Browser screenshot results always named the message tool even when policy had
removed it from the model's effective tool surface. Report the staged sanitized
outbound copy and its path without naming another independently gated tool, so
the result remains actionable without becoming hallucination bait. Update the
vision, fallback, and raw-image result tests to protect that contract.
* fix(google-meet): classify gateway CLI fallback with typed guards instead of message matching
Replaces the message-substring classifier in the Meet CLI's
gateway-unavailable local fallback with the structural
isGatewayTransportError/isGatewayClientRequestError guards (seam shared
verbatim with #125458). Decision surface chosen per Meet semantics:
uncoded transport closes (unreachable sockets, now incl. ETIMEDOUT/
ENETUNREACH) fall back; coded closes incl. 1006 deliberately do not,
since a live gateway may still own Meet browser sessions; unknown-method
fallback stays but requires a structural gateway request error. Adds
ECONNRESET to GATEWAY_UNREACHABLE_SOCKET_CODES so connect-time resets
wrap into the typed unreachable error instead of escaping raw.
* test(process): treat Linux zombies as terminated
* fix(openai): clamp realtime barge-in truncation
Clamp OpenAI Realtime barge-in truncation to the audio duration produced for each assistant item. This prevents interruption failures observed live when 3,700 ms of decoded audio was paired with a 3,760 ms transport-clock timestamp.
* fix(openai): propagate PCM sideband audio format
The browser sideband negotiated PCM16, but its bridge defaulted to G.711 for duration accounting. Propagate the negotiated format so truncation uses the actual PCM duration and closes ClawSweeper's in-scope finding.
* fix(voice-call): survive gateway in-process restart and stop CLI dead-ends
The gateway's in-process restart (SIGUSR1 config reload) reuses the cached
plugin registry, so service stop/start run on the same retained voice-call
registration. Generation fencing from #120289 treated that restart as a stale
actor: stop retired the generation forever, the next start silently bailed,
and every voicecall.* RPC answered UNAVAILABLE "runtime generation is
retired" while the webhook never rebound.
- Registrations now hold a replaceable generation: service start after stop
mints a fresh generation, takes over a running slot owned by a retired
predecessor, and reports start failures to service health instead of
silently returning.
- The voicecall CLI classifies gateway failures with typed guards instead of
message substrings: standalone/store fallback only when the gateway is
genuinely absent; reachable-but-failed (request errors, auth, timeout)
exits with actionable text; a standalone webhook port collision explains
that a running Gateway probably owns the port instead of raw EADDRINUSE.
- Plugin SDK gateway-runtime exports structural isGatewayTransportError /
isGatewayClientRequestError guards (+2 documented surface budget).
- Regression coverage: same-registration stop/start restart, retired-owner
takeover, typed CLI fallback classification, and a real token-auth gateway
server routing voicecall.status through callGatewayFromCli.
* refactor(voice-call): split CLI modules and dedupe gateway fallbacks
Collapse the four duplicated gateway-or-runtime command blocks (speak, dtmf,
end, continue fallback) into one generic runGatewayManagerCommand helper —
the continue command owns its legacy-method fallback and operation polling
via a gatewayCall closure, so the helper carries no per-command policy.
Smoke reuses the shared initiateVoiceCall path instead of a bespoke
fallback.
Split the 988-line cli.ts into concept modules (cli-gateway-call,
cli-call-log, cli-command-io) and drop its grandfathered max-lines
suppression plus the now-stale max-lines and assertion-safety baseline
entries (shrink-only ratchet maintenance).
Behavior-frozen: stdout/exit semantics unchanged; net -2 production LOC.
* fix(voice-call): redact gateway URLs in CLI operational errors
ClawSweeper P1: the operational-error formatter interpolated the raw
connectionDetails.url, so a configured gateway URL with userinfo or query
tokens would print credentials into terminal output. Redact the composed
message once with the canonical net-policy redactor (also covers
remote-controlled close-reason text), exported through the plugin SDK
gateway-runtime subpath (+1 documented surface budget). Regression test
covers a credential-bearing URL in both the URL and message fields.
memory_search corpus=all silently swallowed a returned memory-manager
error (missing sqlite, embedding misconfiguration, corrupted index):
the error branch was gated on !shouldQuerySupplements, so the tool
served wiki-only results in a payload with no error/warning field and
recorded no cooldown — a degraded search read as complete, contradicting
the tool description's own disabled-flag contract. The thrown-error path
already surfaced this; only the returned-error path was silent.
Record the cooldown in both cases and, when supplements still serve,
carry a warning naming the unavailable memory corpus in the payload.
* fix(voice-call): expose realtime/streaming stream paths through tailscale serve/funnel
Tailscale serve/funnel now auto-exposes the realtime and streaming WebSocket stream paths when those audio modes are enabled; previously Twilio <Connect><Stream> could not reach wss://<ts-host><streamPath> and realtime calls dropped after ~1s with no log.
Also: CLI voicecall expose mounts/clears stream paths symmetrically, partial tailscale mounts roll back on failure, and a warning now fires when a call's stream WebSocket never arrives within the token TTL. Reported/observed live by Peter Steinberger.
* fix(voice-call): make tailscale stream exposure atomic
Preserve configured public Tailscale prefixes when mapping realtime and streaming routes, and roll back the full route set when any mount fails. The CLI now uses the same transactional setup path and reports failure instead of accepting a partial exposure.
* fix(voice-call): expose Twilio's configured streaming path
Keep realtime routes under the public webhook prefix they advertise, while mounting streaming.streamPath exactly as Twilio emits it. This preserves the canonical public-to-local route list without expanding the provider API outside the scoped repair.
* fix(voice-call): grace realtime stream disconnects
Share reconnect grace by CallSid and stream ID across classic and realtime streams while cleaning realtime bridge resources immediately.
Log terminal call reasons and document Twilio inbound voice and status callback setup.
* test(voice-call): align reconnect grace after rebase
Preserve the newly landed realtime generation and inactivity coverage while updating its terminal expectations for shared delayed finalization.
Remove the redundant replacement cross-product case so the lifecycle suite remains below the max-lines limit.
* fix(codex): bound no-engine continuity projections to half the context window
A degraded or absent context engine sends fresh-thread continuity through
projectContextEngineAssemblyForCodex with the whole-window projection cap
((window - 20k) x 4 chars), so a large uncompacted transcript renders into
a single turn/start input consuming up to 90% of the model context window.
That turn fills its own native thread, the next turn's token fuse rotates
it, and the following fresh thread re-projects the transcript again -
observed as 11 near-window turn inputs on cold threads in one day
(openclaw/openclaw#125254).
Continuity projections now use a dedicated cap that reserves half the
context token budget, so the fresh thread keeps headroom for later turns
and the existing delta-resume path can actually engage. The active-engine
projection path keeps its whole-window cap unchanged.
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
* fix(codex): size continuity projections from real token cost, not the optimistic estimate
The continuity cap reserved half the context window in tokens but converted
that budget to characters with APPROX_RENDERED_CHARS_PER_TOKEN = 4, so at a
258,400-token window it permitted 516,800 chars. A live projection measured
703,134 chars for 226,146 input tokens, meaning that cap really costs about
166k tokens (~64% of the window), not the intended 129.2k.
Codex reports input tokens only after a turn and bounds turn input by
characters, so the projection cannot be sized in verified tokens before it is
sent. Convert the continuity budget at a conservative 3 chars/token instead,
which holds the reserved half in real tokens at the densest ratio observed.
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
* fix(codex): scope the continuity sizing claim to the density it was measured at
The half-window claim was stated as a guarantee, but the 3 chars/token
conversion rests on one observed projection. Input that tokenizes more densely
(CJK, base64, minified code) still exceeds the reserved half, so the constant is
renamed to CONTINUITY_EMPIRICAL_CHARS_PER_TOKEN and its comment says plainly
that it is an empirical floor rather than a bound.
The invariant test is narrowed to the measured density, and a companion test
pins the break-even ratio at 3 chars/token so the limitation is visible in the
suite instead of implied. Choosing between a guaranteed worst-case bound and
this empirical cap is a maintainer-owned tradeoff, left open on the PR.
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
* feat(codex): size continuity projections from the session's observed token density
Each completed Codex turn now records a calibration sample on the thread
binding: prompt chars actually sent vs the provider-reported input token cost
(uncached + cache read + cache write). The no-engine continuity cap converts
its half-window token budget at that observed ratio instead of a fixed
chars-per-token guess, so capChars / ratio stays at the reserved budget for
any content density - CJK, base64, and minified code included. The sample is
captured before startup rotation so a rotated-away thread's density still
sizes the fresh thread's projection; without a sample the empirical 3
chars/token default applies, and degenerate samples clamp to [0.5, 4].
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
* fix(codex): make continuity calibration monotone - samples only tighten the cap
Red-team finding: a loose sample (up to 4 chars/token) followed by denser
content could size the cap past the empirical default, and stale or
non-continuity samples persist on the binding. Clamping the calibrated ratio
at the empirical default makes every such failure mode degrade to the
uncalibrated behavior instead of past it, and the invariant test asserts
monotonicity across poisoned samples directly.
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
* fix(codex): record continuity calibration only from continuity projections
ClawSweeper P2: calibration ran after every successful turn filtered only by
prompt size, so a dense direct or active-engine prompt could persist a sample
whose density later shrinks continuity history it never measured. The
no-engine continuity appliers now mark the prompt state, finalize gates the
sample on that marker, and a cross-mode regression proves a large direct
prompt records nothing.
Related: #125254
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JKtoZgXWnaAH8rmLiSydpN
---------
Co-authored-by: Marvinthebored <262704729+Marvinthebored@users.noreply.github.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Realtime call teardown previously depended on an object-identity guard
that silently skipped ending the call record when bridge instances were
replaced, and had no transport-liveness backstop: a WS close that never
propagated left an answered call running forever. Bindings are now
socket-bound with current-generation terminal ownership, predecessor
audio is retired on successor admission (the overlapping-voices bug),
a 30s media-inactivity watchdog with 2s grace ends calls whose
transport dies silently, and every realtime call end is logged with
its cause.
* feat(agents): unify agent status into a durable progress_card
Replace the write-only update_plan to-do tool and the fragmented plan
rendering with one durable status artifact per session: progress_card
({plan?, markdown?}, replace-on-write, 8 KiB markdown / 50-step caps).
Cards persist in a lazy-additive session_progress_cards table in the
per-agent DB (no schema-version bump), broadcast progressCard.changed,
and render from the store with exactly one live placement per view
(session rail when visible, else the composer-adjacent bar); transcripts
collapse to one-line receipts, and the sidebar hovercard shows other
sessions' cards inline (markdown + <progress>, DOMPurify allowlist, no
iframes). The three stream-derived plan renderers and their dedup
heuristics are deleted.
Codex runs disable the native plan tool per thread
(tools.update_plan.enabled=false) and receive progress_card via the
dynamic-tool bridge; compaction restore now reinjects the card (steps +
bounded markdown). Card writes still emit the legacy plan stream event so
native apps and channels keep working until their per-platform
migrations. Policy names map update_plan -> progress_card; the shipped
tools.updatePlan=false kill switch is honored.
Net -277 production LOC; -480 test LOC.
* test(agents): regenerate Codex prompt snapshots for update_plan thread-config disable
* chore(protocol): allowlist progressCard.changed for native apps pending card migration
* fix(ci): repair progress card integration checks
* fix(codex): canonicalize native progress cards
* test(gateway): reconcile progress card method order
* test(codex): stabilize native approval fixture
* fix(agents): preserve contextual reply correlation
Carry prepared reply policy into embedded and CLI message tools so targetless direct replies retain transport correlation and Reef thread identity.\n\nCloses #125279
* fix(reef): preserve unthreaded reply correlation
Keep reply-only inbound messages unthreaded unless the transport supplies a thread, while still anchoring new exchanges to their initiating envelope.\n\nRefs #125279
Plugins holding a remote URL had no non-deprecated Plugin SDK path to turn it into managed media; saveMediaSource was only reachable through the deprecated media-runtime barrel while docs pointed at media-store. Re-export it from the focused subpath and move the bundled qa-channel plugin off the deprecated barrel. Maintainer decision: full saveMediaSource (local path + HTTP(S)) is the supported media-store contract.
Fixes#125259
Show an explicit waiting acknowledgment when sessions_yield ends an otherwise-silent interactive turn, while keeping private resume context out of channel delivery and preserving existing visible replies.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(agents): report interrupted tool outcomes honestly
Prevent restart recovery from claiming that an interrupted tool call completed when no successful matching result was recorded.
* fix(agents): keep ambiguous recovery restart-safe
Classify failed replay-unsafe tool results at the shared restart-recovery owner so interrupted side effects remain unavailable until their external state is verified.
* fix(agents): distinguish missing tool outcomes
Carry the existing missing_tool_result fact into projected Codex transcripts so restart recovery restricts only genuinely unknown outcomes, while confirmed failures remain retryable.
Peter Steinberger