## What Problem This Solves
Skill Workshop experience review ran as an isolated agent with a re-rendered 60k-character transcript, its own bootstrap, and a trimmed tool surface. Every review was a cold request: no prompt-cache reuse, a large re-serialized trajectory, and a reasoning/tool profile that differed from the foreground turn. Autonomous updates could target any workspace skill, size limits allowed 40 KB skills to grow unchecked (one deployed workspace hit a 21 KB `SKILL.md`), the collection reviewer had to re-list every skill in its plan and read them under a fixed 24k-character budget (a 16-skill workspace failed every pass with "Read every current skill before reconciling"), collection review only recorded successes so a failing pass retried on every Gateway restart, and `openclaw skills curator status` showed nothing about what the last reviews did.
## Why This Change Was Made
- **Warm fork.** Experience review now continues the finished turn from the same in-memory session prefix (same session identity, bootstrap, skills prompt, tool schemas, `promptCacheKey`), appends one short review message, and runs with `sessionPersistence: "detached"` so nothing it writes reaches the foreground transcript or session record. Only `skill_workshop` executes; every other tool keeps its schema for cache parity and fails with a next-step message. The gate wraps the core tool list before Code Mode / Tool Search compaction, so catalog-hidden tools stay gated while `tool_call`/`exec` remain callable. `src/agents/embedded-agent-runner/run/attempt.skills-policy.test.ts` asserts identical system-prompt and tool digests between a foreground turn and its review, and that `tool_call` reaches `skill_workshop` but not `read` during review. Code Mode swarm globals (`phase`, `log`, `agents.run`) are the `sessions_spawn` capability and gate on the same allowlist, so a review cannot emit foreground lifecycle events or launch collectors. The review shares the foreground session, so it no longer retires that session's MCP runtime on run end; it reuses the warm runtime like any turn. Accepted tradeoff: the review inherits foreground tool construction (it creates the session MCP runtime only when the session has none, and spawns per-attempt LSP servers only when configured), because splitting tool construction from the foreground turn would break the cache-parity invariant this PR exists for.
- **One apply owner, rechecked at the write boundary.** `src/skills/workshop/autonomous-apply.ts` decides apply-vs-pending for both the post-review loop and foreground repair: creates and updates to Workshop-authored skills apply; updates to user-authored skills stay pending with a recorded reason. `applySkillProposalTransition` rechecks path-based ownership under the commit lock for non-operator actors (agent, or no actor), so a claim released after the pre-check cannot let an agent write a user-authored skill; gateway and CLI operators still approve any pending proposal. The user-authored pending write runs under the same commit lock and rereads the row, so an operator apply that lands first is kept. The old `auto-apply.ts` path and name-based `listWorkshopAuthoredSkillNames` are deleted.
- **Sparse collection plan and model-relative reads.** The collection reviewer returns only `write` and `drop` entries; unlisted skills stay untouched, so a 16-skill workspace no longer has to restate 16 keeps. The skill read budget is derived from the model's context window (35%) instead of a fixed 24k characters, with the 10,000-character skill cap still enforced per skill. No windowed read parameters: a skill is read whole or not at all.
- **Weekly cadence.** Collection review runs once every 7 days instead of daily (`REVIEW_INTERVAL_MS`), matching how slowly a skill library actually changes; the attempt is recorded before the model starts, so a failed pass does not retry on every restart.
- **Lean cap.** `AUTONOMOUS_SKILL_MAX_CHARS = 10,000` in `src/skills/workshop/collection-contracts.ts` is enforced by the tool and by collection reconcile; an oversized skill may only shrink. Tool description, experience prompt, and collection prompt were rewritten around procedures-not-records, one trigger per branch, and "NOTHING_TO_LEARN is the correct answer for most turns".
- **Detached runs end quietly.** A detached run writes no transcript or session record and runs under the foreground session key, so `attempt-finalize` now skips agent_end side effects for it: plugins do not observe the review as a foreground turn, and a deep review cannot schedule a successor review of itself.
- **Recorded outcomes.** Experience review records `applied | proposed | nothing | failed` with usage; one attempt per turn, drop on error. `openclaw skills curator status` prints the last collection and experience review outcome.
## User Impact
- Default `auto` mode: experience review reuses the foreground prompt cache and skips message/other tool execution, so review cost is one bounded continuation instead of a cold re-render.
- Autonomous edits touch only Workshop-authored skills; handwritten skills surface as pending proposals for operator approval.
- Autonomous `SKILL.md` results stay ≤ 10,000 characters.
- Collection review runs weekly and succeeds on larger skill libraries. A workspace with no recorded attempt reviews at the first daily check after Gateway start; an upgraded install keeps its recorded last attempt and reviews once it is older than 7 days. The 10,000-character cap applies to the next autonomous write; existing oversized skills are only ever shrunk.
- `openclaw skills curator status` shows the last collection and experience review outcome, time, and tokens.
- Docs: https://docs.openclaw.ai/tools/self-learning, https://docs.openclaw.ai/tools/skill-workshop
Related: #123866, #128871 both patch skills above the reviewer read budget; this PR caps autonomous skill size and restricts oversized skills to shrink-only rewrites.
## Evidence
- Live gateway (Linux, Telegram, `xai` provider) at `133ffe3`: manual experience review on a real foreground turn → `nothing` (usage: 0 uncached input, 13,902 cached, 36 output — the forked review hit the foreground prompt cache); manual collection review over the workspace → `succeeded`; the three oversized skills were rewritten under the cap (21,002 → 9,449; 11,882 → 7,735; 10,300 → 7,413 bytes, originals in `skill-workshop/collection-backups/`), the rest untouched, and the pre-PR "Read every current skill before reconciling" failure is gone. The next foreground turn's automatic experience review recorded `nothing` and `openclaw skills curator status` shows both outcomes. The apply recheck, MCP-runtime change (`26e821f`), and detached agent_end skip (`7e4a506`) landed after that run and are covered by the regression tests below.
- `pnpm test src/skills/workshop src/agents/tools/skill-workshop-tool src/gateway/server-methods/skills` plus `src/agents/embedded-agent-runner/run/attempt.skills-policy.test.ts`, `src/agents/embedded-agent-runner/run/attempt.tool-search-catalog-abort.test.ts`, `src/cli/skills-cli.curator.test.ts`, `src/agents/harness/tool-surface-bridge.test.ts` — green. New regression tests (`service.test.ts` agent-vs-operator apply on a user-authored skill and operator apply kept over a stale pending snapshot; `experience-review.apply.test.ts` no `cleanupBundleMcpOnRunEnd`; `attempt-phase-lifecycle.test.ts` no agent_end for a detached run; `code-mode-swarm.test.ts` swarm globals refused under the review allowlist) fail on the pre-fix code.
- Dependent sweep: 46 test files importing the touched modules — green.
- `oxfmt`, `scripts/run-oxlint.mjs` on changed files, `git diff --check` — clean.
- `pnpm tsgo && pnpm check:test-types` on Blacksmith Testbox — clean.
- Local ClawSweeper review (`gpt-5.6-terra`, high): `133ffe3` raised two findings (ownership recheck removed from the apply boundary; review retiring the shared session's MCP runtime), fixed in `26e821f`; `26e821f` raised one (detached review re-entering agent-end scheduling), fixed in `7e4a506`; `7e4a506` raised three: pending write racing an operator apply and Code Mode swarm globals bypassing the execution gate, both fixed in `e4b4322`; "prove detached review does not start configured MCP or LSP runtimes" is skipped as the cache-parity tradeoff stated above (LSP tool schemas come from the LSP runtime; the MCP runtime is session-owned and already warm). Maintainer decision on `e4b4322`: accepted — a detached review reuses the session MCP runtime and re-spawns configured per-attempt LSP servers exactly like a foreground turn; it still cannot execute them.
- Unrelated CI on `26e821f`/`7e4a506`: `check-lint-core-2` (`max-lines` in `src/gateway/server-methods/models-list-result.ts`, unused param in `models.test.ts`, both from #129332) and `checks-node-compact-small-8` (`doctor-auth.profile-health.test.ts`, Claude CLI auth from #129052) fail identically on `main` run 32857954734; `checks-node-compact-small-31` pins plugin SDK export counts (`4340` vs `4342`) that #129052 moved — this PR touches no `src/plugin-sdk` file. On `e4b4322` the failing set is the same twelve `checks-node-compact` shards that fail on `main` run 32857954734 (`large-5/12/13/14/15/18/22`, `small-8/14/20/21/26`) plus `small-31`; `check-lint-core-2` passes here. None are fixed here; #129357 carries the `main` fix.
- LOC (raw numstat): production +712 / −664 (net +48; the four review-fix commits after the live run add +76 / −26); tests + docs +1269 / −1238 (net +31).
Stop OpenClaw from copying or refreshing Claude CLI OAuth tokens.
Claude CLI now owns native login and refresh state; Doctor removes retired copies while preserving CLI routing.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(cli-runner): drop stock watchdog defaults from CLI backend descriptors (#125045)
The claude-cli and gemini-cli backend descriptors spread
CLI_FRESH/RESUME_WATCHDOG_DEFAULTS into config.reliability.watchdog.
Those copies are byte-identical to the fallback pickWatchdogProfile
already uses, so shipping them only makes `configured` always truthy,
which permanently disables the promotion gate (!configured) and pins
resumed cron/explicit-timeout turns to the 180s resume no-output ceiling
instead of the 600s fresh ceiling.
Removing the blocks lets `configured` be undefined, restoring the
promotion gate. No user-config path is affected: reliability.watchdog
is not exposed in the config schema, the descriptor is the sole source,
and pickWatchdogProfile is the only runtime reader. The codex backend
never shipped the block, so this aligns the three backends.
Co-Authored-By: Claude <noreply@anthropic.com>
* test(cli-runner): add descriptor-through-resolver watchdog proof (#125045)
Drive the real shipped Anthropic/Gemini CLI backend descriptor config through
the production resolveCliNoOutputTimeoutMs resolver to prove the resumed-cron
promotion is live post-fix. Pre-fix the descriptor shipped byte-identical
resume defaults, so configured was truthy and the !configured promotion gate
was dead, pinning resumed cron turns to the 180s resume ceiling (180000).
Post-fix the block is gone and the resolver returns the promoted fresh budget
(480000 = 600000 * 0.8).
resolveCliNoOutputTimeoutMs is re-exported through the test-only
openclaw/plugin-sdk/test-fixtures subpath (excluded from published dist), so
extension tests can drive the real descriptor through the real resolver without
crossing the extension/core boundary.
Co-Authored-By: Claude <noreply@anthropic.com>
* test(anthropic): split watchdog promotion proof into its own file
Move the descriptor-shape and descriptor→resolver real-behavior proof
tests out of cli-shared.test.ts into cli-watchdog-promotion.test.ts.
cli-shared.test.ts exceeded the extensions max-lines budget (1000
non-blank/non-comment lines) once the proof tests landed; the focused
file keeps the same test names and coverage while staying under the
budget. No production change.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(cli-runner): restore inherited resume watchdog promotion
Validate bundled plugin descriptors through their public setup entries and keep custom watchdog contracts intact.
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
* chore(cli-runner): align conflicting provider owner with main
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
* fix(cli-runner): complete inherited resume watchdog promotion
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
* test(cli): keep watchdog regressions in their owning plugins
Keep Anthropic and Google descriptor assertions inside their plugin-owned registration suites, preserve generic resumed/custom watchdog coverage in its core owner, and remove the cross-plugin core test rejected by architecture CI.
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
---------
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(systemd): honor effective drop-in configuration
Read one coherent manager-loaded service snapshot while retaining the managed base definition for repairs. Track drop-in ownership, surface pending daemon reloads, and keep update and doctor rewrites from adopting operator-owned values.
Closes#128929
* fix(systemd): expose reload status in service snapshot
* test(systemd): cover Linux doctor ownership
CI jobs are not installs; unchecked they outnumber operators by orders of
magnitude and make version and platform counts meaningless. A configured
telemetry endpoint still reports, so update-path E2E lanes keep working.
* feat(ui): dock the session progress card beside a wide composer
The progress card had two placements: the companion rail when that side
panel is open, otherwise a collapsed one-line bar stacked inside the
composer box. On a wide chat the composer stays centered at the
transcript width, so the space either side of it sits empty while the
card is squeezed into the composer.
Add a third placement. When the measured free gutter beside the composer
is at least 280px, the card docks into it with its full checklist
expanded; below that it falls back to the existing composer bar, and an
open companion rail still wins. Exactly one placement renders at a time,
now expressed as a closed {card, placement} prop so the composer bar and
the dock cannot both draw the same card.
The gutter is measured from the DOM by a small ResizeObserver controller
rather than derived from the pane width: the transcript width is a
browser-local setting in arbitrary CSS units, and an open side panel
shrinks the conversation column without changing the pane. The dock is
positioned absolutely in that gutter, so the transcript and composer
never shift when a card appears or is dismissed, and its inline-start
edge repeats the composer's own half-width formula through a shared
--chat-composer-side-inset token so the two stay in agreement.
* fix(ui): keep the new-session composer sized outside the chat surface
The composer shell tokens are declared on .chat, but the new-session page
reuses .agent-chat__composer-shell outside it. There the var() had no
value, so the whole width declaration was invalid and dropped, and the
composer stretched full-width instead of holding its 48rem centered box.
Give both tokens their literal fallback at the use site, matching the
neighbouring --chat-thread-max-width. The custodian surface was already
immune because it overrides width outright.
Caught by ui/src/e2e/new-session-page.places.e2e.test.ts, which is the
regression test for this: it failed on the previous head and passes now.
* fix(memory): apply temporal decay to dated files in memory subdirectories
DATED_MEMORY_PATH_RE only matched memory/YYYY-MM-DD.md at the root, so
dated files in subdirectories (e.g. dreaming phase reports at
memory/dreaming/light/2026-05-14.md) were classified as evergreen and
kept full ranking weight forever even with temporalDecay enabled.
Accept a date-only Markdown basename at any depth below memory/, matching
the nested-path shape short-term promotion already recognizes. MEMORY.md,
USER.md, and undated files remain evergreen; non-memory sources keep the
mtime fallback.
Fixes#121046
* refactor(memory): trim issue-specific commentary from dated-path matcher
* fix(memory): decay nested and slugged dated memory files
Co-authored-by: wangmiao0668000666 <wang.miao86@xydigit.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(cli): add openclaw triage for sanitized agent debugging handoffs
Collects read-only doctor findings through a new collectDoctorFindings seam,
reuses the sanitized diagnostics export, writes a bounded 8 KiB debugging
prompt, and prints Claude Code / Codex / embedded agent handoff commands.
Embedded --run gates on an interactive terminal plus a live inference probe.
* fix(doctor): keep word separation when scrubbing multi-line errors
scrubDoctorErrorMessage dropped newlines without substituting a space, so
multi-line errors rendered with glued words in doctor and triage output.
* feat(cli): hand triage prompts straight to a detected coding agent
Interactive triage now detects installed agents via resolveExecutablePath,
offers embedded/Claude Code/Codex/print in a picker, and spawns the chosen
binary with the prompt, propagating its exit code. Embedded inference is
probed only after selection. JSON output adds detectedAgents.
* fix(cli): redact local paths in triage prompts and drop unlaunchable targets
Prompt content now uses canonical path-aware redaction (redactSupportString),
so paths reach external agents as ~/... or $OPENCLAW_STATE_DIR/... instead of
absolute home paths; operator-facing JSON paths and printed commands stay real.
Findings are fitted to the byte budget left after the trailing sections so the
omission notice, bundle path, and privacy statement always survive truncation.
Windows command shims are listed as manual commands rather than offered as a
direct launch that shell-less spawn rejects.
* fix(cli): clarify which Node runtime triage reports
A live agent run flagged the reported version as wrong because the shell
default differed from the runtime executing the CLI.
* docs: route new installs to openclaw triage when setup fails
Getting Started had no recovery path; add one that leads with triage, and
document prompt location, environment inheritance, and exit codes on the
CLI page. Keep the Triage page title untranslated like Doctor.
* docs(copilot): describe the transcript journal, not the removed dual-write mirror
The new copilot plugin doc described transcript persistence through
extensions/copilot/src/dual-write-transcripts.ts with session-scoped
content-fingerprint keys and best-effort failure containment. That module
was removed in #114403 (Jul 27): attempts now own an
attempt-transcript-journal keyed by run id and SDK event id, journaled in
structurally complete groups, and persistence failures fail closed by
aborting the SDK session and flagging replay as unvalidated.
Rewrite the section around the current journal, and update the two
"transcript mirror" references that pointed at it.
* docs(copilot): clarify invalid replay starts a fresh session
Co-authored-by: Santhi Prakash <b.santhiprakash@gmail.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(gateway): bound live health aggregation
Keep verbose health within a Gateway-owned deadline, preserve healthy sibling results when channel probes stall, and document that the CLI timeout controls the Gateway connection.
Fixes#128890
* fix(gateway): retain health probe permits
Keep process-wide health capacity occupied until timed-out plugin work actually settles, and prove repeated refreshes cannot replace five stalled probes.
* fix(agents): preserve CLI session binding on format-class failover
shouldClearFailedCliSessionBinding cleared the stored CLI session
binding for every FailoverError, including format-class failures
(output limit exceeded, parser error, unsupported image input). A
format-class failover means the stream could not be read, not that
the CLI session on disk is invalid — clearing the binding silently
lost the conversation context on the next turn.
Narrow the condition to only session-invalidating failover reasons
(session_expired, auth, auth_permanent) using a whitelist so new
reasons default to preserving the binding.
Fixes#128698
* fix(agents): preserve valid CLI session bindings
* fix(agents): retain CLI binding across format recovery
* fix(agents): scope fresh CLI recovery by backend
* docs(plugins): document CLI recovery policy
---------
Co-authored-by: Vincent Koc <vincentkoc@users.noreply.github.com>
Queue Telegram-visible Mantis proofs at workflow level so only one shared-user run allocates a runner at a time.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* docs(install): surface desktop app download links for users who want a normal app install (#126694)
* docs(install): clarify both desktop apps provision gateways
Keep the contributor desktop-download instructions accurate for both Windows Hub and the macOS app: each can provision a local Gateway during first-run setup or connect to an existing remote Gateway.
Co-authored-by: Finn763 <165816600+Finn763@users.noreply.github.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(macos): honor user accent precedence in config snapshot and live-update chat window
The Control UI user accent (ui.prefs.accent) landed in #128432/#128577 with
precedence user accent -> operator ui.seamColor -> theme default, and the
gateway's talk.config payload already applies it. The macOS app had two gaps:
- ChannelsStore.applyUIConfig read raw ui.seamColor from the config.get
snapshot and clobbered the user accent set from talk.config depending on
arrival order. It now resolves ui.prefs.accent ?? ui.seamColor via a
testable helper mirroring the gateway precedence.
- The native chat window read AppStateStore.seamColorHex once at window
construction, so accent changes never live-updated. MacChatSurface now
reads the @Observable store in body, deleting the one-shot userAccent
plumbing.
Docs: configuration-reference.md documents the precedence for native-app
chrome. Regression test fails pre-fix (snapshot returned the operator seam
color instead of the user accent).
* fix(macos): refresh config from gateway config.changed events
Addresses the review finding that no macOS consumer turned the gateway's
hash-only config.changed broadcast into refreshed shared state, so a
Control UI accent change never reached an open native chat window while
the app ran. ChannelsStore now subscribes to gateway pushes and re-fetches
config.get on config.changed, reconnect snapshots, and sequence gaps.
The refresh applies non-force so an in-progress local settings draft wins
(the gateway rejects stale-hash writes anyway). The in-flight reload queue
gains a closed pending level (none/refresh/force) so a refresh arriving
during a load is coalesced instead of dropped, and a requeued refresh
cannot clobber a dirty draft the way the old boolean force-pending did.
* feat(geolocation): resolve client addresses to a coarse city via a bundled plugin
The Activity identity card could show a client's IP address but not where it
was, so an operator still had to look the address up by hand.
Add a bundled `geolocation` plugin that owns address-to-place resolution behind
one authenticated route, `GET /plugins/geolocation/lookup?ip=`. It downloads a
MaxMind-format database on first lookup into the state directory, answers from
that local copy, and refreshes it monthly, so a lookup never sends an address
to a third party. The Control UI renders the resolved city on the device row
next to the address and the client-reported time zone.
The default source is DB-IP City Lite under CC BY 4.0. That license requires
attribution, so every response carries the credit and the UI renders it next to
the value; the database is downloaded at runtime and never redistributed.
Plugin code and the `maxmind` reader are MIT. No free city-level IP database is
MIT-licensed, so the obligation lives with the data rather than the code, and
`databaseUrl` plus the attribution fields make the source swappable.
No new core provider kind: with one implementation the plugin owns everything
through the existing HTTP-route seam, keeping core plugin-agnostic. A second
provider is what would justify promoting this to a registry contract.
Availability and lookup failure stay distinguishable: a missing or still
downloading database answers 503, never `found: false`. A failed refresh serves
the cached copy, and a body that does not parse as an MMDB is discarded without
replacing a working database.
* fix(docs): correct geolocation config examples and add zh-CN glossary entries
The config examples used `plugins.<id>` instead of the real
`plugins.entries.<id>.config` shape, which the docs config-example
validator and src/config/docs-config-examples.test.ts both reject.
New doc labels also need zh-CN glossary entries.
* chore(labeler): cover the geolocation extension directory
AGENTS.md requires a labeler entry plus a GitHub label for every new
plugin surface; test/scripts/labeler-extension-coverage.test.ts enforces
the labeler half.
* fix(geolocation): address review findings on caching, download bounds, and scope
Cold-start lookups were permanently suppressed. The loader cached one promise
per address including failures, so the 15s browser deadline expiring against a
first download that takes ~46s cached a blank forever, and a mounted row only
looks up again when its IP changes. Lookups now return a discriminated
located/absent/unavailable result: only definitive answers are cached, and the
element retries an unavailable one on a widening 5s/15s/45s schedule.
Download limits ran after allocation. The size check happened only after
`response.arrayBuffer()` had buffered the whole body, and gunzip had no output
ceiling, so a replaced source or a compression bomb could exhaust Gateway memory
before rejection. The body now streams against a compressed ceiling enforced
per chunk, and inflation uses zlib's maxOutputLength.
Cached placements were not scoped to the Gateway. The cache keyed only by
address while endpoint and credentials come from the shared Gateway context, so
a switch could render the previous Gateway's answer. The shared reset hook now
supports multiple subscribers - a single slot silently dropped whichever
registered first - and the geolocation cache subscribes.
Unresolvable ranges no longer trigger a download. Only loopback suppresses `ip`
at connect, so Tailscale carrier-grade-NAT and LAN addresses are recorded and
displayed. No geolocation database contains them, so a tailnet-only or LAN-only
Gateway was downloading 125 MB to answer nothing. The route now answers those
ranges without loading the database, using the already-public
`isPrivateOrLoopbackHost` seam so the SDK surface budget is unchanged.
The quickstart queried a reserved documentation range while showing a located
response, which cannot happen; it now uses a routable address and documents the
not-found case.
* fix(deps): resync the lockfile after dropping the net-policy dependency
The geolocation plugin briefly depended on @openclaw/net-policy before
switching to the already-public isPrivateOrLoopbackHost SDK seam. The
package.json entry was removed without regenerating the lockfile, so the
frozen-lockfile install failed and every downstream CI job failed with it.
Dallin Romney