Simplify-pass folds on the #23254 salvage:
- REUSE (HIGH): append_messages_batch now delegates row serialization to
the pre-existing _insert_message_rows helper (already shared by
replace_messages / archive_and_compact / portability import) instead
of adding a third serialization path (_prepare_message_row +
_MESSAGE_INSERT_SQL are gone). One row-writer for every multi-row
path; the row-ID return was consumed by no production caller, so the
batch returns the inserted count.
- QUALITY (HIGH): the compression-lock + compression-closed admission
guards are extracted into _check_transcript_write_guards, shared by
append_message and append_messages_batch (previously duplicated 23
lines that had already needed targeted fixes, #74478). The role-gated
reasoning filtering is no longer duplicated in run_agent.py — it
lives at its one site inside _insert_message_rows.
- EFFICIENCY (MEDIUM, measured): unbounded seed copies hold one BEGIN
IMMEDIATE for seconds (10k rows ~= 2.4s; FTS triggers dominate) and
monopolize the in-process write lock. append_messages_batch grows a
chunk_rows param; all seed/copy call sites use chunk_rows=500. Same
recovery semantics as the old per-row loops, bounded lock holds.
- REUSE (MEDIUM): the two remaining per-row branch-copy loops found by
the pass (gateway/slash_commands.py /branch, hermes_cli
cli_commands_mixin.py branch) are converted to chunked batches too
(AsyncSessionDB's generic to_thread forwarder covers the async site).
Turn-flush benchmark unchanged after the refactor: 2.43 -> 0.87 ms
median per 5-message flush (64% faster).
Sibling sites of the per-message flush pattern: both branch-seed
paths (session.branch in methods_session.py and the lazy seed persist
in server.py) copied the parent history row-by-row -- one transaction
per row, and a branch seed can be hundreds of rows. Route both through
SessionDB.append_messages_batch. The server.py path also gains real
atomicity: _branch_seed_persisted assumed every row landed, which the
per-row loop could not guarantee.
Salvage of #69926: omit_messages support ported from the PR's
tui_gateway/server.py base onto the post-split methods_session.py
layout. When a Desktop client passes omit_messages=true on
session.resume / session.activate, the RPC returns messages: [] with
messages_omitted: true and an accurate message_count, skipping the
potentially multi-megabyte compression-lineage serialization over the
WebSocket; Desktop hydrates the transcript via the authenticated REST
route in parallel.
The PR's bundled cron-outputs endpoint and codex quiet-timeout bump
were dropped from this salvage as unrelated (invited back separately).
The original PR #77274 used positional slicing (current_history[len(history):])
to detect the model-switch-only mutation. But _append_model_switch_marker
strips prior markers in-place before appending the new one, so when a prior
marker existed (every switch after the first in a session), the net length
delta is zero and the slice produces an empty list — the merge path is dead
code for the common case.
Replace with a content-based diff: strip markers from both the turn-start
snapshot and the current history, then check that the non-marker content is
identical. This correctly handles the strip-and-replace behavior.
Also guard against auto-compression making result["messages"] shorter than
the turn-start history — use the full result as the base when that happens.
Added test covering both no-prior-marker and prior-marker cases.
When a model switch occurs mid-turn, `_append_model_switch_marker()`
appends a marker to session history and increments `history_version`.
The turn completion guard then sees `current_version != history_version`
and discards all agent output — producing empty assistant messages in
the session DB.
Detect when the only history mutation during the turn was one or more
model-switch markers. In that case, merge the agent's new messages
into the current history (which now contains the marker) instead of
discarding them. Genuine desyncs (undo/compress/retry) still surface
the warning as before.
Fixes#76870
Deferred model switches append a marker and bump history_version at
turn start; the dispatcher was snapshotting history before that
mutation, so the version-mismatch guard rejected the turn's own
result as a stale/concurrent write. Move the snapshot to after
_apply_pending_model_switch/_sync_agent_model_with_config, under
history_lock, so the turn's own preparatory mutation is included in
its baseline while the anti-stale guard still catches real external
writes.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Unify @null-runner's selector approach with the merged #77019 keepalive:
the reapers now match live delegations by origin UI sid AND (when the TUI
owns the durable lifecycle, never for gateway-viewer tabs, #60609) by the
durable session_key, so a delegation dispatched from an earlier tab of
the same resumed session still keeps it alive. has_live_for_session also
counts 'stalling' to match the #77019 live-state set. Explicit teardown
(_finalize_session) interrupt semantics unchanged.
Replace the process-global HERMES_CRON_SESSION env var with a per-session
ContextVar so a cron tick in the gateway process cannot leak into unrelated
live gateway/API/TUI turns. The cron scheduler now sets the ContextVar
inside the job's try/finally scope and resets it on cleanup. Gateway, API
server, ACP adapter, and TUI gateway all pass cron_session='' to explicitly
mark their sessions as non-cron, masking any stale process env.
Co-authored-by: hinablue <hinablue@gmail.com>
Closes#37968
Simplify-pass follow-up on the #66355 salvage:
1. _config_settings runs on EVERY trim attempt (before the cooldown
check) and only reads — swap load_config for load_config_readonly.
Deep-copying the whole config per attempt generates exactly the
allocator garbage this module exists to release. Tests re-seamed.
2. Trim-failure logs demoted warning->debug at all 3 periodic sites
(gateway housekeeping, idle reaper, slash worker): sibling failure
branches in the same loops log at debug, and a persistent failure
(e.g. broken import after a partial update) would otherwise warn
every 60s forever.
3. The frame-inspection test now asserts the expected locals exist
before reading them — a rename in _run_prompt_submit fails the test
loudly instead of vacuously passing on None.
Add config-driven glibc malloc_trim for long-lived Hermes processes:
- hermes_cli/mem_trim.py: trim_memory() with configurable cooldown,
RSS snapshot telemetry, and forced-trim INFO logging
- gateway/run.py: periodic trim in gateway housekeeping loop
- tui_gateway/server.py: trim in idle reaper (~every 5 min)
- tui_gateway/slash_worker.py: trim on turn boundary
- run_agent.py: force trim on agent close
- hermes_cli/config.py: context.memory_trim config section
(enabled, cooldown_seconds, log_every_n, info_log_min_delta_mb)
CSA tier-4 reviewed (4 rounds, 0 HIGH/MEDIUM/CRITICAL remaining).
Supersedes PR #63708 + #64591 with enhanced telemetry and gateway/slash_worker coverage.
Builds on #72787's current_root guard (cherry-picked with authorship
preserved). Two further hardenings for #72776:
- require the settled cwd and the workspace to share the same common .git
dir (the shape 'git worktree add' produces), so a git workspace visiting
an UNRELATED repo is a browsing visit, not a re-home (repro'd by
Johnny-xuan in the issue thread);
- never reconcile over an explicitly chosen workspace (explicit_cwd),
while a settle-adopted cwd stays followable via a cwd_from_settle
marker cleared by _set_session_cwd / project switch.
Fixes#72776
`_reconcile_session_cwd_from_terminal()` treats a settled terminal cwd in a
different git working tree than the session's workspace as a relocation. But
when the session's workspace is not itself in a git repo, `_git_repo_root_for_cwd(current)`
is None, so the `landed == _git_repo_root_for_cwd(current)` guard never matches
and the FIRST git directory a tool call steps into hijacks the session: its
cwd/git_repo_root flip to that repo, later tools run from the wrong place, and
the repo's AGENTS.md gets injected into an unrelated conversation.
Only reconcile when both the current and settled cwds resolve to valid, differing
git roots. A non-git workspace visiting a git repo to read a file or run a command
is a browsing visit, not a re-home — matching the docstring's own intent that
`cd`-ing away must not re-home the chat. The intended "follow into a worktree"
behavior is unaffected: that path starts from a git checkout, so current_root is
valid there.
Adds a regression test covering a non-git workspace that touches a git repo.
Each mid-session `/model` switch appended a `[System: The active model for
this chat has changed to …]` user-role marker to the live conversation history
and never removed the prior one. N switches left N stale markers, all re-sent
to the provider on every subsequent turn — the issue measured ~80-120 tokens
each (Chinese+English), so 5 MoA-preset switches burned ~400-600 context
tokens per API call, permanently. Only the newest marker is meaningful (it
names the currently-active model); the rest are pure waste. Fixes#65891.
`_append_model_switch_marker` now strips any earlier markers from
`session["history"]` (in place, under the existing history_lock) before
appending the new one, so the live payload carries exactly one marker. This is
the payload re-sent each turn, and it's self-healing across resumes: whatever
markers a history reload brings back, the next switch collapses to one.
A stable `_MODEL_SWITCH_MARKER_PREFIX` constant is shared by the builder and
the `_is_model_switch_marker` predicate so the two can't drift; the marker
string itself is byte-for-byte unchanged. Existing behavior (role='user' per
the single db.append_message persistence) is preserved.
Scope note: this dedups the in-memory history (the per-turn cost the issue
quantifies). The DB still persists one row per switch for audit/resume; I did
NOT soft-archive prior marker rows because `active=0, compacted=0` is the
repo's rewind/undo sentinel (hermes_state.py get_messages docs), so
deactivating a marker naively would make it look rewound — a follow-up that
needs its own archive semantics.
Tests (tests/tui_gateway/test_model_switch_marker_role.py, +5): second switch
replaces (not stacks) the marker; the issue's exact 5-MoA-switch sequence
leaves one marker naming the last model; real conversation turns are preserved
in order; a stale marker sitting between turns is stripped; history_version
increments once per switch. `pytest test_model_switch_marker_role.py` → 13
passed (8 existing unchanged). Authored on Windows.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013b1XyXitAxV7phGmKWigJX
The `/` menu was a flat first-30 slice of the completer's output, and the
completer emits every registry command before the first skill. On a
230-skill install that meant a bare `/` filled all 30 rows with commands
and offered no skill at all, while `/p` cut off inside the alphabetical
skill block — dropping /proving-a-fix-works (471 invocations) and
/pr-update (160) but keeping /pretext (2).
Spend the limit per kind and rank the skill block by recorded usage
(the same .usage.json count Capabilities shows), most-used first and A-Z
within a tie. A bare `/` is browsing, so bundled skills that shipped with
Hermes and were never opened are dropped as noise; a typed query is a
search, and a search that hides a match is broken, so there nothing is
pruned and the ranking only reorders. An argument stage keeps the order
its own command chose.
The /indicator command was registered in COMMAND_REGISTRY, listed in
/help, offered by tab-completion, recommended by the tips system, and
even documented in config.py — but it had no actual handler. Running
/indicator in the CLI produced "Unknown command: indicator".
Add _handle_indicator_command to CLICommandsMixin that:
- Shows the current indicator style when called with no args or "status"
- Validates the requested style against the shared INDICATOR_STYLES
allowlist (ascii | emoji | kaomoji | unicode)
- Persists the choice to display.tui_status_indicator in config.yaml
via the existing save_config_value helper
- Falls back to session-only when config save fails
The indicator-style allowlist is defined once in hermes_constants as
INDICATOR_STYLES + DEFAULT_INDICATOR_STYLE and imported by all three
consumers (CLI handler, command registry, TUI gateway config handler),
preventing drift between the TUI and CLI validation.
Also adds tests/cli/test_indicator_command.py covering dispatch,
validation, persistence, and registry integration.
Signed-off-by: dongjiang <dongjiang1989@126.com>
Disk-full / ENOSPC / SQLITE_FULL on first-message session persist used to be
swallowed as a debug log while prompt.submit still returned streaming, so the
send vanished with no error. Re-raise those failures, return a real RPC error,
and stamp session_persistence_failed turns with error so clients get a terminal
error frame.
The idle-TTL reaper, the LRU cap, and the WS-orphan reap tear down a
live session without the client asking. Nothing was pushed, so a client
kept a runtime id the backend had already forgotten and only found out
by failing a later prompt. Broadcast session.reclaimed with the runtime
id and the reason; client-initiated closes stay silent.
* fix(vision): mount images/ upload dir into sandboxes and permit host read (#69575)
Desktop, clipboard, and PDF uploads land in the flat top-level
HERMES_HOME/images/ dir, but Docker sandboxes only mounted the cache/
subtree and the vision resolver only permitted host reads from the media
caches. So vision_analyze on any desktop-app upload failed under a Docker
backend with "not reachable inside the sandbox".
- Add ("images", "images") to _CACHE_DIRS so the uploads dir is bind-mounted
into sandbox containers through the existing profile-scoped cache-mount and
reverse-mapping mechanism.
- Add home/"images" to _media_cache_roots() so the non-local host-read
allowlist permits reading uploads directly from the host filesystem.
- Cover the mount entry, the container path mapping, and the Docker-mode
resolver read for a profile-scoped upload.
Co-authored-by: JonthanaHanh <92574114+JonthanaHanh@users.noreply.github.com>
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
* fix(tui_gateway): write image uploads under the session's profile home (#69575)
The attach RPCs (image.attach_bytes, clipboard.paste, pdf.attach) wrote
uploads to the gateway's module-cached launch home via _hermes_home/"images".
Those RPCs run before prompt.submit installs the session's profile HERMES_HOME
override, so in a multi-profile / root-gateway deployment the file landed in
the launch home while the sandbox mount and the vision host-read allowlist
both resolve the session profile's images/ at run time — the agent could
never see the upload it was handed.
Add _session_images_dir(session), which anchors the write on the session's
stored profile_home when present (matching the mount/read scope) and falls
back to the launch home otherwise. Route both write sites through it, keeping
per-profile isolation.
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
---------
Co-authored-by: JonthanaHanh <92574114+JonthanaHanh@users.noreply.github.com>
Co-authored-by: webtecnica <75556242+webtecnica@users.noreply.github.com>
A model picked mid-turn is applied at the next turn start, but the end-of-turn
session.info (_emit_settled_session_info) reads the still-live agent — the OLD
model — and clobbered the optimistic paint, so the UI showed the new model, then
snapped back to the old one when the turn ended, then switched for real on the
next turn. Report the queued pick's model/provider in _session_info while a
switch is pending (it IS the model the next turn runs), so the display stays on
the user's choice through the settle. Cleared once the switch applies.
* feat(tui): Ctrl+O opens the model picker without clearing your draft
Reaching the model picker meant typing /model, which forces you to wipe
whatever you'd already drafted. Bind Ctrl+O to open the same picker overlay
directly, leaving the composer untouched. Ctrl+O is added to the textInput
pass-through allowlist so the composer doesn't swallow it, mirroring the
existing Ctrl+X session-switcher path.
* feat(tui): apply a mid-turn model switch at the next turn instead of rejecting it
Picking a model while a turn was streaming hit a 4009 'session busy' reject:
switch_model() mutates the agent's model/provider/base_url/client in place and
the worker thread reads those every iteration. Now config.set queues the pick
in session[pending_model_switch] and _apply_pending_model_switch applies it on
the turn thread at the next turn start, before any model call — no race, no
interrupt, no waiting on the client rebuild. The TUI paints the pick optimistically
and notes '(applies next turn)'.
CI caught ACP session restore seeing an unexpected _row_id in restored
history — get_messages_as_conversation feeds more than the desktop, and
changing the default shape broke the strictest consumer. Row ids are now
include_row_ids=True, requested only by the gateway's resume/display
projections; ACP restore, export, and inspection get the transcript in its
historical shape.
Reactions live in the existing messages.display_metadata JSON column (no new
table), with iOS Tapback semantics enforced DB-side: one reaction per author
per message, re-tap retracts, different emoji replaces. The desktop catches up
to the reaction contract five platform adapters already ship.
- SessionDB: set/get_message_reaction, latest_message_row_id (role + offset +
require_text so invisible tool-call-only rows are never targeted),
take_unseen_reactions (announce-exactly-once), get_message_role
- message.react RPC: accepts row_id or newest_role for live messages that
haven't learned their durable id yet
- react_to_message tool: desktop-gated (check_fn), defaults to the user's
latest visible message, messages_back for retroactive reactions
- Model context rides run_message only (beside the speech-interrupted note):
the persisted prompt stays clean, so no [The user reacted …] scaffolding in
transcripts, and no cached prefix ever changes
- Resume projection forwards row_id + reactions; _row_id is stripped from
outgoing API copies next to display_metadata
The poll this page's pairing block rode was retired hours earlier by
f8e07a332, which moved Messaging onto platforms.changed. That signal is the
mtime of gateway_state.json — where the gateway persists connect/disconnect
health — and a new pairing request moves none of it. So on an event-capable
backend a pending row stayed invisible until something unrelated
reconnected, and the count badge with it.
Adds pairing.changed to the change watcher, signed off the pending/approved
ledgers across the global store and every profile's own. _rate_limits.json
is deliberately excluded: it moves on every unauthorized DM, including ones
that produce no new row, so signalling on it would refetch for nothing.
The page now refreshes platforms and pairing on their own signals rather
than one combined call, and the legacy visible-tab poll (older backends)
covers both.
The change watcher (#73673) missed one always-on-while-mounted timer: the
Messaging page polled /api/messaging/platforms every 6s for connection
status. The gateway already persists platform connect/disconnect/health to
gateway_state.json, so watch that file's mtime and broadcast
platforms.changed (floored to 5s — the gateway also rewrites the file for
in-flight-count bookkeeping), route it through live-sync like its
siblings, and refresh the page on the tick. Older backends keep the
legacy visible-tab poll verbatim.
Finishes the always-on poll sweep for #73618.
TUI (tui_gateway/server.py _load_approval_mode): now delegates to
tools.approval._get_approval_mode instead of re-reading config raw via
_load_cfg + _deep_merge(DEFAULT_CONFIG, ...) and normalizing locally.
Behavior fix, not pure refactor: the canonical load_config path applies
managed-scope config overlays and ${VAR} env expansion, plus a legacy
max_turns lift, which the TUI's raw YAML read bypassed — under a managed
config that sets approvals.mode, the TUI previously reported/toggled a
different mode than the approval gate actually enforced. Both surfaces
now agree by construction. Name/signature and the mode-vocabulary clamp
are preserved.
Codex (agent/transports/codex_app_server_session.py): read confirmed the
_decide_exec_approval/_decide_apply_patch_approval paths carry NO
Hermes-side mode/timeout reads — the Hermes resolution already flows in
from agent/codex_runtime.py via tools.approval.is_approval_bypass_active()
(auto_approve_* routing) and via the shared approval-gate callback. So no
code extraction was needed; added docstrings pinning that invariant and a
cross-reference on the protocol-semantic choice mapping
(_approval_choice_to_codex_decision), which intentionally stays local.
Adds tests/tools/test_approval_mode_parity.py: cross-surface invariant
test asserting the core resolver, the TUI path, and the codex bypass
derivation agree for synthetic configs (unset defaults, global mode set,
YAML-bool off, malformed values, whitespace/case), plus a delegation-seam
test proving the TUI has no independent config read left.
Note: gateway/run.py has sibling raw reads but is intentionally untouched
(multiple in-flight PRs); flagged as follow-up.
The disease: ~15 scattered raw yaml.safe_load(config.yaml) reads that
silently miss managed-scope overlay, ${ENV_VAR} expansion, profile-aware
pathing, and root-model normalization. Every new config feature needed an
N-site sweep (incident chain 9cbcc0c9c8 → 732293cf87 → b0e47a98f9 →
1928aa0443). This commit assigns every raw read to an owner and adds a
lint-guard test so the class cannot regrow.
New primitive (additive-only change to hermes_cli/config.py):
read_user_config_raw(path=None) — reads the user file EXACTLY as
written; docstring states it is ONLY legal for write-back round-trips
and raw-file diagnostics. Behavioral reads must use
load_config()/load_config_readonly().
BEHAVIOR FIXES (class-a sites migrated to a canonical loader — these
previously read values that could DIFFER from the effective config):
gateway/run.py _try_resolve_fallback_provider → _load_gateway_runtime_config
keys: fallback_providers/fallback_model (provider, model, base_url,
api_key). Drift fixed: a managed-pinned fallback chain was ignored;
an api_key of "${OPENROUTER_API_KEY}" reached the resolver unexpanded.
gateway/run.py GatewayRunner._load_provider_routing → same loader
key: provider_routing. Drift fixed: managed-pinned routing prefs and
${VAR} templates were ignored.
gateway/run.py GatewayRunner._load_fallback_model → same loader
keys: fallback chain. Same drift as above.
gateway/run.py GatewayRunner._refresh_fallback_model
keeps the raw primitive (its last-known-good-on-parse-failure contract
forbids the fail-open loader, which returns {} on a torn write) but now
applies managed overlay + env expansion inline. Drift fixed: chain
edits under managed scope / env templates were previously frozen out.
tui_gateway/server.py _load_cfg (72 behavioral call sites)
now = raw read + managed overlay (pre-existing) + NEW ${VAR} expansion,
split from a new _load_cfg_raw() write-back primitive. Drift fixed:
e.g. custom_prompt: "hello ${VAR}", agent.system_prompt, model,
api_key/base_url templates reached sessions unexpanded. DEFAULT_CONFIG
is deliberately NOT merged (callers treat missing keys as unset;
`_load_cfg() == {}` sentinels and _save_cfg round-trips depend on it).
tui_gateway/server.py _profile_configured_cwd
keys: terminal.cwd of a NON-launch profile. Drift fixed: managed
overlay + ${VAR} expansion now apply (load_config() would resolve the
wrong profile's home, so the raw primitive + inline pipeline is used).
plugins/platforms/telegram/adapter.py _reload_dm_topics_from_config
→ load_config_readonly(). keys: platforms.telegram.extra.dm_topics.
Drift fixed: managed overlay + profile-aware pathing + expansion.
plugins/memory/holographic _load_plugin_config → load_config_readonly().
keys: plugins.hermes-memory-store.*. Same drift class.
WRITE-BACK ROUND-TRIPS (class-b: stay raw BY DESIGN via read_user_config_raw;
merging defaults/overlay would pollute the saved user file):
gateway/slash_commands.py: model persist x2, _save_gateway_config_key,
memory/skills write_approval toggles
gateway/platforms/yuanbao.py auto-sethome
tui_gateway/server.py _write_config_key + all cfg→_save_cfg blocks
(reasoning show/hide/full/clamp, details_mode[.section], prompt)
→ new _load_cfg_raw()
plugins/memory/holographic save_config
RAW-FILE DIAGNOSTICS + presence-sensitive bridges (class-c: stay raw,
now via the shared primitive with an explanatory comment):
hermes_cli/doctor.py x5 (model validation, stale-root-keys, .env drift,
deprecation sweep, memory-provider probe — the latter two keep their
inline managed overlay where they had one)
gateway/run.py _bridge_max_turns_from_config and the module-level
TERMINAL_*/HERMES_* env bridge (bridging merged defaults would export
all of DEFAULT_CONFIG into the environment; both keep their inline
overlay + expansion)
hermes_cli/send_cmd.py env bridge (same presence-sensitivity)
hermes_cli/gateway.py multiplex-conflict probe (reads the DEFAULT root's
config, not the active profile's — load_config is the wrong owner)
hermes_cli/profiles.py / hermes_cli/web_server.py / tools/wake_word.py
multi-profile reads (load_config targets only the ACTIVE profile home)
cron/jobs.py _resolve_default_model_snapshot and cron/scheduler.py
run_job config read keep their existing inline overlay+expansion but
now share the primitive (their fail-open + last-value semantics and
the deliberate no-defaults merge are preserved exactly).
Failure-semantics audit: every migrated site preserves its exact previous
behavior on missing file ({} / early return) and parse failure (raise into
the caller's existing except, warn, last-known-good, or fail-open) —
read_user_config_raw intentionally mirrors bare open()+safe_load semantics
(raises on parse errors, {} only on FileNotFoundError/non-dict root).
Guard: tests/hermes_cli/test_config_read_guard.py scans the tree for
yaml.safe_load within 6 lines of a 'config.yaml' reference outside an
explicit ALLOWLIST (hermes_cli/config.py, gateway/config.py, gateway/run.py
fallback path, hermes_cli/managed_scope.py which reads the MANAGED file,
gateway/readiness.py parse-health probe) and fails on new offenders.
E2E: tests/hermes_cli/test_config_loader_e2e.py runs a subprocess with a
temp HERMES_HOME (config.yaml containing ${E2E_PROMPT_SUFFIX}) plus a
HERMES_MANAGED_DIR overlay pinning agent.reasoning_effort, asserting
tui _load_cfg resolves "hello world"/"high" while _load_cfg_raw +
_save_cfg round-trip the template and user value verbatim with no
managed/default leakage.
Replaces the half-duplex per-playback barge monitors with ONE listener
that runs for the entire agent turn in continuous voice mode: armed at
utterance-submit, disarmed when the turn is fully done (response + TTS
finished). Fixes Teknium's live report that voice interruption never
works: (a) not while the LLM is generating, (b) not while TTS plays.
Root causes:
- HALF-DUPLEX GAP: the barge monitor only spawned when TTS playback
STARTED (cli.py streaming/whole-file paths, gateway _tts_stream_begin).
During LLM generation there was NO microphone listener at all.
- PLAYBACK DEAFNESS: the monitor calibrated its VAD noise floor WHILE the
speaker was blasting TTS (speaker bleed baked into the floor), then
multiplied it by 8x with a 1s strictly-consecutive block requirement —
normal speech could rarely reach the trigger, and the 2s grace
swallowed early interjections.
New model — tools/voice_mode.full_duplex_listen():
- Pre-playback calibration: quiet-room noise floor established at turn
start and HELD through playback (never recalibrated against bleed).
- Phase-aware trigger: generation = floor x voice.barge_in_threshold_multiplier
(new config, default 3.0, justified by synthetic-frame tests);
playback = additionally clamped to a 1500-RMS minimum so bleed alone
can't trip; 4000-RMS ceiling keeps speech always reachable.
- Windowed-majority detection (>=80% of a 300ms window) instead of the
strictly-consecutive counter that reset on intra-word energy dips.
- Grace on playback ONSET only (voice.barge_in_grace_seconds, default
down 2.0 -> 0.5) — suppresses the onset transient, not the mic.
- Debug diagnostics at every decision point (calibrated floor, per-window
RMS above 50% of trigger, trip/no-trip, grace suppressions) — always
logger.debug, mirrored to stderr under HERMES_VOICE_DEBUG=1.
Phase behavior (CLI cli.py + tui_gateway/server.py, same model):
- generation: speech interrupts the in-flight turn via the SAME seam the
typed/Ctrl+C interrupt uses (agent.interrupt()), cuts any pending TTS
pipeline so the stale reply never plays, and submits the captured
interjection (pre-roll capture, first syllable kept) as the next turn.
- playback: cuts TTS (streaming pipeline stop + fallback speak stop
events + file player) and submits the capture.
- stop phrase honored in BOTH phases: mid-generation 'stop' interrupts
the turn AND ends the voice chat (stop everything).
- one listener instance spans generation -> playback (no re-arm race);
double-arm refused (CLI _voice_fd_active / gateway _fd_listener_active).
Gateway specifics: _arm_full_duplex_listener() at _run_prompt_submit turn
start and inside _tts_stream_begin; _speak_text_with_barge registers its
(stop, done) pair in _fd_speak_pipelines so fallback speaks are cut and
tracked; _tts_stream_barge_in_monitor kept as a shim that arms the new
listener. Desktop renderer owns its own mic path (voice-barge-in.ts) and
is unaffected; if desktop backend-mic mode is used it inherits via the
gateway.
Tests: full_duplex_listen synthetic-RMS suite (speech-over-bleed trips,
bleed alone doesn't, quiet floor held through playback, grace window,
multiplier math 3x vs 8x, windowed-majority dips), CLI listener phase
tests (generation interrupt seam, playback cut, lifecycle spans phases,
double-arm, config forwarding, stop-phrase-mid-generation), gateway
generation-phase interrupt + stop-phrase tests. Generation-interrupt
test sabotage-verified.
Premise check on live main: barge-in machinery EXISTS for the per-turn
STREAMING pipeline only — cli.py chat() arms _voice_barge_in_monitor and
tui_gateway _tts_stream_begin arms _tts_stream_barge_in_monitor. What was
actually broken for spoken interruptions:
1. CLI whole-file fallback (_voice_speak_response_async — used whenever
streaming TTS cannot start: sounddevice missing, requirement probe
fails): NO monitor was ever armed, so talking over the reply did
nothing. Now arms _voice_barge_in_monitor in continuous voice mode.
2. Gateway fallback speak (tts_queue None → speak_text thread) and the
voice.tts RPC (desktop-triggered speech): speak_text ran bare, and
its internal streaming dispatch created a PRIVATE stop event nothing
could reach — uninterruptible even by stop_playback(). New
_speak_text_with_barge() runs the same barge monitor beside the speak
thread; hermes_cli.voice.speak_text/_speak_text_streaming accept an
external stop_event so a barge cuts the streaming pipeline too.
Stop-phrase handling and voice.transcript submission are inherited
from the shared monitor (merged #73933 behavior preserved).
3. False barge during TTS (the reason interruption "worked" then
self-cancelled or fired randomly): salvaged PR #71083 by @beardedeagle
(previous commit, kept authorship) — rolling-window VAD floor, 8x
multiplier, 4000-RMS trigger ceiling, barge_in_grace_seconds (2s)
before the mic opens, min-floor clamp. barge_in_grace_seconds is now
documented in DEFAULT_CONFIG.
Desktop spoken barge (renderer mic via voice-barge-in.ts) already covers
both its live-stream and fallback speech paths — verified, no change.
Long thinking/tool stretches in a voice conversation are dead air — the
user cannot tell whether the agent is alive. New: quiet, repeating soft
bubble blips while the agent works and no speech audio is flowing.
- tools/voice_mode.py: numpy-synthesized blips (no binary assets) — two
alternating low pitches (G4/E4) with pitch glide + smooth attack/decay
envelopes, ~0.8-1.2s randomized spacing, volume = voice.beep_volume * 0.5.
start_thinking_sound(should_play=...) / stop_thinking_sound() daemon-loop
lifecycle; macOS-TCC-safe (sounddevice output gated there → silent skip,
no per-second afplay churn). New mark_audio_output_active()/
is_audio_output_active() ref-count wraps play_audio_file and the
streaming OutputStream sentence writes so "audio is flowing" is accurate.
- Config: voice.thinking_sound (default true) off-switch.
- cli.py: starts when a voice-mode turn begins, per-blip gate skips while
TTS speaks / mic records / barge capture owns the mic; stopped in the
chat() finally on every exit path.
- tui_gateway/server.py: same lifecycle around _run_prompt_submit turns
(voice mode on), gated on is_audio_output_active + continuous capture.
- Desktop: renderer owns voice-conversation audio, so a matching WebAudio
implementation (src/lib/thinking-sound.ts, same envelope/pitches) runs
while conversation status === "thinking"; honors voice.thinking_sound
(via config store) and the shared sound-mute toggle; stops instantly on
speaking/listening/end.
One owner for the wording: voice_stop_hint() in tools/voice_mode.py —
sources the phrase from voice.stop_phrases (first entry) so a custom
phrase renders correctly, and returns "" when the feature is disabled
(stop_phrases: []) so no surface shows a hint.
- CLI: printed in /voice on output (style-matched dim notice).
- TUI: voice.toggle action=on now carries stop_hint; the Ink client
renders it in the "Voice mode enabled" block (older gateways omit
the field — no hint, no crash).
- Desktop: the renderer voice loop never touches tools/voice_mode.py,
so the phrase is read from config (voice.stop_phrases → $voiceStopPhrase
store, seeded in use-hermes-config) and shown as an info toast when a
voice conversation starts. i18n: en/ja/zh/zh-hant/ar.
The continuous-voice no-speech counter (3 strikes -> voice off) counted
every silent capture cycle unconditionally. During a long agent turn
(thinking/tool-calling for minutes) or while TTS is speaking, the user
is CORRECTLY silent — those cycles ended the voice chat under them.
- hermes_cli/voice.py: new set_voice_busy_probe() seam + _voice_activity_held()
(TTS-playing via the existing _tts_playing Event, agent-busy via the
registered probe). Both the continuous-loop strike path and the
force-transcribe single-shot strike path skip counting while held.
Fail-open: a broken probe counts cycles as before.
- tui_gateway/server.py: registers _any_session_running() as the probe
on voice.record start (voice is process-global; any running session holds).
- cli.py: classic CLI strike path skips counting while _agent_running
or TTS playback is in flight.
Stop phrase and barge-in still work during the hold (own paths).
Includes a fixture fix for the #71083 cherry-pick: the fake tools.tts_tool
module needs _load_tts_config (main's tts_streaming imports it).
Replace one-shot VAD calibration with a rolling deque window that
continuously recalibrates the noise floor throughout TTS playback,
preventing false barge-in triggers from stale calibration. Add a
grace period before VAD activates so TTS playback establishes first.
Suppress duplicate text rendering when token streaming is enabled.
Mirror the barge-in and TTS stream stop logic to the TUI gateway path
so both CLI and gateway use the same VAD semantics.
Rolling-window VAD:
- 90th percentile of rolling window (~3s) for noise floor
- 8x multiplier (was 5x) for TTS volume variation headroom
- 4000 RMS trigger ceiling so genuine speech can still trip
- min_floor clamped to SILENCE_RMS_THRESHOLD * 2
- sustained_ms=1000, calibration_ms=800
Barge-in grace period (barge_in_grace_seconds, default 2.0s):
Delays VAD activation so TTS playback establishes before the mic opens.
Duplicate render suppression:
When streaming_enabled, pass display_callback=None to
stream_tts_to_speaker so the token stream is the sole display path.
TUI gateway mirror:
Mirror _tts_stream_stop and _tts_stream_barge_in_monitor changes to
tui_gateway/server.py so both code paths use the same VAD parameters,
grace period, and TTS CUT diagnostic logging.
Profile-scoped session DB and MoA progress events in tui_gateway/server.py
were necessitated by the TTS pipeline changes affecting session state
and event routing.
Normal-exit flag and TTS CUT diagnostic logging at all cut paths.
Regression tests:
- test_quiet_then_loud_playback_does_not_trip
- test_8x_multiplier_absorbs_tts_volume_spikes
- test_trigger_ceiling_lets_genuine_speech_trip
- test_silence_calibration_does_not_false_trip_on_tts
- test_tts_stream_stop_latches_interruption_for_next_turn
- test_tts_stream_stop_after_natural_finish_does_not_latch
- Profile-scoped session DB tests (10 tests)
Saying OR typing a configured stop phrase (voice.stop_phrases, default
"stop") now ends the voice chat everywhere, not just classic CLI PTT:
- hermes_cli/voice.py: new explicit on_stop_phrase callback through
start_continuous/stop_continuous. The force-transcribe path previously
DISCARDED the stop phrase silently — with auto_restart=False the client
re-arms the next capture, so the conversation never ended. Both halt
paths now fire on_stop_phrase (fallback: on_silent_limit for legacy
callers) as user intent, distinct from the no-speech timeout.
- tui_gateway/server.py: voice.record wires on_stop_phrase and emits
voice.transcript {stop_phrase: true} after flipping HERMES_VOICE(_TTS)
off and stopping streaming TTS — same teardown as /voice off. The TTS
barge-in monitor stop-checks its transcript too. prompt.submit consumes
a TYPED bare stop phrase at the server-side choke point when voice mode
is on (returns {voice_stopped: true}, no turn starts).
- ui-tui: voice.transcript {stop_phrase} ends voice mode with a clear
'voice chat ended' notice (distinct from the no-speech-limit message);
submitPrompt releases the busy latch on a consumed voice_stopped reply.
- cli.py: _typed_voice_stop in process_loop — typing a bare stop phrase
while voice mode/continuous is active ends voice mode instead of
sending 'stop' to the agent; typed 'stop' outside voice mode is
unchanged. Voice transcripts skip the check (already stop-checked).
- desktop: interceptsTypedVoiceStop — the composer's onSubmit ends the
live voice conversation (same path as clicking end on the pill) when a
bare stop command is typed with no attachments; renderer-owned loop, so
handled client-side like the existing spoken isVoiceStopCommand.
- tools/voice_mode.py: transcribe_recording never lets the Whisper
hallucination filter swallow a configured stop phrase (e.g. 'bye'
configured as a stop phrase is both a hallucination-blocklist entry and
a stop phrase — stop-phrase check now wins).
Tests: continuous-loop signal (sabotage-verified), force-transcribe stop
signal + legacy fallback, hallucination-filter ordering, typed-stop CLI
unit tests (voice on/off/longer text), prompt.submit typed-stop gateway
tests, TUI vitest for stop_phrase event handling, desktop vitest for the
typed-stop interceptor.
Uvicorn's 16 MiB default drops one-shot base64 remote attachments before
Hermes sees them. Raise ws_max_size to fit the 256 MiB attach reader after
base64 expansion, and bump DESKTOP_BACKEND_CONTRACT to v5 so older remotes
surface skew instead of silent disconnects.
Co-authored-by: Börje <borje@dqsverige.se>
kshitij
joncaldwell90
Daniel Blank
kshitij
JonthanaHanh
joaomarcos
Teknium
null-runner
xrazai
kshitij
Ryder Freeman
PRATHAMESH75
fyzanshaik
hans
brooklyn!
dongjiang
Austin Pickett
Gille
beardedeagle