`hermes acp --setup-browser` installs agent-browser into the Hermes-managed
node prefix (~/.hermes/node/bin/agent-browser), which isn't necessarily on
PATH. doctor only checked PROJECT_ROOT/node_modules and PATH (shutil.which),
so it false-negatived with "agent-browser not installed" even though the
binary was present and runnable. Mirror dep_ensure._has_hermes_agent_browser()
by also checking HERMES_HOME/node/bin and the legacy
HERMES_HOME/node_modules/.bin path, each gated by agent_browser_runnable().
Tested with tests/hermes_cli/test_doctor.py (added positive + not-runnable
cases) and pytest tests/hermes_cli/test_doctor.py -q (66 passed).
The exact kwargs snapshot broke when VAD hardening added keys — the
test's real contract is 'null stt.local: must not crash or force
language/prompt'. Baseline kwargs are pinned by the dedicated suite.
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.
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.
Streamed response text carried a 4-space _STREAM_PAD indent and the
final-response Rich Panel used padding=(1, 4), so every line selected
out of the terminal came with leading whitespace. Both now render
flush-left (pad empty, panel padding=(1, 0)); the table-realignment
width budgets were widened to match.
/copy now writes the ORIGINAL message text through native clipboard
tools (pbcopy / PowerShell Set-Clipboard via base64 / wl-copy / xclip /
xsel — same fallback chain as the TUI's writeClipboardText), falling
back to OSC 52 only when no native backend succeeds. This is the
TUI-equivalent answer to soft-wrap mangling: the clipboard gets the raw
text, not the rendered layout.
_has_agent_browser()'s new managed-Node rung calls
shutil.which('agent-browser', path=...); tests that monkeypatch
shutil.which globally with 1-arg lambdas raised TypeError when their
code path reached the browser readiness probe (test_post_setup_gating,
test_setup_model_provider).
Three GUI Capabilities-tab defects reported on Windows:
1. Browser rows stuck on 'Setup required' after a successful setup run.
Root causes, all in the readiness probe (not the installer):
- _has_agent_browser() never searched the Hermes-managed Node dir
(%LOCALAPPDATA%/hermes/node / $HERMES_HOME/node/bin) where the
Windows install lands, and probed node_modules/.bin/agent-browser
as the extensionless POSIX shim, which fails exec on Windows
(WinError 193) — now resolved via PATHEXT-aware shutil.which
against both rungs, mirroring _find_agent_browser().
- Cloud rows (Nous Subscription Browser Use, Browserbase, Browser
Use, Firecrawl) declared post_setup: agent_browser, whose
readiness gate requires a LOCAL Chromium build the cloud never
uses — switched to the cloud-scoped 'browserbase' hook (CLI-only).
- _agent_browser_installed() could read browser_tool's stale cached
'Chromium missing' result from before the install ran in the
spawned post-setup process — cache now dropped before probing so
the pill flips to Ready right after a successful run.
2. No way to tell which backend is active, and clicking a row to read
its details silently rewrote config. Row click now only
expands/collapses; activation is an explicit 'Use this backend'
button, the active row carries an 'Active' pill, and the expanded
active row says 'This is your active backend'.
3. OpenAI TTS showed one model and one voice. The options were always
defined but rendered through a native <datalist>, which filters by
the field's current value — a field already set to a valid option
suggested only itself. Replaced with a real combobox (Input +
dropdown) that lists every option, and voice suggestions now track
the selected model per the OpenAI TTS docs: tts-1/tts-1-hd = 9
voices, gpt-4o-mini-tts = 13 (adds ballad, verse, marin, cedar).
The Skills Hub 'Marketplace' tab showed a single useless entry: Anthropic
changed .claude-plugin/marketplace.json to bundle-shaped plugins whose
source is './', so all plugins collapsed to one identifier pointing at the
repo root, and the second marketplace repo (aiskillstore/marketplace) is
gone (404). Everything in anthropics/skills is already surfaced by the
GitHub tap as the Anthropic tab, making this source fully redundant.
Removes ClaudeMarketplaceSource and all wiring: source router, index
builder (crawl + floors + sort order + rate-limit messaging), extract
labels/install/URL mapping, hub UI tab, web server labels, CLI limits,
docs (en + zh), the legacy index-cache snapshot, and test fixtures.
Stale skills-index entries with source 'claude-marketplace' still install
fine: HermesIndexSource fetches via resolved GitHub paths generically.
Replace 20 lines of manual os.open/O_EXCL/fdopen/fsync/os.replace with the
existing atomic_json_write() from utils.py, which is already used by 6+
modules and handles temp-file creation, fsync, atomic replace, mode
control, and owner preservation. The only novel helper (_fsync_directory)
is retained — atomic_json_write does not do directory fsync.
Update test_flush_write_failure_leaves_no_recovery_file to monkeypatch
utils.os.replace (the new call path) instead of gateway.shutdown_flush.os.replace.
test_default_bind_serves_both_families asserts an IPv6 listening socket
exists, which is the environment's capability, not the adapter's — CI
runners with IPv6 disabled would flake. Probe an actual ::1 bind (not
just socket.has_ipv6, a compile-time constant) and skip cleanly when
the host has no usable IPv6 stack.
The LINE adapter's webhook server defaulted to host="0.0.0.0", which
binds IPv4 ONLY. On IPv6-only private networks — notably Fly.io 6PN,
where the hosted edge router reverse-proxies LINE ingest to
<app>.internal:8646 over an fdaa: IPv6 address — nothing is listening
on the dialed address: connection refused → customer-visible 502 when
LINE's console verifies the webhook (NS-603).
This is the same bug the generic webhook adapter fixed in d542894ad;
the LINE adapter was never updated to match. Fix mirrors that commit:
- DEFAULT_HOST = None → asyncio/aiohttp create_server binds one socket
per address family (v4 + v6), regardless of the bindv6only sysctl.
"::" is NOT a valid substitute — Fly machines set bindv6only=1, so
it would yield an IPv6-only socket and break IPv4 loopback probes.
- Empty-string host collapses to None; LINE_HOST/extra.host still pin
a specific bind address.
- reuse_address=False scoped to macOS only (BSD wildcard-socket
traffic-splitting footgun), mirroring 9420ad946.
- The three outbound-media guards compared webhook_host == "0.0.0.0"
by string equality; extracted to _missing_public_url() which treats
None/0.0.0.0/::/"" as "no fetchable hostname" so the LINE_PUBLIC_URL
requirement still fires under the new default. _media_url() falls
back to 127.0.0.1 instead of interpolating 'None' into URLs.
Tests: dual-stack default decision table (default None, empty→None,
pinned preserved, LINE_HOST override), behavioural both-families bind
proof via runner.addresses, and the media public-URL guard matrix.
88 passed, ruff clean.
Companion router fix (hermes-agent-router) routes /webhooks/line and
/line/webhook|/line/media to :8646 over 6PN; both are needed for
end-to-end hosted LINE delivery.
Fixes NS-603
show_status now reads get_nous_auth_status_local(); the test_status.py
mocks still patched the old live-resolve entry point, so the patched
dict was never consumed (CI slice 8/8 red on the salvage PR).
Follow-up widening of the /api/status fix: add get_nous_auth_status_local(),
a refresh-free auth-store snapshot (local invoke-JWT decode only), and use it
on the read-only display surfaces that previously called
get_nous_auth_status() -> resolve_nous_runtime_credentials() -> live OAuth
refresh POST:
- hermes_cli/status.py (hermes status auth-provider panel)
- hermes_cli/doctor.py (hermes doctor auth-provider checks)
- hermes_cli/portal_cli.py (hermes portal status display)
- hermes_cli/web_server.py /api/portal endpoint and the accounts-tab
provider card dispatcher (_resolve_provider_status nous branch)
Action paths (login flows, portal operations needing a live credential)
keep using get_nous_auth_status(). Part of NS-592.
resolve_sidecar_dir() probes the filesystem (touch/unlink) and can mirror
sidecar files to HERMES_HOME. Doing that as a module-import side effect
meant plugin discovery, `hermes --help`, and test collection all paid a
filesystem probe (and possibly a mirror copy) just for importing the
photon adapter or CLI.
Convert _SIDECAR_DIR/_NPM_ERROR_LOG in adapter.py and cli.py to lazy
cached accessors (_sidecar_dir()/_npm_error_log()); resolution now
happens on first actual use. Existing tests that monkeypatch the
_SIDECAR_DIR module global keep working — the accessors honor a
non-None value. Adds a regression test proving import performs no
resolution.
The Photon iMessage sidecar needs node_modules under
plugins/platforms/photon/sidecar/, but hosted/managed images keep the
whole install tree under an immutable /opt/hermes — every install and
self-heal path (setup CLI, stale-deps reinstall, cold install) died on
EROFS, and hosted users have no shell to work around it.
Three-layer fix, mirroring the WhatsApp bridge resolver pattern:
1. Bake the deps into the image. The Dockerfile now runs npm ci for the
sidecar in the layer-cached dependency stage (deterministic installs
from the committed lockfile; the postinstall spectrum-ts patch runs
at build time). Hosted happy path needs no runtime install at all.
2. New sidecar_paths.resolve_sidecar_dir() decides where the sidecar
runs from: PHOTON_SIDECAR_DIR override > writable source dir (dev
installs, unchanged) > read-only dir with baked fresh deps (managed
image) > mirror to $HERMES_HOME/photon/sidecar (writable data
volume) when deps are missing or stale in a read-only tree. The
mirror refreshes changed source files on image updates while
keeping node_modules, so the existing lockfile-staleness self-heal
works there.
3. connect() can now cold-install: _start_sidecar() runs the bounded
npm ci bootstrap when node_modules is missing instead of raising
immediately, and check_requirements() reports available when a
self-install is possible (npm present + writable resolved dir) so
the gateway actually creates the adapter on hosted instances. A
failed bootstrap still raises the actionable error, which connect()
surfaces as the retryable SIDECAR_FAILED fatal state on the
dashboard.
Tests: resolver decision table (env override, in-place, mirror,
refresh, fail-open), cold-install lifecycle paths, and a Dockerfile
contract test guarding the baked-deps + no-chown invariants.
Fixes NS-606.
The ownership guard let a sentinel with pid=None pass as self-owned, so
an exiting life could clobber evidence of unknown provenance with a
clean-exit claim. Tighten: only rewrite when the sentinel pid matches
os.getpid() exactly; pid=None (or malformed) is left untouched. Adds
tests for the pid=None no-op and the own-pid rewrite paths.
Hosted agents that die uncleanly (kernel OOM kill, SIGKILL, whole-VM
death) leave no trace: shutdown_forensics only covers graceful signals,
gateway-exit-diag.log only covers exit paths that actually run, and the
VM reboot wipes dmesg before anyone can capture it. NS-608 (BlueAtlas
hourly crash cycle, July 12-15) took days of manual log correlation to
classify because nothing recorded 'the previous life ended violently'.
Add gateway/lifecycle_ledger.py — a sentinel state machine persisted to
<HERMES_HOME>/state/gateway.lifecycle.json:
- start_gateway() claims the sentinel (phase=running) right after the
PID-file/runtime-lock claim, and reports any prior life that never
reached an exit path as gateway.previous_unclean_exit in
gateway-exit-diag.log + a WARNING log line.
- Every exit funnel marks the sentinel exited with a reason:
_exit_after_graceful_shutdown (graceful_shutdown), the shutdown
watchdog (shutdown_watchdog), and the loop-liveness watchdog
(loop_liveness_watchdog).
- Ownership-guarded for --replace takeovers: a live matching owner is
never reported dead, and the old life cannot clobber the
replacement's freshly claimed sentinel on its way out.
The 30s loop heartbeat now embeds a cheap /proc memory sample (own RSS,
MemAvailable, swap used) so every unclean-death report carries a
'memory N seconds before death' snapshot; the detector flags
suspected_oom when the last sample shows <64MiB or <5% available.
container-boot.log lines gain prior_exit=clean|unclean|unknown per
profile, stamping unclean container deaths into the volume-persisted
boot log where support can grep for them.
Tests: tests/gateway/test_lifecycle_ledger.py (16 cases) + 4 new
container-boot annotation cases. Existing watchdog/forensics/boot
suites all green; ruff clean.
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.
_install_dependencies now routes through tools.lazy_deps.install_specs
(NS-605); the force-reinstall test still stubbed
hermes_cli.tools_config._pip_install, so its spy list stayed empty
(CI slice 3/8 red on the salvage PR).
Widen NS-605 to the two remaining direct-install sites in the hindsight
plugin, which still shelled out to 'uv pip install --python
sys.executable' and therefore failed (EROFS/EACCES) on immutable hosted
images with sealed venvs, and lost packages on redeploy:
- post_setup dependency install (~L835): install_specs() with ok /
blocked-reason / stderr handling matching honcho/mem0.
- initialize()-time hindsight-client auto-upgrade (~L1240):
install_specs(); blocked installs log the gate reason with the manual
command instead of a raw subprocess error, and init proceeds.
Audited every other memory plugin (supermemory, byterover, holographic,
openviking, retaindb) for direct pip/uv install subprocess calls: none
remain — their deps flow through plugin.yaml pip_dependencies or
lazy_deps.ensure().
Tests: TestClientAutoUpgradeRoutesThroughLazyDeps — upgrade goes through
install_specs with the exact spec (regression guard asserts no
subprocess.run), blocked upgrade is non-fatal and surfaces the gate
reason. Updated TestPostSetupEnvEncoding stubs to the new install path.
Installing a memory provider (Honcho, mem0, hindsight, ...) from the
dashboard Plugins page failed on hosted deployments with a permission
error: the setup endpoint shelled out to
`uv pip install --python sys.executable`, which targets the sealed
read-only venv under /opt/hermes (immutable hosted image, NS-579/#49113).
The correct mechanism already exists: tools/lazy_deps.py redirects
installs to the writable durable target on the data volume
(HERMES_LAZY_INSTALL_TARGET=/opt/data/lazy-packages) when the venv is
sealed (HERMES_DISABLE_LAZY_INSTALLS=1), appends the target to the END
of sys.path (core venv always wins collisions), and constrains shared
deps to core-venv versions. The dashboard installer simply never used
it.
Fix:
- tools/lazy_deps.py: new public install_specs() — installs arbitrary
manifest-declared pip specs through the same environment routing as
ensure(): venv-scoped by default, durable-target on sealed images,
refused with an actionable reason when gated off (config kill switch
or sealed venv without a target — never surfaces raw EROFS/EACCES).
Specs are validated with _spec_is_safe(); post-install it invalidates
import/metadata caches so availability rechecks in the same process
see the new packages without a restart. Never raises.
- hermes_cli/web_server.py: _install_memory_provider_pip_dependencies
now calls install_specs() instead of building its own uv/pip
subprocess. Blocked installs surface the gate reason in the setup
results; the response's status block reflects post-install
availability (stale 'missing deps' state clears immediately).
- hermes_cli/memory_setup.py, plugins/memory/honcho/cli.py,
plugins/memory/mem0/_setup.py: CLI setup wizards routed through
install_specs() too — same sealed-venv failure mode, same fix.
No hosted setup path writes to /opt/hermes anymore; provider discovery
and installation now use the same environment (sys.path activation is
shared with the lazy-install bootstrap in hermes_bootstrap).
Tests:
- tests/tools/test_lazy_deps.py: TestInstallSpecs — gating matrix
(sealed+no-target blocked with immutable-deployment reason, config
kill switch, sealed+target proceeds), spec-safety rejection before
any subprocess, venv-scoped vs --target command display, failure
stderr passthrough, never-raises contract.
- tests/hermes_cli/test_web_server.py: setup endpoint routes pip
through lazy_deps (regression guard asserts no direct 'pip install'
subprocess), blocked-reason surfacing, same-response availability
recheck clears stale missing state.
Fixes NS-605 (Plain T-1111).
speak_text (hermes_cli/voice.py — the TUI/gateway one-shot TTS entry
point) now checks resolve_streaming_provider() first: when the
configured provider has a chunked streamer, the reply is spoken through
the same stream_tts_to_speaker pipeline CLI voice mode uses, so audio
starts on sentence one instead of after whole-file synthesis. No
streamer (edge/piper/etc.) or a streaming failure falls back to the
existing whole-file path unchanged — one dispatcher, zero parallel
streaming implementations.
Refs: #58930
Follow-up integration for the #47588 salvage, aligning the new streamers
with the post-campaign invariants:
- All streaming key lookups go through _resolve_key -> tts_tool.
_resolve_provider_key -> resolve_provider_secret (config > env/.env >
credential pool, profile-scoped) — never bare get_env_value. xAI
resolves via resolve_xai_http_credentials so OAuth users stream too.
- _capped(): every provider's chunk iterator is bounded at 16 MiB per
sentence, mirroring _read_tts_response_bytes' bounded-upstream-body
invariant on the sync paths.
- Tests updated for the resolver contract + new coverage for credential
routing and the cap.
Salvaged from PR #47588 and rebased onto the post-campaign streaming core:
the StreamingTTSProvider ABC/registry and the ElevenLabs/OpenAI streamers
already live on main (tools/tts_streaming.py), so this ports the pieces
main lacked:
- GeminiStreamer: streamGenerateContent?alt=sse -> base64 PCM chunks
(24 kHz mono int16), reusing main's DEFAULT_GEMINI_TTS_* constants.
- XAIStreamer: WebSocket wss://api.x.ai/v1/tts -> binary PCM frames,
async->sync bridged via the _collect_async test seam.
- tts.streaming.provider config knob: pin one streamer, or 'auto' to
walk the priority list elevenlabs -> gemini -> openai -> xai. Unset
keeps the never-swap-the-user's-voice default.
- docs/streaming-tts.md: architecture, capability matrix, how to add
a provider.
- Unit tests for the knob, SSE parsing, and the WS bridge; key-gated
E2E tests (skipped without credentials).
Refs: #47588
Add an opt-in streaming-audio adapter seam to BasePlatformAdapter so
voice-capable gateway platforms (LiveKit, Discord voice, future adapters)
can consume LLM output as streaming PCM audio before the full response
completes, dropping perceived voice latency from ~2-3.5s to ~500-800ms.
Adapter contract (gateway/platforms/base.py):
- AudioFormat dataclass: declared sample_rate, channels, sample_width
- StreamingTTSHandle: opaque handle with audible/aborted flags
- supports_streaming_tts / begin_streaming_tts / write_streaming_tts
/ finish_streaming_tts / abort_streaming_tts
- All default to unsupported/no-op so existing adapters are source-compatible
- Per-turn _streaming_tts_completed_chats set suppresses duplicate whole-file
auto-TTS when streaming succeeded; cleared after turn completion
Gateway consumer (gateway/streaming_tts_consumer.py):
- StreamingTTSConsumer: bridges sync agent deltas to async adapter audio sink
- Uses existing SentenceChunker (no competing parser)
- Thread-safe bounded queue; on_delta never blocks the agent worker thread
- Resolves configured streaming provider via resolve_streaming_provider()
- Serialises clause playback in order; flushes tail on completion
- Pre-audio failure: completed=False (falls back to whole-file TTS)
- Post-audio failure: completed=True, partial=True (no replay from start)
- Abort is idempotent; late chunks silently dropped
- Per-turn state isolated across concurrent chats
Gateway integration (gateway/run.py):
- message_type parameter threaded through _run_agent -> _run_agent_inner
- StreamingTTSConsumer created when voice input + auto-TTS + provider active
- Delta callback teed to both text stream consumer and TTS consumer
- TTS-only delta callback installed when text streaming is off
- finish() called from executor; wait_complete() in async context after
- Barge-in aborts the consumer at all three interrupt detection points
- Runner-level _send_voice_reply suppressed when streaming TTS completed
Tests (tests/gateway/test_streaming_tts_consumer.py):
- 15 focused tests: adapter defaults, lifecycle, ordered chunks,
unsupported/No-streamer fallback, abort idempotency, late-chunk drop,
pre/post-audio failure, concurrent-turn isolation, think-block suppression,
queue backpressure
Does not touch desktop/TUI code or add config flags. Plugin TTS provider
stream() metadata gap (#47896) is explicitly out of scope — built-in
ElevenLabs/OpenAI PCM streamers are the first consumers.
Refs: #60671, #47896
Sibling fix for #65977 — _model_flow_bedrock_api_key used only
get_env_value for AWS_BEARER_TOKEN_BEDROCK, missing pool-backed
keys. Now uses _resolve_api_key_provider_secret like the other
flows.
index.mjs now imports sibling .mjs helpers (send-format, stream-staleness);
the fixture copied only index.mjs so the sidecar died on module resolution
before reaching the health endpoint.
Maintainer rework of #45580 (issue #54036) on top of the contributor's
cherry-pick, which targeted spectrum-ts 3.1.0 while main pins 8.0.0:
Sidecar (primary detection, new):
- stream-staleness.mjs: pure decision rules, executable under node.
* classifyProbeRejection: only a not-found-shaped rejection of the
synthetic-id read counts as a completed round-trip (ALIVE); any other
rejection is INCONCLUSIVE — never alive. The original /probe treated
ANY rejection as alive, which was too loose.
* shouldProbe: probe only after 10+ min of stream silence (configurable
via PHOTON_STREAM_SILENCE_PROBE_MS; <=0 disables) with a cooldown.
* isZombieSuspect: zombie only on silence past threshold AND a
probe-proven live channel. Silence alone NEVER degrades (shared lines
can be quiet for hours); inconclusive probes NEVER degrade (network
may be down — the iterator will throw and the re-subscribe loop
recovers on its own).
- index.mjs: track last inbound-iterator yield (noteInboundYield), run a
30s watchdog tick, and on a confirmed zombie feed markStreamDegraded ->
the existing exit-75 restart path. /healthz gains a stream.staleness
block (silentForMs, threshold, lastProbeOutcome, zombieSuspected).
/probe reworked to strict semantics: 200 only on a proven round-trip,
503 with outcome hung|inconclusive otherwise.
Adapter (second layer, reworked):
- _probe_once returns tri-state alive|hung|inconclusive; only a hung
sidecar HTTP call counts toward the respawn counter — inconclusive
resets nothing and triggers nothing.
- default probe_interval_seconds 60 -> 600 (conservative; avoid restart
storms on quiet lines).
- _monitor_sidecar_health surfaces zombieSuspected from /healthz as a
warning; the fatal UPSTREAM_STREAM_DEGRADED path is unchanged and fires
when the sidecar escalates.
Tests: test_zombie_stream_watchdog.py executes the real node decision
module and drives the adapter against mocked /healthz responses;
test_presence_watchdog.py updated for the tri-state probe.
Also adds contributor mappings for nickkarhan (#53283) and vaibhavjnf
(#45580).
Follow-up to the URL markdown fix: extract the /send builder decision into
sidecar/send-format.mjs and rewrite test_url_send_path.py to execute the real
module under node (format+text in -> chosen builder out) instead of regex-
grepping index.mjs source, which is a banned test pattern in this repo.
spectrum-ts's live-stream consumer (consumeLive in spectrum-ts 3.x) only
reconnects when its inbound async iterator throws or ends. A half-open
("zombie") gRPC socket — where the TCP connection stays ESTABLISHED but the
peer is gone (NAT idle-timeout, network blip, laptop sleep) — makes the
iterator hang forever: no error, no end. The SDK exposes no gRPC keepalive
knob (createClient takes only {address, tls, token}; grpc.keepalive_time_ms
defaults to -1 = pings off), so the inbound stream silently dies and stays
dead until the gateway is restarted. Symptom: the agent's iMessage line goes
"online but deaf" — Photon's cloud-side fallback answers users with "the agent
isn't online right now" and inbound never reaches the gateway.
Fix, entirely in the code we own (no SDK fork):
- Sidecar gains a POST /probe endpoint that drives a cheap unary read
(space.getMessage on a synthetic id) over the SAME gRPC channel the inbound
stream uses. A live channel round-trips in ms (server returns not-found,
which is success for liveness); a zombie hangs. It sends nothing to any user
and creates no chat (space.get is local in shared/dedicated mode; only the
message read touches the wire).
- The adapter runs a presence watchdog: it probes on an interval, skips the
probe when natural inbound traffic already proved liveness within the
window, and after N consecutive failed probes respawns the sidecar — a fresh
Spectrum() re-subscribes the stream and re-registers presence. Successful
probes double as application-level keepalive, helping prevent the zombie
from forming at all. Respawn is lock-guarded against double-spawn and the
watchdog is torn down cleanly on disconnect.
Behavioural settings live in config.yaml (extra), bridged to env per the
.env-is-secrets-only convention:
probe_interval_seconds (60), probe_timeout_seconds (10),
probe_max_failures (3). A non-positive interval disables the watchdog.
Tests: tests/plugins/platforms/photon/test_presence_watchdog.py covers config
resolution, the disable switch, probe alive/dead(500)/timeout/no-client, the
core N-failures->one-respawn detection, success-resets-failures, stop-then-
start respawn ordering, and lock-guarding — all without spawning Node or
hitting the network.
Contributed by Vaibhav Sharma (X: @vabbyshabby).
The `clarify` tool's multiple-choice prompts flattened to a numbered text
list on Photon/iMessage, even though iMessage has a native poll bubble and
spectrum-ts already exposes it via the `poll()` content builder. Two gaps
caused the flattening:
* Outbound: the sidecar only had `/send` (text); there was no way to send
a poll, so the base adapter's numbered-text fallback was used.
* Inbound: `normalizeContent()` handled only text/attachment/voice, so a
poll vote (`poll_option`) was dropped on the floor ("[Photon content
type not handled: poll_option]") and never resolved the clarify.
Fix, end to end:
* Sidecar: import `poll` from spectrum-ts; add a `/send-poll` route
(`space.send(poll(title, ...options))`); serialize inbound `poll_option`
(the vote: chosen title + selected bool) and `poll` content in
`normalizeContent()`.
* Adapter: override `send_clarify` — for choices, send a native poll via
`_sidecar_send_poll` and call `mark_awaiting_text` so the gateway's
existing pending-clarify text-intercept resolves the answer; open-ended
clarifies keep the plain-text path. Inbound `poll_option` selections are
dispatched as a plain-text MessageEvent carrying the chosen option
(deselections / empty votes are dropped). If the poll send fails (an
older sidecar without `/send-poll`, or a send error) it falls back to the
numbered-text clarify, so nothing regresses on a half-upgraded restart.
No new model tool, no new env var, no core change — the capability lives at
the platform edge. The poll vote reuses the existing clarify text-intercept
resolution path, so no new gateway resolution mechanism is introduced.
Tests: tests/plugins/platforms/photon/test_poll_clarify.py — inbound vote ->
choice text, deselection/empty-vote dropped, send_clarify sends a poll +
enables text-capture, open-ended stays text, and poll-failure falls back to
the text list. Full photon suite green.
Contributed by Vaibhav Sharma (X: @vabbyshabby).
Maintainer follow-up to the #51193 salvage:
- _send_with_retry: permanent classes (auth_or_config, target_not_allowed)
now short-circuit BEFORE the unconditional plain-text fallback resend,
including when a retry attempt surfaces one — no more double-sends of
permanently-failing requests.
- sidecar classifySidecarError: new structured code target_not_allowed for
Spectrum's 'Target not allowed for this project' AuthenticationError
(shared/free-tier lines cannot initiate outbound sends to new targets).
Classification applies to every handler sharing the catch-all
serverError path (/send, /send-attachment, /react, /typing, ...).
- _standalone_send now parses the structured error body too (it reads
sidecar responses independently of _sidecar_call) and returns
error_class/retryable alongside the message.
- target_not_allowed maps to a canonical user-facing message in both
paths; raw upstream error text never leaks through the structured code.
Closes the actionable halves of #50971, #51897, #52794.
Photon's Node sidecar intentionally hides raw handler exceptions, but the Python adapter still needs a safe failure class and retryability bit so delivery retries do not collapse into an opaque generic 500.
Constraint: Sidecar responses must not leak raw stack traces or private exception text
Rejected: Retry every internal sidecar error | masks permanent auth/config failures
Confidence: high
Scope-risk: narrow
Directive: Keep sidecar error text generic; extend safe error classes instead of exposing raw SDK failures
Tested: uv run --with pytest-timeout pytest tests/plugins/platforms/photon/test_overflow_recovery.py -q
Tested: uv run --with pytest-timeout pytest tests/plugins/platforms/photon -q
Tested: uv run ruff check plugins/platforms/photon/adapter.py tests/plugins/platforms/photon/test_overflow_recovery.py
Tested: python3 -m py_compile plugins/platforms/photon/adapter.py tests/plugins/platforms/photon/test_overflow_recovery.py
Tested: node --check plugins/platforms/photon/sidecar/index.mjs
Tested: git diff --check
Tested: python3 scripts/check-windows-footguns.py --diff origin/main
Not-tested: Live Photon/Spectrum delivery against a real iMessage account
Related: #50971
`PhotonAdapter._start_sidecar` is `async`, but it ran the Spectrum
mixed-attachment patch script with a bare `subprocess.run(...)`: it spawns
node and *waits* for it, with `timeout=10`. Executed inline that holds the
shared gateway event loop for the whole window, so no other platform's
messages, heartbeats, or sessions are serviced until it returns.
The same function already establishes this exact invariant twenty lines
above, where the stale-dependency reinstall hops to a worker thread:
# Runs off the event loop so a cold install can't freeze every other
# platform's traffic.
if _sidecar_deps_stale():
await asyncio.to_thread(_reinstall_sidecar_deps)
The patch spawn never got the same treatment. It is not startup-only
either — `_start_sidecar` is called from `connect()`, which takes
`is_reconnect`, so an ordinary Photon reconnect (network blip, sidecar
death) re-runs it and stalls a live gateway that is actively serving
Discord/Telegram/Slack traffic.
Dispatch it via `asyncio.to_thread` like its sibling. Same off-the-loop
class as the inbound-image decision (#66688) and the cron-fire verifier.
Adds a regression test asserting the spawn executes on a worker thread
rather than the loop thread.
`_reap_stale_sidecar` is `async`, but it identified the processes holding
the sidecar port with two blocking helpers called inline:
* `_find_listener_pids` -> `subprocess.run(["lsof", ...], timeout=5.0)`
* `_pid_is_sidecar` -> `subprocess.run(["ps", ...], timeout=5.0)`, once
per candidate pid
so the inspection can hold the shared gateway loop for 5 + 5·N seconds
while nothing else on it is serviced. It only runs once the /healthz probe
finds something already listening — the orphaned-sidecar recovery path —
and `_reap_stale_sidecar` is awaited from `_start_sidecar`, which runs on
every reconnect (`connect(is_reconnect=True)`). The stall therefore lands
on a live gateway that is still serving every other platform, right when a
crashed sidecar has already left an orphan behind.
Move the whole inspection to one `asyncio.to_thread` hop (one hop rather
than N+1 round trips). The reaping semantics are untouched: SIGTERM for
verified orphans, SIGKILL escalation, and both foreign-listener
RuntimeErrors behave exactly as before.
Same off-the-loop class as the inbound-image decision (#66688) and the
cron-fire verifier.
Adds a regression test asserting both the lsof lookup and the per-pid ps
check execute on a worker thread rather than the loop thread.
Fixes#73159.
When the Photon sidecar exits unexpectedly, _supervise_sidecar()
(running as self._sidecar_supervisor_task) correctly detects
SIDECAR_CRASHED and calls self._notify_fatal_error(). The Gateway's
fatal-error handler answers that by calling adapter.disconnect(),
which calls _stop_sidecar() -- from INSIDE the very task that's
currently executing this whole chain.
_stop_sidecar()'s cleanup unconditionally cancelled
self._sidecar_supervisor_task. Cancelling the currently-running task
raises CancelledError at its own next await point (inside
_notify_fatal_error() or _stop_sidecar() itself). Since
asyncio.CancelledError inherits from BaseException (not Exception),
the Gateway's `except Exception` guards around the fatal-error handler
don't catch it -- the handler aborts before ever reaching the
"queue platform for background reconnection" step. Photon then stays
permanently in `retrying` state until the whole process is manually
restarted, even though detection worked correctly and the underlying
transient upstream outage had long since recovered.
Fix: in _stop_sidecar()'s cleanup, check whether
self._sidecar_supervisor_task is asyncio.current_task() before
cancelling it. A task cannot legally cancel itself in any useful way
anyway (the cancellation only takes effect at its own next await,
which is exactly the corruption described above) -- when we're inside
the supervisor's own call stack, it's already in the process of
finishing on its own once _notify_fatal_error() returns, so skip the
cancel and just clear the reference.
This mirrors an existing precedent in the same file: disconnect()
already guards its OTHER task-cancellation (self._sidecar_health_task)
the same way (`if task is not asyncio.current_task()`), just not the
supervisor task in _stop_sidecar().
Per the issue's own note that existing tests mock out
_notify_fatal_error() entirely (so this integration chain was never
exercised), added two tests that drive the REAL chain: one runs
_supervise_sidecar() as an actual asyncio task with a real
_notify_fatal_error() that calls the real disconnect() -> _stop_sidecar(),
confirming the task completes without CancelledError and that the
post-disconnect reconnect-queue step actually executes; a second
confirms the OTHER call path (external cleanup, a different task)
still correctly cancels a running supervisor exactly as before.
Reverting only the adapter.py fix (keeping the new test) reproduces
the exact CancelledError from the bug report, confirming this is a
genuine regression test.
2 new tests pass; 113/113 in the full tests/plugins/platforms/photon/
directory (no regression to sidecar lifecycle, overflow recovery, or
health-monitoring behavior).
Follow-up to #69112. That PR hardened the shared gateway dispatch path in
gateway/run.py against the *caller* being cancelled. A second, self-referential
cancellation specific to PhotonAdapter sits one layer underneath it and survived
that fix.
PhotonAdapter is the only platform adapter that awaits _notify_fatal_error()
inline, on the same task that detected the fault. Both _monitor_sidecar_health
and _supervise_sidecar run as self._sidecar_health_task /
self._sidecar_supervisor_task, and the notification routes into
GatewayRunner._handle_adapter_fatal_error_impl, which tears the adapter down via
_safe_adapter_disconnect -> disconnect(). disconnect() then cancels
self._sidecar_health_task and awaits it -- which, when the health task is what
raised the notification, means disconnect() cancels its own caller several
plain-await frames up.
disconnect()'s `task is not asyncio.current_task()` guard does not catch this.
The current task where that guard evaluates is the wrapper
_await_adapter_cleanup_with_timeout creates around disconnect() via
asyncio.ensure_future, not the health task further up the chain, so the guard
passes and the cancel lands.
CancelledError stopped subclassing Exception in Python 3.8, so the
`except Exception` that wrapped the inline notify call never saw it. The health
task died silently mid-handoff: no log line, no "exception never retrieved"
warning (cancellation is normal asyncio), and no retry. The platform stayed
stranded until the gateway was restarted by hand.
Fix: dispatch the notification onto a new task, the same pattern
DiscordAdapter._handle_bot_task_done already uses for this reason. disconnect()
can then cancel the health/supervisor task freely without that cancellation
reaching the code still running the handoff, so the handoff always reaches the
reconnect queue. Both Photon fatal call sites are converted: the health-poll
path (observed wedging) and the sidecar-crash path (same shape, not yet
observed). gateway/run.py is untouched.
Observed twice on a self-hosted gateway, ~4h38m and ~52min of silent inbound
outage, both cleared only by a manual restart, both post-dating #69112's merge.
In each case the fatal log line appears and `queued for background reconnection`
never does.
Tests: new tests/plugins/platforms/photon/test_fatal_notify_self_cancel.py
covers the self-cancellation (fails with CancelledError without this change),
that the dispatch does not block its caller, that a failing notification warns
rather than raising, and a source guard against reintroducing either inline
await. Two assertions in test_overflow_recovery.py that drove these coroutines
directly now drain pending tasks before asserting delivery, since the
notification is deliberately no longer awaited inline.
Prepared with agent assistance and reviewed before submission.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
