## 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>
* perf(gateway): stop sessions.files.list convoying on cold transcript folds
Raise the touched-files fold cache from 16 to 256 entries so >16 viewed
sessions no longer thrash it into repeated full-transcript refolds, use
in-process insideGitCheckout instead of spawning git rev-parse per call,
and reuse one prepared fs-safe workspace root per request instead of
re-opening it per touched file.
* perf(gateway): make session PR subscriptions non-blocking with poll-aligned caches
controlUi.sessionPullRequests.subscribe now registers the replace-set,
pushes cached snapshots, hydrates cold keys in the background (bounded
concurrency 4) and responds immediately instead of awaiting serial git
and GitHub batteries per session key. Local-git caches grow 10s->75s and
the GitHub success cache 60s->90s so the 60s poll hits warm caches
instead of re-running the full battery for every watched session every
minute; failure and rate-limit semantics are unchanged.
* perf(skills): memoize workspace skill discovery on the watcher snapshot seam
loadSkillEntries now serves repeat loads from a bounded 64-entry cache
keyed on the chokidar snapshot version plus config fingerprint, instead
of synchronously rescanning every skill root 4-6x per inbound message.
The session snapshot cache serves fresh sessions read-through, and the
skill precedence collision warning fires once per snapshot version
instead of ~9k times/day.
* perf(gateway): bound sessions.preview and chat.message.get transcript reads
sessions.preview reads a bounded recent tail (64-256 events/1MiB, one
8MiB escalation when filtered messages starve the 12-item preview) and
yields between batch keys instead of synchronously materializing every
transcript in a 64-key loop. chat.message.get resolves visibility via
the message-ID-anchored reader instead of loading the full projected
transcript per request.
* perf(ui): consolidate Control UI boot chunk graph for HTTP/1.1 gateways
The Control UI boot flow (app shell + sidebar + chat route) lazily loaded
~124 automatic chunks in one burst after the gateway handshake, which the
gateway's HTTP/1.1 transport serializes into ~24 six-connection round-trips
on high-latency links (Tailscale, remote gateways).
Add a measured boot-module manifest (ui/config/control-ui-boot-modules.json,
regenerated via pnpm ui:boot-manifest:gen) and a control-ui-boot codeSplitting
group that merges exactly that module set into a handful of chunks with
recursive dependency inclusion. Lazy islands (locales, ghostty-web, novnc,
non-default routes) keep their own chunks; stale manifest entries degrade
gracefully back to automatic chunking.
Measured on the built dist with the mocked gateway (chat route, 3 runs):
unique boot JS requests 140 -> 45, raw boot JS 3751 -> 3717 KiB, chat
composer interactive at simulated 50 ms RTT ~1600 ms -> ~575 ms.
Largest-CSS budget rises 45 -> 47 KiB for the merged boot CSS; startup JS
gzip baseline ratchets down (345049 -> 339214 B) as consolidation shrinks
the startup graph.
* chore(ui): refresh boot module manifest after rebase onto current main
* fix(ui): stop the pending lazy shell action replay loop starving boot
When a pending lazy shell action (command palette open, panel toggle)
replayed while the shell was still splash-gated, the dispatched event had
no rendered element to consume it and re-entered requestLazyElement in a
microtask cycle: request -> load -> replay -> dispatch -> request. The
cycle starved tasks (Gateway WebSocket messages included), so the boot
never finished and the recovery e2e froze on the splash screen.
Gate replay on the element actually being rendered: the controller skips
the action after load until the host's render root contains the tag, and
restorePendingLazyAction skips dispatch while a defined element is still
render-gated. The host retries after every completed update, so the replay
fires on the update that first renders the element. Regression test fails
on the pre-fix controller.
* fix(ui): re-anchor the scope-upgrade details popover before opening
wa-popover resolves its `for` target once per property change and never
re-resolves a missing or replaced anchor. The trigger with the shared id
can render after the popover's first update (the header trigger ships with
the lazy chat chunk), leaving the opened popover permanently invisible:
active popup with a native [popover] part stuck at UA display:none because
showPopover() never ran without an anchor. Re-arm the watcher when opening
while the anchor is missing or disconnected.
* test(ui): compare settled layouts in device-scope stability assertions
The 0.5px no-move assertions sampled geometry that later reflowed when the
details surface's first render fetched glyph subsets, reporting sub-pixel
drift the open never caused. Burn in the one-time open per context and
sample the baseline adjacent to the click.
* fix(ui): map the keyboard shortcuts dialog in lazy replay gating
Current main added the keyboard-shortcuts lazy shell event; the replay
gate's exhaustive event-to-element record needs its entry.
* chore(ui): refresh startup budget baseline after rebase onto current main
* refactor(anthropic): explore official Claude Agent SDK runtime
* refactor(anthropic): replace handwritten Claude sessions with SDK
* refactor(anthropic): collapse SDK live-session ownership
* refactor(anthropic): simplify SDK ownership and preserve live skills
* fix(anthropic): fence cancelled SDK runs before process startup
* fix(anthropic): harden SDK approvals, lifecycle, and packaging
* refactor(anthropic): own SDK process trees and streamline runtime
* fix(anthropic): repair rebased packaging and legacy test fixtures
* feat(gateway): generic operator roles for non-maintainer access
Adds gateway.roles: named role bundles over a closed capability set —
sessions.others (none/view/suggest/write), an agents allowlist, and an
operator-scope ceiling. Roles are person-level (additive user_profiles.role
column, SQLite stays at v9); users.setRole (admin-only) assigns them. With
no gateway.roles config, behavior is unchanged for solo deployments.
Enforcement is deny-by-default from a host-minted actor identity
(system vs operator+profileId on server-only client.internal, never
accepted from the wire) and covers every entry point: WS RPCs, OpenAI-compat
and Responses HTTP, tools invocation, cron, questions, usage, task
suggestions, session catalog/sharing/reads. The agents allowlist gates both
session creation and run-start on existing sessions. Subagent completion
announce and descendant wake mint explicit system authority so role
boundaries never silently drop parent notifications.
The enforcement surface is expressed through a narrow policy vocabulary
(operatorSessionCap, hasOperatorBoundary, authorizeSessionSharing) rather
than per-handler policy internals.
* fix(gateway): heal PR CI after rebase onto main
- Break import cycles: extract GatewayOperatorRoleActor leaf contract; merge
session-group-mutation-targets into session-sharing-target-input.
- Split sessions-suggestions.test.ts (max-lines) into a visibility suite.
- Add users.setRole to the 2026.8 train registry test and regenerate the
Kotlin protocol client.
- Startup UNAVAILABLE gating now precedes session authorization: session
stores are not loaded during startup, so authorization reads would deny
with a misleading non-retryable error.
- sessions.assignOwner keeps its documented visibility-authorized contract
when no operator role caps the caller; view/suggest-capped roles still
cannot reassign foreign session ownership.
- Test stubs updated for main's socket readyState guard (#128144) and the
system-authority arg on channel-native resets.
* test(gateway): chat.send pending-profile dispatch carries its required session target
chat.send requires a non-empty sessionKey at the protocol level; the mutation
pipeline now rejects targetless frames before profile-dependent dispatch, so
the pending-profile test must send a realistic frame.
* fix(memory): enforce canonical SecretRef resolution
Make Gateway runtime snapshots the exclusive owner of memory SecretRef materialization. Bind embedding credentials and headers to provider-owned destinations, and fence per-agent stale reuse by the provider destination/auth contract.
Release note: Memory search resolves secret references through configured provider policy and keeps embedding credentials scoped to their intended destination.
* fix(lmstudio): preserve resolved memory headers
memory remote headers are already materialized by the Gateway snapshot and now bypass SecretRef re-resolution; provider-owned headers retain canonical resolution; final loopback request proof covers literal preservation and precedence.
* fix(memory): bind stale credentials to auth owners
Resolve memory adapter credential owners from snapshot manifest metadata, conservatively fail cold when metadata is absent, and prove Gemini/Google destination changes plus zero-egress unresolved refs.
* fix(memory): scope compatible embedding credentials
Apply destination ownership to the core compatible adapter while preserving destination-owned credentials and intentionally unauthenticated endpoints. Distinguish loopback principals, consolidate duplicate security tests, and verify the final credential boundary through a live isolated Gateway request.
* test(memory): align destination auth precedence
* fix(memory): bind credentials to query identity
Include URL query parameters in embedding destination ownership so provider credentials and headers never cross tenant boundaries.
* fix(memory): preserve query-bound embedding destinations
* fix(compaction): re-compact when last entry is a compaction record but context still exceeds window
prepareCompaction short-circuited to undefined whenever the last session
entry was a compaction record, treating it as a no-op signal. That proxy
does not hold: the retained context (prior summary + kept recent turns +
system prompt + injected files) can still exceed the compaction threshold,
so a session can wedge above 100% of its window while the compactor
reports nothing to do. It only becomes eligible once enough new turns
append that the compaction record is no longer last.
Drop the last-entry-is-compaction-record short-circuit so the existing
re-compaction path stays reachable: prepareCompaction continues to walk
the retained tail, and when there is compactable content the prior
summary flows through previousSummary into UPDATE_SUMMARIZATION_PROMPT.
The empty-branch and last-entry-is-reset no-op guards are preserved, and
the existing messagesToSummarize.length===0 guard still returns undefined
when there is genuinely nothing new to summarize.
Related to #120290
* fix(compaction): harden re-compaction state handling
---------
Co-authored-by: Altay <altay@hey.com>
Ayaan Zaidi