hermes
aliases) to the bundled file, ensures openWakeWord base feature models
are fetched for custom paths too, and updates config defaults + docs
from hey_jarvis to hey hermes.
Makes the wake word a tri-surface feature with one configurable owner.
- wake_word.surface ("auto" | "cli" | "tui" | "gui") + shared
wake_surface_enabled() gate consulted by every surface, so exactly one
place owns the listener and the new session it opens.
- tui_gateway: wake.start/stop/pause/resume/status RPCs + a wake.detected
event, sharing one server-side detector for both TUI and desktop. The
detector yields the mic to voice.record (pause on capture start, resume
on terminal) and to the desktop's browser mic (wake.pause/resume).
- TUI (Ink): arm wake.start on gateway.ready; on wake.detected open a
fresh session and start voice capture.
- Desktop (Electron): arm wake.start on connect; on wake.detected open a
fresh session.
- CLI now gates on wake_surface_enabled("cli"); /wake status shows surface.
- Tests for the surface gate; docs cover the surface knob + cross-surface.
Adds an opt-in, on-device hotword listener for the CLI. With
wake_word.enabled (or /wake on), Hermes listens in the background for a
wake phrase; on detection it starts a fresh session, captures one
utterance through the existing voice pipeline, and answers — the
"Hey Siri" pattern.
- tools/wake_word.py: provider-pluggable detector (openWakeWord, free
local default; Porcupine, premium) over the shared 16 kHz sounddevice
capture path. Background daemon thread with pause/resume so it yields
the mic during a voice turn.
- CLI wiring: startup listener (off-thread), on-wake flow, an idle
watchdog that resumes the detector after each turn, cleanup hook, and
a /wake [on|off|status] command.
- config.yaml wake_word section; PORCUPINE_ACCESS_KEY as an optional
secret. Engines lazy-install via the [wake] extra.
- Hands a transcript to the input queue exactly like voice mode, so no
system-prompt/cache mutation. No new core model tool.
- Tests (mocked, no live audio/network) + feature docs.
- Replace _load_cron_jobs_for_config_warning with lazy import of
cron.jobs.load_jobs — picks up BOM handling, corruption repair,
and context-local store resolution for free
- Re-add model.name to axis mapping (was dropped during merge);
model.name is a legacy alias for model.default
- Fix grammar: '1 enabled unpinned cron job have' -> 'has'
- Pass user_config to cron_model_drift_guard_enabled in set_config_value
to avoid a redundant load_config() re-read of the file just written
When an operator changes the global model/provider config, warn that
unpinned cron jobs with stored snapshots will fail-closed on their next
run. Adds a cron.model_drift_guard config opt-out (default true) for
fleets that should deliberately track changing global defaults.
Addresses #59031. Original PR #59177 by @doncazper.
Saying EXACTLY a configured stop phrase (default: 'stop') and nothing
else now ends the voice conversation instead of being sent to the agent
as a prompt. Match is deliberately strict — whole utterance,
case-insensitive, surrounding punctuation stripped — so 'stop doing
that and try X' still reaches the agent.
- tools/voice_mode.py: is_voice_stop_phrase() + voice.stop_phrases
config loader (default ['stop'], [] disables, malformed config falls
back safely).
- hermes_cli/voice.py: shared continuous loop (TUI + desktop) halts on
a stop phrase exactly like the silent-cycle limit (fires
on_silent_limit so every UI turns voice off); stop_continuous
force-transcribe path swallows the phrase without counting a silent
cycle.
- cli.py: classic CLI push-to-talk/continuous path and barge-in
utterance path disable voice mode on a stop phrase.
- Config default voice.stop_phrases: ['stop']; docs updated.
Tests: tests/tools/test_voice_stop_phrase.py (27 tests,
sabotage-verified: loop test fails when detection is disabled);
voice suites green (110 + 44 passed).
Whisper auto-detection frequently misidentifies short/accented clips,
which users experience as voice notes transcribed in the wrong language
(Teknium + CTO both hit this). The unified resolver from #73067 made a
global hint possible; this makes it the DEFAULT so stock installs stop
guessing. Non-English users set stt.language once; '' restores
auto-detect for multilingual use.
Deep-merge gives existing configs the new default automatically (no
_config_version bump needed); any explicit per-provider or global
language setting still wins.
Class-level fix for the 'STT transcribes the wrong language' issue family
(#55551, #50181 and siblings). Previously language handling was per-provider
chaos: local honoured stt.local.language, Groq/OpenAI/Mistral/DeepInfra sent
no language hint at all, xAI silently forced 'en', ElevenLabs used its own
language_code key, and there was no global setting.
- New _resolve_stt_language() helper: stt.<provider>.language >
stt.language (new global key) > HERMES_LOCAL_STT_LANGUAGE > auto-detect.
- Threaded through ALL providers: local, local_command, groq, openai,
mistral, xai, elevenlabs, deepinfra (shared OpenAI handler), command
providers, and plugin dispatch.
- xAI no longer forces English when nothing is configured (auto-detect).
- Mistral Voxtral now receives a language hint when configured.
- stt.groq.model is now honoured from config (previously env-only).
- DEFAULT_CONFIG gains stt.language, stt.groq, stt.xai, stt.mistral.language.
- Tests: tests/tools/test_stt_language_resolution.py (11 tests, sabotage-
verified) + full transcription suite green (236 passed).
Builds on cherry-picked contributor work from #19786 (@zombopanda),
#23161 (@materemias), #50684 (@BlackishGreen33).
Add optional stt.openai.language config for OpenAI transcription. Forward non-empty hints to the API while preserving auto-detection when unset. Document the config-only setting, add its default, and cover configured and unset request arguments.
Per Teknium: the caps should bound a single agent loop, not accumulate
over the whole session. Rename SessionCapConfig -> LoopCapConfig and the
config section session_caps -> loop_caps; move the counters into
reset_for_turn (invoked per turn via turn_context) so each turn starts
with a fresh budget; retune defaults 200 -> 50 (a single turn issuing 50
web searches / spawning 50 subagents is already pathological). Block
codes session_*_cap -> loop_*_cap and messages updated to drop the
/new-resets-the-budget guidance (irrelevant now that it resets per turn).
Tests flipped: the old persists-across-turn-resets assertion becomes
resets-each-turn.
Add per-session lifetime caps on web_search calls and subagent spawns
(defaults 200/200, matching Claude Code v2.1.212). Unlike the existing
per-turn tool-loop guardrails, these count over the whole session and
reset only when a fresh agent is built (/new, /clear). Hitting a cap
blocks the offending call and halts the turn cleanly.
- agent/tool_guardrails.py: SessionCapConfig + session counters on the
controller (in __init__, not reset_for_turn, so they persist across
turns). before_call() enforces caps first, independent of
hard_stop_enabled. delegate_task batches count each task.
- hermes_cli/config.py: tool_loop_guardrails.session_caps defaults.
- docs + tests (unit + E2E validated against a real AIAgent).
After N consecutive guardian denials in a session the deny message escalates to a hard-stop instruction. Inspired by ChatGPT Work auto-review circuit breaker.
DEFAULT_CONFIG declared "kanban" twice. Python keeps only the last
literal for a duplicate key, so the first kanban block was silently
dropped and its "auto_subscribe_on_create": True default never made it
into DEFAULT_CONFIG. The consumer in tools/kanban_tools.py masks the
miss with cfg_get(..., default=True), but the documented default was
absent from DEFAULT_CONFIG (so config templates / 'hermes config show'
omit it), and the duplicate key is a standing hazard: any future key
added to the first block would also vanish.
Merge auto_subscribe_on_create into the single canonical kanban block.
Adds a regression test asserting both default sets survive and a guard
against any duplicate top-level DEFAULT_CONFIG key.
The default max_iterations/agent.max_turns budget was set when long
agentic runs were rare; complex tasks now routinely exceed 90 tool
calls. Raise the default to 500 across every surface that hardcodes
the fallback: AIAgent constructor, DEFAULT_CONFIG, CLI resolution
chain, gateway env bridge, cron scheduler, and TUI gateway. Explicit
user config values are unaffected (deep-merge preserves them; no
_config_version bump needed).
Docs (en + zh-Hans), CLI help text, tips, and pinned tests updated
to match.
Teknium review changes on the tiered policy:
1. threshold_pct default 10 -> 5 (listing budget = min(5% of context,
listing_max_tokens)); unknown-context fallback 20K -> 10K.
2. Tier 2 no longer leaves the model blind: when even names-only doesn't
fit, the bridge description now carries a one-line-per-server summary
('cloudflare (3320 tools)') plus an instruction to search FIRST rather
than substitute a generic tool or claim the capability is missing —
the measured tier-2 failure mode (core-tool substitution) at zero
meaningful token cost (~50 tokens/server).
3. Listing degradation is now PER SERVER, largest first: one oversized
server (Cloudflare) collapses to its summary line while small
co-attached servers (Linear) keep their full per-tool listings
('mixed' form). Previously global: attaching Cloudflare next to
Linear silently cost Linear its listing. Greedy fit is deterministic
(size then label) so the rendered block stays byte-stable per catalog
— prompt-prefix cache safe.
E2E on real captures (defaults, 200K ctx): linear alone -> tier 1 full;
unreal alone -> tier 2 groups (5% budget) / tier 1 names at 1M;
cloudflare alone -> tier 2 groups; linear+cloudflare -> tier 1 MIXED
(linear fully listed, cloudflare summarized). 48/48 tests.
Tier 0: no MCP/plugin tools -> everything eager (pass-through).
Tier 1: deferred tools whose catalog listing fits min(threshold_pct%
of context, listing_max_tokens) -> bridge + skills-style
listing, degrading to names-only over budget.
Tier 2: listing over budget even names-only (Cloudflare's flat API
surface: 3,320 tools, names alone ~32K tokens) -> bare bridge,
discovery through tool_search only.
The old activation threshold (defer only when schemas > threshold_pct
of context) let mid-size catalogs ride eager and pay full schema cost;
with servers like Cloudflare (~597K tokens of schema, would not even
fit a 200K window) the binary gate is the wrong shape. Activation is
now driven purely by deferrable-tool presence; threshold_pct is
repurposed as the listing budget's context-relative leg.
- AssemblyResult gains tier + listing_form for observability
- listing_max_tokens default 4000 -> 20000 (cap 60000) so an 830-tool
catalog keeps a names-only listing while Cloudflare-scale drops to
bare bridge
- E2E verified against real captures: Linear 24 tools -> tier 1 full,
Epic UE 5.8 830 tools -> tier 1 names-only, Cloudflare 3,320 tools
-> tier 2 (both 200K and 1M context)
Deferred MCP/plugin tools become invisible once the tool_search bridge
activates — live benchmarking (48 runs, Claude Haiku 4.5) showed models
substituting visible core tools (terminal/web_search/browser) for deferred
capabilities or declaring them nonexistent instead of searching: 16/24
task success vs 24/24 with eager loading.
Skills never had this failure mode because every skill keeps a ~21-token
name+description listing line in the system prompt. This ports that exact
pattern to the tool bridge: when tool_search activates, a grouped manifest
of every deferred tool (name + first sentence of description, clipped to
60 chars, grouped per MCP server / toolset) is embedded in the tool_search
bridge description.
- tools/tool_search.py: build_catalog_listing() with deterministic
ordering (byte-stable across assemblies -> prompt prefix stays
cacheable); token-budget fallbacks full -> names-only -> legacy bare
count; bridge_tool_schemas(listing=...) embeds it and instructs the
model to skip tool_search when the exact name is visible (one fewer
round-trip per use)
- config: tools.tool_search.listing auto|on|off (default auto),
listing_max_tokens (default 4000, clamped 200..20000); legacy bool
shapes keep working
- tests: 8 new tests (config parsing/clamps, short-desc clipping,
deterministic rendering, budget fallbacks, bridge embedding, assembly
on/off paths); full file green (47 passed)
- docs: tool-search.md config table + rationale
- scripts/tool_search_livetest2.py: benchmark harness v2 with real
per-call token accounting (normalize_usage spy) and a third 'listing'
mode for A/B/C comparison
2107b86024 flipped compression.in_place to True but left both explanatory
comments reading "Default False during rollout". The contradiction is
load-bearing: it is why two recent PRs (#71747, #48951) were built on the
premise that agent.session_id still rotates at every compaction.
Comment-only; no behavior change.
Some OpenAI-compatible endpoints — notably Tencent Copilot
(copilot.tencent.com) — only accept streaming chat requests; any
non-streaming call returns HTTP 400 (code 11101, 'Non-stream chat
request is currently not supported'). The main conversation loop already
streams, so interactive chat works, but every auxiliary task (title
generation, compression, web extraction) used the non-streaming path and
failed on each call.
_provider_requires_stream() detects stream-only endpoints
(copilot.tencent.com built in, plus user-configurable
auxiliary.stream_only_base_urls substring markers in config.yaml).
Matching sync auxiliary calls route through _create_with_progress
(force_stream=True) and async calls through the new
_acreate_with_stream, aggregating the chunk stream — including tool-call
deltas and reasoning deltas — into a complete response via the shared
_ChatStreamAccumulator.
Salvaged from PR #60686 by @kudi88 onto the progress-aware streaming
machinery from #71508, addressing both sweeper-review gaps: the async
path now consumes the stream with 'async for' (awaiting create() and
iterating synchronously raised on AsyncOpenAI streams), and tool-call
deltas are reassembled instead of dropped (MCP passes tools= through
call_llm). Under force_stream there is no silent non-streaming retry —
a stream-only provider rejects those by definition, so the original
error surfaces to the normal recovery chains.
The gateway's pre-agent session-hygiene compression killed the summary
call at a fixed 30s wall-clock deadline (compression.hygiene_timeout_seconds),
regardless of whether the summary model was hung or merely slow. A reasoning
model happily streaming a large summary was cut off mid-generation, the user
got '⚠️ Context compression timed out after 30.0s', and a 300s failure
cooldown left the session oversized — a doom loop for slow-but-healthy
auxiliary models.
Timeouts are now liveness-based instead of wall-clock-based:
- agent/auxiliary_client.py: new thread-local aux_progress_hook. When
installed (only by context compression today), the primary call_llm
attempt streams (stream=True) and aggregates chunks back into a complete
response, ticking the hook per chunk. The configured timeout then acts
per stream read (idle) instead of as a total budget. Providers that
reject streaming fall back to the plain non-streaming call; auth/payment/
rate-limit/transport errors propagate unchanged into the existing
recovery chains. Codex Responses (per SSE event) and Anthropic Messages
(per stream event, via the new create_anthropic_message on_stream_event
callback) tick the same hook from inside their wire adapters.
- agent/conversation_compression.py: CompressionCommitFence gains
touch_progress()/seconds_since_progress(); compress_context() installs
fence.touch_progress as the progress hook around the compress call.
- gateway/run.py: the hygiene wait loop treats hygiene_timeout_seconds as
an INACTIVITY budget — while the fence reports fresh progress the wait
extends, bounded by the new compression.hygiene_total_ceiling_seconds
(default 600s, clamped >= the idle budget) so a degenerate trickle
stream still dies. The timeout warning now says the summary model
produced no output, which is the only case that still triggers it.
- config/docs: hygiene_total_ceiling_seconds added to DEFAULT_CONFIG and
configuration.md; hygiene_timeout_seconds documented as inactivity-based.
Tests: tests/agent/test_aux_progress_streaming.py (hook plumbing, stream
aggregation incl. tool-call deltas and reasoning deltas, rejection
fallback, ceiling kill, fence progress surface); two new gateway tests
prove a slow-but-streaming worker survives past the fixed timeout
(sabotage-verified: fails with the old fixed deadline) and a
forever-trickling worker is still cut off at the ceiling.
Leg 2 of #63078: prompt.submit returns {"status":"streaming"} immediately and
runs _start_agent_build + _wait_agent(timeout=30s) behind it. The deferred
build (MCP discovery with per-server retry backoff, synchronous model-metadata
HTTP, skills scanning) routinely outlives 30s on cold starts; on timeout
run_after_agent_ready emitted an error EVENT and returned without ever calling
_run_prompt_submit — the user's first message was permanently discarded while
the build finished successfully in the background. The desktop's optimistic
row eventually cleared with no visible error: the blank first session.
New _wait_agent_for_prompt replaces the flat cliff for the deferred prompt
path only (_sess()'s RPC-blocking _wait_agent keeps its 30s contract):
- The pending prompt stays attached to the (already off-RPC) run thread and
is delivered the moment the still-running build completes — a slow build is
no longer message loss.
- The wait runs in 5s slices so a cancel (session.interrupt / churn) is
honored promptly; the cancelled path returns None and defers to the
caller's cancel branch (the #65567 emit) for user-visible messaging.
- Past 30s the client gets ONE keyed notification.show ('Still starting the
agent…', key=agent-build-slow, desktop toast / TUI status bar), cleared on
delivery — patient, never silent.
- Permanent failure only when the build itself fails: agent_error set at
ready, the build thread died without signalling ready (fail fast via the
new _agent_build_thread handle instead of sitting out the cap on a corpse),
or the bounded cap expired on a genuinely hung build. The cap defaults to
600s and is tunable via agent.build_wait_timeout in config.yaml (no new
env vars); the error message states the message was not sent.
Tests: slow-build delivery with zero error events; the keyed progress notice
shown once and cleared; build-failure surfacing exactly one error event with
the real reason; dead-thread fail-fast; cancel honored mid-wait; config
override + fallback semantics; cap expiry message. The compute-host fallback
test stubs the new waiter alongside _wait_agent.
