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.
Two bugs surfaced by the desktop wake conversation:
1. Runaway loop: wake -> voice -> resume -> wake fired again within
~200ms. openWakeWord keeps its rolling feature buffer across
pause/resume, so on resume it immediately re-scored the "hey jarvis"
captured before the pause and re-fired, reopening a session and
restarting voice in a tight cycle. Reset the engine buffer on every
detector (re)start so resume begins from clean audio.
2. Empty-transcript toast: a silent re-listen returns
success:false / "… STT returned empty transcript", which the desktop
transcribe endpoint turned into a 400 -> thrown error -> "Voice
transcription failed" notification on every silent gap. Treat an empty
transcript as no-speech: return {ok, transcript: ""} so the voice loop
quietly re-listens. Real failures still 4xx/5xx.
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.
CI shard 4 caught 4 failures the local baseline diff missed (local env
pollution made them look pre-existing): _make_compressor() constructs
under a get_model_context_length mock but the lazy deferral (#32221)
pushed resolution past the with block, so the 96K window / 75% floor the
tests rely on never materialized. Resolve inside the mock.
Review-pass findings on the lazy-init deferral:
- No-op guard: the codex app-server usage callback assigns
compressor.context_length on EVERY response (same window each time).
The setter unconditionally invalidated the derived budgets, wiping
runtime corrections applied directly to threshold_tokens /
tail_token_budget (aux-context threshold sync) — those persisted on
main's eager init. Same-value assignment is now a no-op.
- Re-floor on genuinely new window: the setter invalidates budgets but
previously kept the stale threshold_percent, so a codex window switch
recomputed threshold_tokens from the new window with the old model's
floored percent. Re-apply the raise-only small-context floor from
_base_threshold_percent so percent and tokens derive from the same
window (guarded with getattr for object.__new__ test instances).
- Init log extracted to _emit_init_summary_once() and also fired from
the setter path, so a consumer assigning context_length before any
read no longer strands the startup line forever.
- threshold_tokens getter resolves the window into a local before
reading threshold_percent — correctness no longer depends on
left-to-right argument evaluation order.
- reasoning_timeouts: fix inaccurate 'tuples are immutable' comment
(the container is a list; safety comes from build-once-at-import),
document why the slug stays in the tuple.
Adds TestContextLengthSetterCoherence (3 tests): same-value assignment
preserves overrides; new-window assignment re-floors both directions.
With the selective prompt-cache copy (#57046), un-marked messages on the
decorated api_messages list share their nested content parts with the
persistent conversation history — the per-message copy in
conversation_loop is shallow and decoration now deep-copies only the
marked messages. try_shrink_image_parts_in_messages previously wrote the
re-encoded image INTO the aliased part/source dicts, so an
image-too-large retry on an Anthropic route would silently replace the
original image bytes in agent.messages (and persist the degraded copy).
Replace the in-place writes with copy-on-write: rebuild the content list
with fresh part/source/image_url dicts and reassign msg['content'] — a
top-level write on the per-call copy that never reaches history.
Adds two regression tests simulating the aliasing; both fail against the
old in-place implementation (mutation-verified).
Baseline diff vs clean upstream/main (182 pre-existing environmental
failures on both sides) showed exactly 6 PR-caused failures, all the same
mechanism: tests construct ContextCompressor under a
get_model_context_length mock and read threshold_percent/threshold_tokens
after the with block, or build via object.__new__ and assign
context_length AFTER the derived budgets (the setter now resets the
lazily-cached budgets).
- test_compression_small_ctx_threshold_floor: resolve inside _make()
- test_cjk_token_estimation: resolve inside mock
- test_per_model_compression_threshold: resolve inside mock (2 tests)
- test_pre_compress_memory_context: assign context_length before
threshold/tail/summary budgets; add summary_target_ratio
TestThresholdTokensCap and TestLazyContextResolution landed on main after
the #38991 lazy-init base; they construct ContextCompressor under a
get_model_context_length patch and read threshold_tokens after the with
block. With deferred resolution the probe now fires lazily, so resolve
inside the mock (same pattern as the rest of the suite) and give the
lazy-resolution mock a real return_value.
Self-review found the deepcopy->shallow-copy change in
apply_anthropic_cache_control (#57046) had no test pinning the
"prompt caching is sacred / never mutate the caller's list" invariant.
The old test_returns_deep_copy only exercised the single marked message,
never an un-marked shared reference, so a regression that deep-copied too
little would pass the whole suite.
Add two tests: (1) caller list + every element left byte-identical after
the call, un-marked middle messages returned as shared references, marked
messages fresh copies, and mutating a returned marked message does not leak
upstream; (2) structural byte-equivalence vs a reference full-deepcopy
implementation across both native_anthropic modes and two TTLs.
Mutation-verified: neutering the per-message deepcopy makes test (1) fail.
The lazy-init change in #32221 deferred get_model_context_length() out of
ContextCompressor.__init__, but the "Context compressor initialized" log
(emitted only when quiet_mode=False) reads self.context_length,
self.threshold_tokens and self.tail_token_budget. Those property reads
resolve the deferred value, so the synchronous model-metadata probe still
ran inside __init__ on the interactive-CLI path — the exact blocking the
PR removed, just narrowed to the non-quiet path.
Emit the informative line once, on first context-length resolution, so
construction stays non-blocking on every path. Add a regression test
asserting no probe fires in __init__ with quiet_mode=False and that the
init log is emitted exactly once on first access.
The structural AST walker in test_compression_session_id_persistence.py
only recursed into control-flow children found via iter_child_nodes(stmt).
When a session_entry.session_id assignment lives inside an `else` block
whose statements are all assigns (no nested If/Try/etc to trigger
_walk_node), the assignment was invisible to the walker and the test
florped with 'No assignments found'.
Walk the stmt itself when it is a control-flow node so its
body/orelse/finalbody (and Try handlers) are always expanded, regardless
of whether iter_child_nodes yields an inner control-flow child.
Manual /compress already persists the rotated child transcript first and
only then repoints the live session_entry; a False rewrite_transcript
return keeps the entry on the original session_id so the conversation
stays reachable. Session hygiene auto-compress did the opposite: it
rebound session_id (and lease/topic) first, then called rewrite_transcript
without checking the return value. On a failed write the live entry
already pointed at an empty child SID and the turn continued — permanent
silent conversation loss. Persist first; rebind only after success.
Phase 2 review found two sibling sites with the same bug class:
- truncate_before_user_ordinal in prompt.submit counted display_kind
timeline rows as user turns, shifting the truncation target
- rollback.restore used the old pop-loop pattern that would pop a
display_kind marker instead of the last real exchange
Both now use the same predicate (role==user and not display_kind)
matching list_recent_user_messages, /undo, /retry, and CLI resume.
Added tests for both paths.
list_recent_user_messages and the in-memory /retry + session.undo walkers
treated every role=user row as a real user turn. Timeline bookkeeping
(model_switch, async_delegation_complete, auto_continue, hidden) is stored
that way, so /undo soft-deleted from a marker and /retry re-sent opaque
bookkeeping text. Exclude display_kind the same way CLI resume counting and
the prompt.submit ordinal path do.
Replaces the O(N) list_sessions_rich histogram in /api/sessions/stats
with a single GROUP BY query, reducing response time from ~575ms to
<1ms on large databases.
Original PR #48921 by @liuhao1024. Salvage fixes based on review
feedback from teknium1 and @wernerhp:
1. Preserve try/except guard — a DB error still degrades to empty
by_source instead of failing the whole stats response.
2. GROUP BY COALESCE(source, 'cli') — the original GROUP BY source
could emit duplicate 'cli' keys (NULL group + literal 'cli' group)
that the dict comprehension silently dropped.
3. Add exclude_children=True — list_sessions_rich excludes subagent
runs, delegates, and compression continuations by default; the
bare GROUP BY counted all rows, inflating source counts.
Aggregate shape (exclude_children/include_archived/limit params)
adapted from closed duplicate #61120 by @mijanx.
Closes#48914
Co-authored-by: mijanx <mijanx@users.noreply.github.com>
Every LLM call built a fresh openai.OpenAI wire client (new httpx pool,
TCP+TLS handshake, measured 19.2ms p50 / 35.5ms p95 per call at ~5 calls
per tool-loop turn) and closed it when the request finished. Cache ONE
reusable wire client on the agent, keyed by the effective client kwargs:
- _create_request_openai_client hands back the cached client when the
effective kwargs are identical; any change (credential rotation,
provider failover, vision default_headers) evicts and rebuilds.
- Only a request that produced a response reports a reuse close reason
(request_complete / stream_request_complete); error unwinds report
*_error_cleanup and really close, so a retry after a request error
always builds a fresh pool.
- Cross-thread aborts poison the slot: a pool whose sockets were
shutdown(SHUT_RDWR) from a stranger thread is never reused (#29507) —
the owner-thread close discards it and the next create rebuilds. The
holder read and the abort are atomic (under the holder lock) at all
three abort sites, so a late abort can never poison the NEXT request's
checked-out client.
- Worker-side interrupt breaks close the half-read SSE stream on the
owning thread before building the partial response; a failed close
poisons the slot (otherwise each interrupt leaked one checked-out
connection until the pool hit PoolTimeout). run_codex_stream gets the
same poison-on-close-failure handling.
- Single checked-out slot (in_use): a concurrent call gets an untracked
client with the old per-request lifecycle.
- release_clients() / close() really close the cached client when idle;
if a worker has it checked out they abort the sockets and detach the
slot, deferring the FD release to the worker's own close.
- MoA facade and Mock passthroughs never enter the cache; max_retries=0
is preserved on all request clients.
Follow-up to PR #64171. The _TOKEN_DELTA_* classification must exactly
cover update_token_counts' keyword surface: an unclassified kwarg is
silently kept only from the first delta of a merged run. Introspect the
live signature and fail with a pointed message when a future kwarg is
added without classification (or a classified field is removed).
Follow-up to PR #64171. The writer loop and flush_token_counts'
caller-drain both set _token_writer_busy BEFORE popping the queue —
that ordering is what makes flush's lock-free fast path (reads queue
then busy, no cond held) sound. _stop_token_writer's leftover drain did
it backwards (clear queue, then set busy), leaving a few-bytecode window
where a concurrent flush could observe 'empty and idle' and return True
with the popped batch still unapplied. Shutdown-only staleness, no data
loss — but the protocol now matches at all three drain sites.
Test: concurrent flush during a stop-drain mid-apply must time out
(False), never report drained.
Follow-up to PR #64171. Two writer-death hardening gaps:
- queue_token_counts only spawned the writer when the thread object was
None, so a writer that died from an unexpected exception could never be
replaced: deltas piled up on the uncapped deque until a reader's flush
drained them synchronously. Respawn on 'not thread.is_alive()' instead.
(atexit re-registration on respawn is safe: unregister removes all equal
bound-method registrations and the drain hook is idempotent.)
- _coalesce_token_deltas ran outside the per-delta try/except in
_apply_token_batch, so a merge bug (e.g. an unclassified future kwarg
summing None + 0) would escape and kill the writer thread. Wrap it and
fall back to applying the raw batch — coalescing is an optimization,
never load-bearing.
Tests: coalesce-failure fallback + dead-writer respawn.
Every API call in the tool loop persisted its token/cost delta by
calling SessionDB.update_token_counts() synchronously on the turn
thread — a BEGIN IMMEDIATE sessions UPDATE plus a session_model_usage
upsert, measured in production at p50 3.3ms / p95 70.4ms per call and
up to 299ms against a cold multi-GB state.db. The tool loop stalls for
that long between calls, N times per multi-tool turn.
SessionDB gains queue_token_counts(): same signature and semantics as
update_token_counts(), but the critical path is a deque append plus a
condvar notify. A lazily started daemon thread applies deltas in
enqueue order through the existing update_token_counts ->
_execute_write path, so the established self._lock / BEGIN IMMEDIATE /
jitter-retry discipline is unchanged. When a backlog forms, adjacent
same-route incremental deltas coalesce into one UPDATE: token and
api-call fields sum, cost fields sum None-preservingly (an all-None
run stays None so COALESCE keeps the stored value), and absolute=True
deltas never merge and act as ordering barriers. Route equality is
required for a merge because those fields feed COALESCE backfill, the
last-non-None-wins status fields, and the per-model usage attribution
key — a merged apply is row-equivalent to sequential applies.
Correctness and durability:
- flush_token_counts() gives read-your-writes to token/cost readers
(get_session, list_sessions_rich, _get_session_rich_row,
list_gateway_sessions, InsightsEngine.generate) — a plain attribute
check when nothing is queued. The writer sets its busy flag before
popping the queue so the lock-free fast path can never miss an
in-flight batch.
- update_session_model / update_session_billing_route /
update_session_meta write the sessions row synchronously, bypassing
the queue, so they flush it first: a still-queued first-of-session
delta carries the pre-switch route, and applying it after the switch
UPDATE would trip the first_accounted_route branch (api_call_count
== 0 plus a route mismatch) and resurrect the old model/provider.
- AIAgent._persist_session flushes at turn finalize and every
error-exit persist point; close() stops and drains the writer before
the WAL checkpoint; an atexit hook (registered on first enqueue,
unregistered on close so closed instances are not pinned until
interpreter exit) drains at shutdown. Worst-case crash loss is the
in-flight call's delta — the same window as the old inline write.
- A flush trusts a live stop-flagged writer (its loop drains before
exiting) and only drains on the caller's thread when the writer is
dead or never started, claiming the same busy flag so concurrent
flushes wait instead of racing an in-flight batch.
- After close() has stopped the writer, queue_token_counts applies the
delta inline instead of parking it on a queue nothing will drain; a
closed-connection failure then raises at the call site, which
already guards for it, exactly like the old synchronous path.
- Writer apply failures are logged and never raise into a turn; the
writer thread survives and keeps applying.
Call sites switched to the queue: the per-call site in
agent/conversation_loop.py and both codex app-server sites in
agent/codex_runtime.py. In-memory per-turn counters
(agent.session_estimated_cost_usd etc.) stay synchronous, so live turn
displays never see the queue.
Tests: tests/agent/test_async_token_accounting.py (19 tests: enqueue
ordering, absolute-as-barrier, backlog coalescing with exact sums,
coalesced-vs-sequential row equivalence, merge unit rules, None-cost
preservation, read-your-writes, flush vs stop-flagged/concurrent
drains, inline apply after writer stop, close/atexit durability,
_persist_session drain, writer failure isolation);
tests/run_agent/test_token_persistence_non_cli.py updated to the
queue_token_counts contract.
Critical fixes to salvaged PR #73020:
- Use SessionDB.append_message instead of raw INSERT INTO messages.
The original used wrong column names (session_key/created_at vs
session_id/timestamp) and bypassed FTS indexing, session metadata
updates, display_kind, and all other columns append_message handles.
- Use get_hermes_home() instead of hardcoded Path.home()/'.hermes'.
Profile-aware path resolution under HERMES_HOME override and active
profile isolation.
- Add 11 tests covering flush, recovery, serialisation, edge cases.
- Change await self._send_read_receipt to asyncio.create_task to avoid
blocking message dispatch on slow bridge responses (matches BlueBubbles
pattern). Up to 5s per-message latency eliminated.
- Update tests: assert_called_once_with instead of assert_awaited_once_with
since the receipt is now scheduled, not directly awaited.
- Add send_read_receipts documentation to whatsapp.md following the
BlueBubbles docs pattern.
- 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.
Skip hashing active copies when the bundled origin is unchanged, build rename and optional migration indexes lazily, and reuse one optional-skill directory index per sync.
Add regression coverage for unchanged copies, deferred modification detection, lazy rename recovery, and optional provenance scans.
Pass the display type into the turn so the row is born typed, and
recognize the note's fixed prefix in _history_to_messages so the
untyped rows already on disk stop painting as user bubbles.
The auto-continue recovery note was typed only after run_conversation
returned, so its row sat untyped for the whole turn — and permanently
when the continuation was itself killed, which is the case it exists
for. persist_user_display_kind stamps the type on the live message
before the crash persist writes it, in the same insert as the content.
The flush also carries display_metadata through, which it was dropping.
Aggregator providers (nous, openrouter) serve dozens of models across many
labs. Add a featured=True enricher to build_models_payload that attaches a
featured_models shortlist to each row: within every lab keep the newest
_FEATURED_PER_LAB models by models.dev release_date, ranked among that row's
own models — never against the current date, so the choice is stable as
models age. Same-date ties fall back to curated list order. Single-lab
providers get an empty list. Derived live from the models.dev catalog already
loaded on this path; no hand-maintained allowlist.
A slash-skill invocation is persisted expanded — activation note plus the
entire skill body. _history_to_messages is the single display projection
every surface reads, so that payload rendered as a chat bubble anywhere a
session was resumed.
Project it here onto the invocation the user typed, and tag skill/bundle
dispatches with the same string so the live send matches. Rewind and
regenerate replay from what the transcript shows, so re-expand the
invocation server-side before running the turn: the replayed prompt is
identical to the original and no client ever holds the body.
The desktop's queue promises "run this AFTER the current turn", but the
promise broke on a race the user can't see: a drain that fired when the
client observed idle while the server was still unwinding the turn landed
in _handle_busy_submit, which applied busy_input_mode — redirecting or
interrupting the live turn with text the user explicitly queued. That's
why force-sending the queue felt like a dice roll: the same gesture
steered, interrupted, or queued depending on a millisecond settle race.
prompt.submit now carries queued:true on every fromQueue drain (composer
auto-drain, send-now, background drain), and the gateway's busy path
honors it by forcing queue semantics — never steer, never redirect,
never interrupt. Lose the race and the text simply waits its turn, which
is what queueing meant all along.
Covered on both sides: a gateway test proving a queued drain cannot touch
redirect/steer/interrupt on the live agent, and the desktop drain tests
now assert the flag rides every prompt.submit shape (direct, background,
resume-retry).
A mid-stream steer persists an interrupted-turn checkpoint so the model knows
its reply was cut off. That scaffolding — "[This response was interrupted by a
user correction.]" and the "Visible response before the interruption:" header —
was written straight into message content, so every reload painted the raw
machinery as an assistant bubble (and merged it into the preceding tool-call
bubble). Steered transcripts became unreadable.
Reuse the existing display/replay split instead of inventing new surface:
- Carry the scaffolded form in the server-only api_content sidecar (the exact
bytes replayed to the provider), keep content the user's/agent's real words.
- When nothing reached the screen there is no clean form, so mark the row
display_kind=hidden — replayed to the model, dropped by every transcript
surface, exactly like compaction-reference rows.
- Honor display_kind=hidden in the gateway's _history_to_messages projection
(it only sniffed the [System: convention), so the checkpoint can't leak
through the live/resume path to the TUI/CLI either.
The model still receives the full interrupted context on the wire; the
transcript shows the partial reply and the user's correction.
Figma's mcp.figma.com register endpoint is a client_name allowlist
(Claude Code / Codex succeed; Hermes Agent 403s) and returns a client
secret while advertising auth_method=none, then requires the secret on
token exchange. Auto-set client_name + client_secret_post for Figma
hosts, pass oauth cfg through login/add paths, force interactive OAuth
for hermes mcp login from non-TTY desktop shells, and ship a catalog
entry. Proven: hermes mcp login figma → 26 tools.
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).
