* refactor(voice): promote realtime voice policy to talk owner
Move the exact-speech protocol (typed consult-call outcome instead of
string scraping), agent-proxy session instructions, and consult/wake-name/
barge-in policy defaults from the Discord plugin into src/talk, exported
additively via openclaw/plugin-sdk/realtime-voice. Wake-name enablement now
keys on a supportsActivationNameGating provider capability instead of a
hardcoded provider id. Follow-up to the voice split (#122479).
* fix(talk): bind exact-speech bypass to retained session state
Per ClawSweeper review on #122668: the protocol marker is untrusted model
tool-call text; only a parsed answer matching a currently retained
exact-speech text may select the privileged replay path. Unretained
marker calls route to a normal consult.
Keep readable native images when another attachment fails resolution.
Suppress only unresolved media facts across persisted and transient runner paths.
Co-authored-by: SunnyShu0925 <shu.zongyu@xydigit.com>
* feat(ai): continue stateful OpenAI SSE turns
Use the prior OpenAI Responses ID for native stateful HTTP/SSE turns, retry missing server state with full history in the same turn, and keep stateless/custom routes unchanged.
Co-authored-by: Markus Hartung <mail@hartmark.se>
* test(ai): assert Responses retry attempt metadata
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Make the process action enum authoritative for model-facing validation and direct executor calls, rejecting malformed actions before session lookup.
Refs #69582.
Punchcard-Session: clear-orchard-timber-c2
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Co-authored-by: adone0 <vladyslav.yavorskyi@outlook.com>
* feat(gateway): recover channels and health promptly after host sleep
A dependency-free thaw detector rides the existing 30s maintenance tick:
when the process resumes after being frozen >=45s beyond cadence (laptop
sleep, VM pause, SIGSTOP), the gateway restarts running channel accounts
(dead sockets otherwise take up to ~35 minutes to notice), refreshes
health/presence, and resets the event-loop histogram so the freeze does
not read as degradation. Admission is rechecked before every recovery
side effect; a suspension beginning mid-recovery re-pends the thaw, and
timed-out channel stops complete their two-call restart in one pass.
The macOS app cooperates: NSWorkspace sleep/wake observers in
GatewayConnectivityCoordinator best-effort prepare a local gateway
suspension before sleep and resume it on wake, never blocking sleep.
The lease is bound to the route that prepared it and always cleared on
wake; route or mode changes across sleep drop it to self-expiry.
Live proof: SIGSTOP 85s on an isolated dev gateway -> 'host thaw
detected: process was frozen ~57683ms', channels restarted, health ok,
eventLoop degraded=false after thaw.
* fix(macos): resume a sleep lease whose prepare response arrives after wake
A prepare completing after didWake previously discarded the lease id,
fencing the gateway until the two-minute expiry after micro-sleeps; the
late response now resumes immediately. Document the conservative
route-token drift tradeoff.
* fix(macos): retry wake resume after refreshing the dead post-sleep transport
After real sleep the WebSocket is usually dead exactly when resume runs;
refresh the endpoint first, then attempt resume up to three times with
bounded delays, clearing the lease only on success or exhaustion. A new
sleep cycle aborts in-flight retries.
* fix(gateway): bound plugin stopAccount so channel stops cannot wedge recovery
stopChannel awaited plugin stopAccount unbounded; a never-settling stop
hung the thaw restart (and health-monitor sweeps) and held the
single-flight recovery guard forever. Race it against the existing
5s stop timeout; the timed-out path flows into the established
recoveryStopTimedOut two-call restart contract. Regression wedges
pre-fix.
* refactor(gateway): move thaw channel restart off ChannelManager and fence mid-pass
restartRunningChannelAccounts is a standalone helper over the public
manager surface with a shouldContinue probe checked before every stop
and start, so a suspension committing while an account stop is awaited
leaves later accounts untouched. Regression covers the mid-pass close.
* fix(gateway): sanitize late writes from an abandoned stopAccount
An abandoned (timed-out) stopAccount can settle after its replacement
started; route its late setStatus writes through the existing
stale-task sanitizer so they cannot repaint or tear down the
replacement. Regression fails pre-fix.
* refactor(infra): extract shared git exec and verified snapshot-copy helpers
Moves the worktrees git wrapper to src/infra/git-exec.ts (with optional
maxOutputBytes for large buffered reads) and the online-backup/sanitize/
VACUUM/verify snapshot step into src/snapshot/openclaw-snapshot-copy.ts so
snapshot backends share one hardened copy path. Behavior-identical moves;
all importers updated.
* feat(snapshot): git-backed versioned SQLite snapshot engine
Deterministic per-table JSONL dumps (PK-ordered, lossless bigint/blob
encoding), verbatim DDL preservation, virtual/shadow-table skipping with
FTS rebuild on restore, secret-table redaction policy, manifest with
per-table row counts and content hashes, and restore verification by
re-serialization. Unchanged data produces no commit.
* feat(backup): recorded runs, freshness surfacing, and scheduled git backups
Every backup attempt is recorded in the previously writer-less backup_runs
table (bounded to 200 rows). openclaw status gains a Backups overview row
and JSON payload; doctor prints an informational hint when no successful
backup is recorded or the newest is stale. New commands: backup git
init/create/log/verify/restore and backup enable/disable, which provision
one idempotent gateway cron job running scheduled git backups.
* fix(state): stop bumping schema_meta.updated_at on unchanged opens
updated_at now records when schema metadata actually changed instead of
when the database was last opened; unconditional bumps dirtied the row on
every open and defeated no-change backup detection.
* docs: document versioned git backups, scheduling, and backup freshness
* fix(backup): satisfy CI ownership checks
* fix(backup): complete CI contract coverage
* fix(backup): complete credential table redaction
* fix(backup): isolate git repository ownership
* fix(backup): persist push degradation
* fix(backup): atomically converge schedules
* fix(status): isolate backup freshness environment
* fix(status): carry scan environment to freshness reads
* fix(backup): harden Git repository ownership
* docs(backup): document Git repository safety
* fix(backup): non-creating outcome log and origin preflight for pushed schedules
Recording a backup outcome never bootstraps an absent state database (a
failed backup on a fresh host would otherwise create a blank DB that a
retry treats as real input), and backup enable --push now requires the
repository to have an origin remote, pointing at backup git init --remote
instead of scheduling permanently degraded pushes.
* refactor(worktrees): use shared git exec helpers
* refactor(worktrees): remove unused git buffer wrapper
* refactor(worktrees): consume buffered git helper
* feat(backup): redact pushed schedules by default
Unattended recurring pushes retain credential-bearing tables durably in
remote Git history, so backup enable --push now defaults to
--exclude-secrets; --include-secrets is the explicit full-fidelity
override (still warned). Local non-push schedules keep full fidelity for
complete restores.
* fix(backup): redact audit HMAC and OAuth pending state; tolerate absent backup_runs
Adds audit_identity_keys (audit HMAC key) and mcp_oauth_pending_authorizations
(live OAuth callback state) to the redaction inventory, and makes read-only
backup freshness treat a same-version database without the additive
backup_runs table as no recorded backups instead of failing before a
writable open converges the schema.
* fix(backup): restrict schedules to local gateways
* fix(snapshot): harden Git restore and redaction
* fix(backup): block pushes of adopted history
* fix(backup): contain commits and pairing secrets
* test(secrets): cover code-mode nested exec store env; document harness boundary
* docs(secrets): warn that store env does not reach external agent harnesses
* feat(gateway): add gateway-host desktop source behind desktop.host lab
Introduce the host as a first-class desktop source so operators can view
the machine OpenClaw runs on, not just cloud-worker environments:
- protocol: desktop.observe / desktop.launch with a discriminated
DesktopSource union (host | environment) plus an additive auth hint;
EnvironmentSummary gains a top-level desktop flag
- config: desktop.host { enabled, port?, passwordFile? }, Labs-gated
- rfb-probe: pure RFB version/security-type parser used to detect an
already-running loopback VNC server and classify its auth
- host-source: attaches to 127.0.0.1:<port>, refuses unauthenticated
(None) and unsupported (VeNCrypt) servers, and refuses ARD with the
supported alternative until the macOS milestone
- host-guidance: per-OS enablement text so no path dead-ends
- doctor + status report host desktop availability and auth type only
worker.desktop.observe/launch stay as delegating aliases with identical
behavior. Also drops the now-unused WorkerDesktopTunnels type export.
Live-verified against macOS Screen Sharing: probe reads RFB 003.889,
returns security types [30,33,36,35], classifies ard-account.
* test(gateway): probe RFB handshakes through the socket boundary
The probe's banner and security-offer parsers were exported solely so
unit tests could call them, which the dead-export gate rejects and which
tests internals rather than behavior. Keep them module-local and drive
the probe through a scripted loopback server instead.
The boundary tests also cover what pure-function vectors could not:
handshakes split across packets, legacy RFB 3.3 single-word security,
server-rejected handshakes, early hangups, and connect timeouts.
* feat(ui): let the Desktop panel view this machine, not just cloud workers
The Desktop panel was gated on a cloud-worker session placement, so an
operator running OpenClaw locally had no way to see the machine hosting
their main session even with a VNC server running on it.
Availability now follows the advertised desktop.observe method plus
operator.admin instead of session placement, and the picker lists every
environment whose summary reports a desktop, with the gateway row shown
as "This machine". Sources are passed to the generic desktop.observe /
desktop.launch RPCs; the app launcher stays worker-only. When a host
attach needs a password the gateway did not supply, the panel prompts and
keeps the value in memory for that connection only.
Adds the hostDesktop Labs toggle for desktop.host.enabled.
* fix(scripts): keep the env-var ratchet usable in shallow checkouts
The env-var budget check resolved its base ref, then hard-failed when
`git merge-base` found no shared ancestor. Shallow clones and grafted
agent checkouts resolve origin/main but truncate the history behind it,
so an advisory growth ratchet took down the whole check:changed gate
with "Could not resolve env-var count merge base for: origin/main".
Only the growth comparison needs a baseline, and the script already has
a no-baseline path. Treat git's exit 1 with empty output (no shared
ancestor) as that case and say so on stderr; a genuine failure still
exits 128 and still throws, and the absolute count-vs-budget check runs
either way.
* test(ui): measure the inline-code chip against its line box
The inline-code spacing test compared the chip's height to a prose text
rect, so it silently measured the monospace font's default line spacing.
That is ~17px on macOS and several px shorter on Linux, so the assertion
passed on CI and failed locally at 4.5 against a 3.75 bound -- after the
bound had already been widened once to chase browser font metrics.
Compare the chip to the paragraph's CSS line box instead, which is what
"the chip must not disrupt the line" actually means and is platform
independent. The horizontal gap stays as-is: it is em-derived padding
plus border, and it is the assertion that catches detached punctuation.
Verified both directions on macOS: the file is fully green, and
restoring the pre-fix 0.15em/0.35em padding still fails the gap
assertion at 5.41.
* feat(gateway): view macOS Screen Sharing from the Desktop panel
Modern macOS only offers ARD account authentication for Screen Sharing,
so the host desktop source refused every Mac. The Gateway now performs
the ARD handshake itself against the loopback server and hands the
browser a plain RFB 003.008 no-auth handshake, so the operator's macOS
account password authenticates the desktop without ever reaching the
browser, the observe result, a URL, or a log.
- rfb-preauth: ARD (type 30) Diffie-Hellman with MD5-derived AES-128-ECB
credentials, and VncAuth (type 2) bit-reversed DES, both under a single
10s negotiation deadline; Apple's RFB 003.889 maps to 3.8
- observe-bridge: runs pre-auth before splicing and starts the view-only
filter at clientInit, since the browser handshake is consumed here;
worker tokens keep the original version start phase
- host-source: attaches ARD, requiring per-observation credentials that
live only in the one-shot observer token and are dropped after use
- doctor: offers an explicitly confirmed sudo launchctl repair when
Screen Sharing is off, and prints the System Settings path otherwise
Live-verified against this Mac's Screen Sharing: the DH exchange and
credential framing are accepted and the server returns SecurityResult.
The VncAuth DES vector is confirmed against OpenSSL independently.
* fix(agents): resolve session-stable reply mode for direct resolver turns
Heartbeat wakes and other direct getReplyFromConfig callers bypass
dispatch, so no session-stable delivery mode reached their CLI binding
facts and prepare computed no messageToolPolicyHash, while dispatched
chat turns hashed the stable mode. resolveCliSessionReuse treats that
one-sided hash as a policy change, hard-invalidating the binding on
every chat<->heartbeat transition and leaving heartbeats to run in
fresh CLI sessions with no conversation history (historyPrompt=none).
Extract the synthetic-turn stable-mode resolution that
prepareAgentCommandExecution already used into
resolveSessionStableReplyMode, and apply it in prepareReplyRunContext
whenever a synthetic turn arrives without dispatch's injected mode, so
every turn kind on a session derives the same binding facts and hash.
Fixes#121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): derive synthetic binding facts from the dispatch policy owner
Consolidate the synthetic-turn stable-mode resolution onto dispatch's own
policy pieces instead of a simplified parallel resolver:
- resolveVisibleRepliesPolicy is extracted from dispatch prepare-context
(configured visible-replies guard + harness default chain) and shared, so
the fallback and dispatch cannot drift on harness defaults, live-vs-entry
chat facts, or ExplicitDeliverRoute.
- The stable fact now applies dispatch's messageToolAvailable downgrade via
the canonical tool-policy resolvers: tool-only delivery with a policy-denied
message tool records automatic, matching dispatch.
- Synthetic turns no longer fall back to their effective turn mode: a
response-tool heartbeat's message_tool_only is per-turn enforcement, not
session policy, and previously bypassed the stable fallback entirely.
Regression coverage: response-tool heartbeat facts match dispatched turns,
and a message-tool-denied config downgrades the synthetic stable fact.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): keep the session-stable reply mode sender-independent
Dispatch's stable-mode downgrade used the turn's sender-aware message-tool
availability, so a sender-scoped message denial hashed the stable policy as
automatic on chat turns while sender-less synthetic turns hashed tool-only —
the same binding reset loop on another axis.
resolveStableMessageToolAvailability becomes the one sender-independent
owner: dispatch passes it as sessionStableMessageToolAvailable for the
stable-mode resolution only (effective per-turn enforcement keeps the
sender-aware verdict), and synthetic binding facts already consume it.
Computed only when the visible-replies candidate is message_tool.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(agents): derive stable reply facts from session surfaces, not wake plumbing
Third-pass review fixes on the #121485 consolidation:
- Strip system-event wake providers ("heartbeat", "cron-event") from the
stable context and resolve origin-less entries as internal, so synthetic
facts take the same internal-channel branch dispatch's live webchat turns
do instead of falling through to harness defaults.
- Fall back to persisted session facts (delivery channel/origin, groupId,
groupChannel/subject, accountId) for the sender-independent availability
stack, so bare-ctx callers like command prepare resolve the same
group/account-scoped policies as dispatched turns; dispatch now passes its
session entry too.
- Collapse the redundant second mode resolution into the availability
downgrade, surface dispatch's injected stable mode separately from
resolvePromptSourceReplyMode so the synthetic slot has one owner, and
reuse resolveTurnModelOverride instead of an inline copy.
- Refresh docs/.generated/plugin-sdk-api-baseline.jsonl: closure hashes only
(17 exports, zero declaration changes, verified) — plugin-sdk reply/meeting
entrypoints transitively bundle the reply pipeline this branch refactors.
Regression: origin-less entry heartbeat resolves internal-automatic facts.
Refs #121485
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* chore: adapt to main module moves and split API baseline
Mechanical rebase refresh onto d3696f8d76: subagent-capabilities moved to
src/agents/subagents/spawn/, and the plugin-sdk API baseline is now
per-entrypoint content hashes (#122082) — regenerated for the reply-pipeline
closure this branch touches.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* refactor(memory-core): split index manager along the inheritance chain
* fix(memory): restore the manager lazy boundary
* refactor(memory-core): delete dead manager plumbing
* chore(lint): ratchet max-lines baseline after memory manager split
* fix(memory): keep the manager out of the runtime facade
* test(memory): cover transitive manager facade imports
Vincent Koc