* fix(agents): strip external-content nonce before loop-detection resultHash
External-content wrappers carry a fresh anti-forgery nonce per result
(createExternalContentMarkerId). The loop detector hashed the wrapper text
including the nonce, so identical failing wrapped tool calls never reached
the no-progress block threshold — the detector warned but never blocked,
leaving the agent to loop until it gave up (#130210).
Strip the wrapper nonce before hashing in extractTextContent, mirroring
the existing stripVolatileSendIds treatment of per-call send ids (#89090).
Only the hash is normalized; the delivered wrapper keeps its nonce, so
anti-forgery is unaffected.
Closes#130210
Co-Authored-By: Claude <noreply@anthropic.com>
* test(agents): validate protected error marker capture
Narrow the generated marker capture before recording it so the real network-error loop regression also passes its owning test-type graph.
---------
Co-authored-by: ruel225 <ruel225@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Make ask_user use native Telegram controls when one single-select question can be represented directly, while keeping multi-select and batch questions on the text fallback. Add native Other-to-ForceReply input, preserve Gateway-owned option ordering across Telegram, Slack, and Discord, and prevent later progress from overwriting the delivered question.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(ui): highlight source code in chat and session diffs
* test(ui): isolate roster bootstrap from slug lookup
* fix(ui): preserve both languages in renamed diffs
Keep explicit CLI choice precedence while preserving existing pnpm, Bun, and Yarn preferences when onboarding is rerun. Retain npm defaults for fresh config and the narrower explicit CLI choices. Fixes#130433.
* feat: make model persistence scope explicit
* fix: keep unauthorized model directives as plain text
An unauthorized sender's inline directives are cleared to plain text before
they are acted on, but the persistent write target was derived from the
directives as they arrived. An unauthorized `/model <provider/model> -a` or
`-g` therefore reached the owner-authority error return instead of the
plain-text path every other directive takes.
Derive the target, the authority flag, and the persist flag after the
clearing. Authorized senders skip the clearing, so their behavior is
unchanged; the only closure that captures the target is invoked well after
the new declaration site.
Covers both persistent flags with a regression test that fails on the prior
head with "Agent and global model defaults require owner authority or
operator.admin scope." and passes here.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* feat(models): make model selection scope configurable
Preserve current model-selection defaults when the optional scope preference is unset. Add explicit session, agent, and global overrides without broadening config-write authority.
Co-authored-by: Marvinthebored <peter@lindsey.jp>
* test(models): simplify scope expectation fixtures
---------
Co-authored-by: Marvinthebored <peter@lindsey.jp>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(approvals): typed approval scope summaries on channel cards
Approval owners can attach a closed ApprovalScope union (message-send,
payment, external-post) describing an action's blast radius. The gateway
sanitizes it once at the producer boundary, the core view model renders a
Scope metadata row so Slack/Discord/Google Chat cards show it unchanged,
shared text builders cover Telegram/WhatsApp/Signal/iMessage/Matrix, and
the durable presentation carries it additively for operator surfaces.
Scope is display-only, never authorization; missing scope keeps today's
cards.
* fix(approvals): emit native ApprovalScope union and clamp recipient previews
Name the three scope variants as registered protocol schemas so the Swift
generator emits the ApprovalScope discriminated union the presentation
structs reference, and commit the regenerated GatewayModels.swift. Clamp
recipient previews to the declared recipientCount at the sanitize boundary
so a count of 1 with 2 previews can no longer render inconsistently.
Addresses both ClawSweeper findings on #130116.
* refactor(approvals): extract text sanitizer to break the exec-approvals import cycle
check:architecture flagged approval-scope joining the exec-approvals SCC
through exec-approval-command-display. Move the self-contained display
sanitizer into a leaf module (exec-approval-text-sanitize) with no
exec-approvals imports and migrate all sanitize importers; command-display
keeps only the payload-typed command/preview resolver.
* chore(plugin-sdk): ratchet public surface budgets down after sanitizer extraction
The approval display sanitizers left the publicly reachable SDK graph when
they moved to the exec-approval-text-sanitize leaf: exports 4343 -> 4338,
callable exports 2582 -> 2578. Shrink-only budget pin.
The tool page shipped with the feature but nothing pointed at it: the
tools index still listed only ask_user under human input, and neither the
secrets CLI page nor ask-user mentioned it. Add those three references.
The page also missed the trap the card invites: the allowed-hosts field
is editable, and clearing it stores a secret that can never be
substituted at egress, so the request looks successful and the credential
is unusable. Warn where the operator edits it, say what a skipped or
expired prompt means for the agent, and drop wording left over from when
requests could ask for an env value.
* feat(secrets): agent-requested credentials the model never sees
The new main-session secrets tool lets the agent request a credential by
name: the human enters the value in a masked question card (Control UI,
/ask/<id> deep link, iOS/macOS/Android), and the gateway diverts the
answer straight into the shared secret store at question.resolve. The
record, broadcast, waitAnswer, tool result, transcript, and model context
only ever carry a synthetic stored marker.
- protocol: additive secretStore binding, secretStoreExisting replacement
metadata, and resolve-time secretStoreAllowedHosts (since 2026.8)
- gateway: store-bound question validation, admin-gated minting (blocks
questions-scope self-answer escalation past secrets.store.set), shared
redaction-first store write service reused by secrets.store.set
- tool: secrets request/list/delete; write-only by design, delete carries
verified agent runtime identity; channel delivery is link-only so chat
text is never captured as a secret
- Control UI: masked composer card with requester identity, store banner,
editable allowed hosts, replacement warning, retry-on-validation-error,
a standalone /ask/<id> page, and a named startup-JS baseline bump
- mobile: SecureField / password transformation for isSecret questions,
no answer echo in terminal summaries; new native string registered in
the locale-refresh inventory (generated artifacts stay workflow-owned)
- regression: claimed harness secret input stays out of session transcripts
Live-proven on an isolated dev gateway: real model turn, masked entry via
Playwright, value present only in secret_store_entries, absent from every
transcript, log, and the DOM.
* chore(protocol): regenerate protocol models and tool display
* fix(cli): read image string options through a typed helper
PR #129463 added four commander option narrowings in image.ts without
SAFETY coverage, leaving the assertion-safety ratchet red (21 > 17) for
every branch on current main. Replace the casts with a typeof-checked
read so the assertions are removed rather than annotated; each value is
still validated by its normalizer. SAFETY comments cannot work in this
file: the ratchet's raw scanner never rescans template tokens, so
comments after the first substitution template are unreadable to it.
* chore(protocol): refresh Swift models against current main
* chore(i18n): re-baseline the native inventory on current main
* docs(secrets): state the default-on tool policy and how to disable it
* fix(secrets): tell the model what the store actually does
The shipped tool description named the three actions and nothing else,
and no parameter carried a description. The model could not tell that
request blocks a human, that reason is shown to that human, what secret
and env select, or - the silent-failure case - that a secret stored with
no allowedHosts can never be substituted, so a successful request could
produce a permanently unusable credential. Move the description to the
presets module beside ask_user and document every parameter.
* refactor(agents): share one blocking-question lifecycle between tools
ask_user and secrets each carried their own registration, wait, and
cancel logic, and they had diverged: ask_user recovers an answer that
lands between its wait timeout and the cancel, while secrets discarded
it and reported no_answer even though the Gateway had already stored the
credential. One shared canceller and answer reader fixes that race for
both, folds the two divergent gateway-call types into one, and drops two
type assertions in favour of the canonical record guard (ask_user's
assertion baseline shrinks 11 -> 8).
Net +49 production lines: the shared module costs more than the
duplication it removes, and buys the correctness fix plus a single owner
for question lifecycle.
* fix(ui): keep the allowed-hosts field readable as an input
Main's composer restructure moved the free-text input styling into the
option-row context, so the store-request hosts field - which sits outside
a row - lost its border and read as static text. It is the one field the
operator is meant to review and edit before releasing a credential, so
give it its own border and focus ring.
* fix(secrets): close two credential-boundary holes in agent requests
Requests are now protected-secret only. list renders env values, so an
agent could request kind=env, watch a human type it into a masked box
under a no-visibility promise, then read it straight back; the tool text
even claimed values are never returned. Environment values stay operator
-set in Settings or the CLI, where they are agent-readable by design.
Store-bound questions are also bound to the run that requested them. The
resolve path authorized only the answering client, so a terminated or
replaced agent run could still have a credential written on its behalf -
the recorded runId was provenance, not closure-bound authority. Minting
now requires a runId and resolution revalidates that exact live run
immediately before the store write, with no await in between, failing
closed as QUESTION_REQUESTER_INACTIVE.
Both reported by ClawSweeper as P1 credential-boundary findings.
The schema already rejects a hand-declared openclaw or clawd profile
that sets neither. Stop telling operators those fields auto-assign.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(skills): expose the collection review as a single cron-driven run
The weekly Skill Workshop collection review lived in a bespoke gateway
maintenance loop with its own interval, due-state check, and gateway admission
wrapper. Replace it with `runSkillCollectionReviewForAgent`, one run that keeps
the workspace claim, identity check, and status recording and reports
`{ status, summary }` to its caller. The cron timer that will own the schedule
already runs under gateway root work admission, so the extra wrapper and its
drain test go away.
* feat(cron): schedule the skill collection review as a system-owned job
Fold the weekly Skill Workshop review into cron the same way the heartbeat
monitor is: a `skillCollectionReview` payload kind, one system-owned job per
workspace (first agent of each canonical workspace) reconciled at gateway
startup, enabled only when the workshop runs in `auto` mode, on a 7-day
`every` schedule with the heartbeat phase spread. The timer invokes the
review runner directly; cron clients cannot create, patch, or remove
system-owned payloads, and the job shows up in `openclaw cron list`.
* fix(cron): warn when skill reviews are disabled
* fix(cron): preserve skill review failure details
* fix(cron): narrow serialized payload variants
* test(cron): tighten skill review coverage
* fix(gateway): reconcile skill reviews on config reload
* fix(cron): fence revoked skill reviews
* fix(cron): satisfy system payload type coverage
* fix(ui): protect system-owned cron jobs
* test(gateway): prove review revocation fence
* fix(gateway): revoke skill reviews before config publish
* fix(scripts): see SAFETY comments after template substitutions
The assertion ratchet scanned each file with a raw ts.createScanner, which
reads the `}` closing a template-literal substitution as a block close. The
scanner desynced there, so every `// SAFETY:` comment after a file's first
`${...}` was invisible and its annotated assertions were counted as bare.
Track substitution brace depth and rescan the closing brace the way the parser
does. The fix uncovers already-annotated assertions in eight files, so the
baseline shrinks accordingly.
* fix(skills): record skill usage again and retire dead curator tables
Skill lifecycle curation shipped in 2026.7.1 with two producers: a trusted
`skill.used` consumer writing `skill_usage`, and a daily sweep aging skills
into `skill_lifecycle`. The weekly collection review (#121653) replaced that
policy and deleted both producers, but left every reader in place. Since then
`skill_usage` has had no writer, so curator status reported `lastUsedAtMs:
null` and `useCount: 0` forever, and `skill_lifecycle` had no writer either, so
pin/unpin/restore either threw "not found" on fresh installs or, worse,
"succeeded" on upgraded ones while gating nothing at all.
`skill_workshop_proposal_origin_runs` was never read in any revision; proposal
provenance is authoritative in `record_json`.
Restore the usage producer at its owner and retire what has no owner:
- `skill.used` events populate `skill_usage` again, registered with the
collection-review maintenance it belongs beside. Curator status derives
curated skills from applied Workshop create proposals and reports real usage.
- Weekly review receives bounded `useCount` / `lastUsedDaysAgo` evidence, with
prompt text stating usage supports keeping a procedure and never alone
justifies a drop.
- State schema 10 drops `skill_lifecycle` and
`skill_workshop_proposal_origin_runs`. Previously archived skills return to
the active collection, where review judges them by content; the migration
logs how many. Reconcile now clears usage only for skills it actually drops.
- `skills.curator.pin`/`unpin`/`restore` stay registered for existing clients
but fail with an explicit retirement message instead of silently doing
nothing.
Retirement code moves to openclaw-state-db-table-retirements.ts to keep the
schema-repair module under max-lines; that split is a pure relocation.
Production delta is +23 raw: about -46 for the change itself, +44 for the file
split, +25 for the ratchet fix in the previous commit.
* fix(ollama): skip local checks for hosted web search
* fix(ollama): explain hosted search API key requirements
* docs(ollama): separate hosted and local search setup
One shared agent could not give different people different execution
isolation: sandbox selection saw only the agent's mode and the session key,
never who created the session. Teams that share an agent had to choose
between sandboxing everyone or no one.
Add a closed, optional role axis sandbox?: "inherit" | "required"
(default "inherit", so existing deployments are unchanged). The Gateway
resolves it from the authenticated session creator and stamps it write-once
onto genuinely new session entries; the agents layer reads that immutable
fact and forces a sandbox even when the agent config says mode: "off".
Anchoring on the creator keeps a session's isolation stable for its lifetime:
a maintainer replying inside a guest session cannot un-sandbox it, and a
caller whose role requires a sandbox is refused inside an unstamped session
rather than silently inheriting host execution. Reset, rewind, rollover and
fast-path all carry the requirement forward so a session cannot be reset out
of its sandbox.
Because exec injects the agent's managed GitHub identity only for gateway-host
execution and blanks GH_TOKEN/GITHUB_TOKEN otherwise, a forced-sandbox session
also loses access to the shared agent's credentials — the property that makes
one shared agent safe for untrusted guests.
Session entries are JSON; no SQLite schema, schema-version, or protocol change.
Docs also correct the sandbox backend list, which omitted the first-class
podman backend.
Teach show_widget authors to fit the host iframe and reflow multi-column content on narrow viewports. Mirror the contract in the public widget guide.
Co-authored-by: roboclaw-bot <309084314+roboclaw-bot@users.noreply.github.com>
* feat(approvals): mint scoped standing grants for cron allow-always
When an operator resolves allow-always for an approval raised by a cron
job's isolated run, the Gateway now mints a scoped standing grant in the
same SQLite transaction that resolves the approval, instead of writing an
unbounded command digest into the JSON allowlist. Subsequent occurrences
of that job execute the exact approved operation (command text, cwd, env
hash) without prompting while the grant revalidates against authoritative
rows: 30-day expiry, revocation, the cron job still existing with the same
config revision, and the minting approval row still holding allow-always
all fail closed back to the normal prompt. Non-cron allow-always behavior
is unchanged.
- New first-use lazy STRICT table operator_approval_standing_grants in the
shared state DB (declared canonically, no schema-version bump; older and
downgraded readers stay valid without it).
- The cron run owner records run -> {agent, job, config revision} in a
process-local registry at run start; exec.approval.request stamps the
cron source and exact operation binding onto the approval at creation,
so nothing is ever inferred from session keys or run ids.
- The gateway exec host consults grants only when policy would prompt;
ask=always, security=deny, mutable file operands, heredoc, strict
inline-eval, and audit-suppression approvals keep prompting. Grant use
updates last_used_at_ms/use_count and emits the exec approval security
event with the grant and minting approval as lineage.
- Abort-wins guard: a run with an abort tombstone never mints.
* feat(approvals): deliver cron exec approvals to approval clients and wait inline
The standing-grant mint path was unreachable end-to-end: #128031 made
cron approval requests register with delivery fully suppressed, so the
shared owner expired them as no-approval-route within milliseconds, and
even a delivered card would have died seconds later when the isolated
run finalized on the approval-pending handoff and authority-close
cancelled the parked approval.
Cron approval requests now carry deliverToApprovalClientsOnly: the
shared delivery owner broadcasts them to connected websocket approval
clients (Control UI, TUI) but skips internal chat approval runtimes,
forwarder/iOS delivery, and turn-source routes, so the per-occurrence
chat spam #128031 removed stays removed. With no approval client
connected, the request still expires no-route into the existing
headless denial. The gateway exec host additionally waits inline for
cron-triggered approvals (the same treatment native chat channels got
in #93918), keeping the isolated run and its delegated authority alive
for the full approval window; cron jobs are single-flight, so at most
one card per job is pending at a time and allow-always ends the
recurrence by minting the standing grant.
Live-proven on a hermetic gateway: card delivered with ~30-minute
window, run waited 72s for the operator click, allow-always executed
the occurrence and minted the grant with no JSON allowlist digest,
the next occurrence ran promptless (use_count 1, no new approval row),
and editing the job failed closed back to a fresh prompt.
* chore(protocol): regenerate Swift models for deliverToApprovalClientsOnly
* fix(approvals): consume standing grants at the spawn boundary; keep node cron headless
Review findings from ClawSweeper on #129526:
- Grant authority is now recorded only at the final effect. The consult
path validates without recording a use and returns a
revalidateBeforeExecution closure (the mutable-file-binding seam) that
consumes the grant immediately before runExecProcess; any invalidation
during awaited pre-spawn work (job edit/delete, revocation, parent
approval reversal) denies with next-step text instead of executing on
stale authority. Regression proves consult leaves use_count at 0 and a
reversed minting approval denies at the boundary.
- Cron approval-client delivery is scoped to host=gateway. Node-host cron
cannot mint or consume grants yet, so it keeps the fully suppressed
headless policy from #128031 instead of raising cards whose allow-always
could not stick; node-host grant support stays a named follow-up.
* test(agents): complete plugin-metadata-snapshot mock factories
Five explicit vi.mock factories for current-plugin-metadata-snapshot.js
exported only getCurrentPluginMetadataSnapshot. Under isolate:false shard
composition the incomplete mock can bleed into siblings that import the
real module — model-resolution-consistency.test.ts failed on CI with
'No withPluginMetadataSnapshotScope export is defined on the mock'.
Spread importOriginal so every binding prod touches stays exported, per
the repo mock-factory rule; only the snapshot getter stays overridden.
* chore: drop accidentally committed pinned swiftlint binary; ignore .build/
scripts/install-swift-tools.sh installs pinned Swift tools into
.build/swift-tools per the lint-swift.sh remediation hint; the 36MB
binary must never ride a commit. Remove it and ignore the directory.
* test(agents): shield model-resolution-consistency from leaked snapshot mocks
The agents-embedded shard still failed after completing five factories:
~20 more test files across the repo mock current-plugin-metadata-snapshot
with incomplete explicit factories, and under isolate:false composition
any of them can strip withPluginMetadataSnapshotScope from this file's
imports. Give the victim a file-local identity mock (importOriginal
spread) that always wins, and revert the static-catalog factory edit that
tipped that grandfathered file over the max-lines cap — the repo-wide
factory completion belongs to a dedicated sweep.
* fix(agents): enforce session permissions for shared terminal input
* chore: keep terminal release notes in pull request
* fix(agents): fence full terminal input to active run authority
## 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).
* 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.
* 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
Remove the bundled OpenProse plugin and /prose command now that upstream owns the maintained Agent Skill. Preserve /prose as migration documentation and let Doctor clean stale plugin configuration.
BREAKING CHANGE: The bundled OpenProse plugin and /prose command are removed.
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
The exec docs claimed a per-call `host=node` request is always allowed
from `tools.exec.host=auto`, gating only `host=gateway` on whether a
sandbox runtime is active. The runtime treats both identically:
`isRequestedExecTargetAllowed` rejects `node` and `gateway` alike from
`auto` while a sandbox is available, and allows both when it is not.
Align the three pages that carried the asymmetric claim with the
test-locked runtime contract, and point readers at the explicit
`tools.exec.host=node` path the rejection error already recommends.
Refs #61009
* fix(dashboard): open widget links through the host instead of sandbox popups
Widget frames granted allow-popups/allow-popups-to-escape-sandbox to every
rendered widget, including grantState "none" widgets that never reach an
operator approval card. CSP has no navigate-to, so a popup bypassed the
connect-src 'none' containment those widgets otherwise run under. Scripted
window.open was blocked only by the document guard's open lock, which is
emitted solely when blockDescendantFrames is set, so two unrelated knobs had
to agree for the boundary to hold.
Restore both sandbox strings and route user-clicked links through the existing
ticket-bound bridge: a trusted click on a ticketed board widget's http(s)
target="_blank" anchor forwards to the Control UI, which opens the tab with
noopener,noreferrer. Links keep working for every rendered widget, the host
owns the scheme filter instead of trusting widget-authored rel attributes, and
scripted popups are impossible by construction rather than by coincidence.
Also collapses the unreachable controlUiBaseUrl branch and corrects both docs
pages, including the claim that links opened for every rendered widget, which
was never true for MCP-App board widgets.
Follow-up to #127315; refs #127314.
* fix(dashboard): open widget links through the Control UI external-url owner
The widget bridge opened links with a local window.open helper, which the
lint:ui:no-raw-window-open boundary guard rejects. Delegate to
openExternalUrlSafe(), the Control UI owner for external navigation: it
applies noopener/noreferrer, detaches the opener handle, and keeps one
canonical open path. The bridge keeps a narrower absolute-http(s) gate,
since widget-supplied links must not reach blob: URLs that the shared
external-link policy permits.
* fix(dashboard): route middle-button widget links and honor canceled clicks
Address ClawSweeper review on #128125.
The wrapper only handled primary click, so with popup permission removed from
both sandbox layers a middle-button activation silently did nothing. It also
listened on capture, so a widget handler's preventDefault could not cancel the
link the wrapper then opened itself.
Match the Control UI activation contract in ui/src/app/native-link-routing.ts:
listen for click and auxclick, gate on !defaultPrevented && !shiftKey &&
!altKey && (click+button 0 || auxclick+button 1), and listen on bubble so
widget cancellation stays effective. A middle click now routes for any widget
link, not just target=_blank, since middle-click always means new tab and
leaving plain links unrouted would keep the same silent-no-op class.
Move the wrapped-document test into src/canvas/wrap.test.ts: it exercises
buildWidgetDocument, so it belongs with its owner, and the added assertions
would otherwise push widget-tool.test.ts past the max-lines cap that AGENTS.md
forbids suppressing.
Reconcile code and command aliases after trusted-policy and hook rewrites so explicit blank or non-string mutations fail closed, including simultaneous valid rewrites. Add owner-boundary regression coverage and document the contract.
Treat blank code and command aliases as absent while preserving mismatch rejection when both aliases contain different instructions. Keep trusted hook and policy rewrites normalized at the Code Mode owner boundary.
Co-authored-by: Marvinthebored <marvin.assistant@lindsey.jp>
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix: surface Codex input prompts across runtimes
Codex structured input now reaches bounded Gateway questions in native and ACP runs, with exact turn ownership, explicit unsupported outcomes, and cancellation fencing. Consume the published ACPX elicitation support.
* chore: align elicitation helpers with current guards
Use protocol-specific helper names required by current main and update the reservation regression to the generalized input owner.
* fix: formalize structured input SDK surface
Expose one documented, frozen agent-harness structured-input contract with runtime and subpath coverage, replacing the accidental function-property API.
* fix: satisfy elicitation architecture gates
Register the real-process ACPX fixture as an executable test root and move shared structured-input types into the boundary leaf to keep Knip and Madge clean.
* fix: remove structured input lint suppression
Preserve the rejected control and invisible-character ranges with an explicit code-point check so the production suppression inventory stays closed.
ruel225