* [AI] fix(context-engine): bound accepted-turn cap to the current turn range
The accepted-turn transcript read applied the 8 MiB / 20k-event cap to
the entire active prefix instead of the admitted turn range, so durable
context engines permanently blocked every later small turn once session
history exceeded the cap. Restrict the cap check to rows at or after the
admission message position while preserving the full messages +
prePromptMessageCount payload contract.
Fixes#121623
Co-Authored-By: Claude <noreply@anthropic.com>
* [AI] test(context-engine): clean up large-prefix regression temp dirs
ClawSweeper P3: the two new accepted-turn regression cases created
multi-megabyte SQLite fixture directories that were never removed.
Track them and delete them in afterEach.
Related to #121623
Co-Authored-By: Claude <noreply@anthropic.com>
* oc-136: make durable turn commits turn-local
* oc-514: use canonical temp-dir cleanup helper
* oc-514.7: version turn-local durable commits
* oc-514.9: preserve legacy durable rows
* oc-514.8: refresh plugin SDK API baseline
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Josh Lehman <550978+jalehman@users.noreply.github.com>
* fix(agents): remove stale media completion reply mechanics
Make generated image, music, and video completion wakes follow the current tool-agnostic visible-reply contract, preserving every structured attachment without NO_REPLY or transport-specific instructions.
Refs #121933.
* test(agents): remove stale media helper
Delete the unused assertion helper left behind by the table-driven generated-media completion coverage.\n\nRefs #121933.
* test(tooling): wait for profiler descendant pid
Require the PID file to contain a valid positive process ID before asserting profiler descendant cleanup, closing the create-before-write CI race.\n\nRefs #121971.
A final reply whose platform send was accepted but whose response was lost
previously ended in silence. Custody that stays unknown after a claimed send
now records durable pendingDeliveryNotice debt; the next same-route turn
delivers one "could not confirm delivery" notice and acknowledges it into the
transcript. Restart recovery completes ambiguous sessions with the same debt
instead of a fire-and-forget notice; the debt survives reset and rollover, and
suppressed notice sends retain it instead of faking delivery. Permanent typed
no-send rejections settle as terminal suppression (no replay, no false
notice); retryable ones restore prepared custody for safe replay. Google Chat
media-only rejections use the typed no-send contract; Telegram native-command
replies join pending-final custody.
Fixes#80362
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Unset heartbeat.target silently collapsed into the explicit "none" opt-out: heartbeats ran every 30m by default, elected notifications were dropped with only an in-memory event, and health read fine. Unset now resolves to "last" (the most recent conversation); explicit target: "none" keeps its internal-only contract. Polls skip pre-model with reason no-route while no route exists yet, and status/doctor surface the waiting-for-route state. Deliberate maintainer-owned default cutover: existing installs without a configured target start receiving heartbeat alerts in their last conversation after upgrade.
Fixes#121880
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Align copyable onboarding commands with the mandatory risk acknowledgement and an explicit health disposition for config-only automation.\n\nRefs #121951.
* refactor(skills): split workspace loading into concept modules
* refactor(skills): delete dead seams and duplicate helpers
* chore(lint): ratchet max-lines baseline after workspace split
* chore(plugin-sdk): refresh API baseline
* refactor(skills): internalize discovery result type
* test(tooling): drop stale unit-fast fixtures
Re-lands the reverted #121507 on the current custody contracts. The reply dispatcher now owns direct-send pending-final custody: claim before provider I/O, terminal settlement for delivered/suppressed/failed outcomes, proven no-send stays replayable, ambiguous evidence fails closed — so Gateway restarts can no longer duplicate an already-accepted final reply.
Proof: ClawSweeper local review clean, exact-head ci-gate green, live Telegram E2E (one turn, one final, no duplicates).
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* feat(ui): edit a queued chat message in place
The pencil on a queued row, or a double-click, lifts the message and its
attachments into the composer. The row stays where it is, marked Editing,
so the position the corrected message will take is visible rather than
promised. The composer shows an edit banner with an X to cancel.
Cancelling touches storage not at all, because the row never left the
queue. Sending retires the original and hands its position to the
replacement, which lands in the same slot. While an edit is open the
drain refuses that row, so a reconnect cannot deliver the text the
operator is visibly rewriting; the queue behind it waits, which is the
same contract the held position promises.
* fix(ui): retire an edited queued row only with its replacement
The send retired the source row first and admitted the replacement second,
so a rejected store write left the operator with neither: the original gone,
the replacement never persisted, and the attachments it dropped already
released. Admission now carries the source id and swaps both rows in one
store write, so a rejected write changes nothing and the edit stays open on
the row that is still there, which is what cancelling already promises.
Filtering the source before the cap check also keeps a replacement
admissible on a full queue.
The edit token recorded only the session key, but an outbox is scoped by
session and agent, and a raw global session keeps its key while the selected
agent changes underneath it. The token now carries the agent that owned the
row and every reader validates the whole scope, so a send after a switch
cannot retire a row in the outbox it left behind.
* fix(ui): hold an edited queued row across every pane
Panes share one outbox and one drain lane while composer state stays per
pane, so the hold that keeps a row from being delivered while it is rewritten
was invisible to the pane that drained it: a session event in a second split
pane resumed every stored outbox, saw no edit, and sent the original text.
The hold is now read across every pane subscribed to the outbox, which keeps
one source of truth — each pane's own scope-validated edit — and needs no
lifecycle of its own, so closing a pane or switching agent releases it by
simply no longer being an active edit.
The payloads an edit owns now travel with its token. The write that admits
the replacement retires the source row and tells every pane synchronously, so
by the time the send released the images the replacement had dropped, the row
holding them was already gone and nothing was released. That leaked the
payload and its object URL for the rest of the browser session.
* fix(ui): release attachments from cancelled queue edits
* fix(system-agent): repair setup-chat execution identity, harness fallback, and error surfacing
Always inject the reserved "openclaw" execution entry for off-roster system-agent runs.
Keep implicit harness runtime selection implicit so Codex refusal falls back to the OpenClaw harness, while explicit policy remains fail-closed.
Log openclaw.chat turn failures and surface their underlying causes in the UNAVAILABLE response.
* chore(plugin-sdk): refresh agent harness API baseline
* test(system-agent): align fixtures with optional harness routes
* fix(system-agent): keep execution roster out of TUI catalog lookup
* refactor(system-agent): keep verified state types internal
* fix(system-agent): separate setup policy and execution config
Recognized workspace paths in chat now read as their basename behind a
file-type glyph, so a file reference is identifiable before it is read.
Markdown, package manifests, TypeScript and other code, TSX/JSX components,
config/data, shell scripts, and images each get their own mark; anything else
falls back to a plain document. Paths sharing a basename keep the smallest
trailing suffix that tells them apart, and Windows paths keep their own
separator.
Classification wins over authoring syntax: a path written in backticks is a
file link first, so it drops the inline-code chip and renders exactly like a
bare path beside it. Code spans that are not file links keep the chip.
The full path stays addressable: it drives the file panel, the new tooltip,
and the message Copy action (which returns the original Markdown).
Author-written labels in [label](path) links are never rewritten, and text
that is not a recognizable path stays plain prose.
The glyph is painted as a masked ::before like the sibling GitHub mark, so it
stays out of the accessibility tree and out of copied text and follows the
link color in every theme. Extension classification moves to one shared
resolver that the file preview modal now uses too, replacing its own
code/text extension set.
* longcat: align baseUrl with docs, fix cacheWrite pricing, add brand icon
- Point baseUrl at the documented https://api.longcat.chat/openai/v1 path
instead of the undocumented unversioned alias route.
- Fix cacheWrite pricing to 0: the LongCat pricing page has no separate
cache-write charge (matches deepseek/moonshot/zai catalog conventions).
- Update the pricing page link to the current docs path.
- Add the official LongCat brand icon (provenance recorded in
ATTRIBUTION.md) and display-name mapping so the Control UI shows proper
branding instead of the letter-badge fallback.
- Migrate the persisted legacy default baseUrl via a plugin-owned doctor
configRepair contract: onboarding persists
models.providers.longcat.baseUrl and the runtime reads the stored value,
so the contract rewrites exactly the former default to /openai/v1,
preserves custom endpoints, and warns via a legacy-config rule.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* fix(longcat): correct provider metadata
* fix(longcat): migrate persisted stock pricing
* style(longcat): format doctor repair
* fix(longcat): preserve doctor migration after compat cleanup
* fix(longcat): avoid map spread in doctor repair
---------
Co-authored-by: yuzehui02 <yuzehui02@meituan.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>
* fix: exec timeout field does not state its unit, so callers pass milliseconds
`exec.timeout` is in seconds, but its sibling `yieldMs` is in milliseconds and
the `process` tool's identically named `timeout` is also in milliseconds. Those
two tools are used together in one workflow: exec backgrounds a long command,
then process polls it.
The unit exists only in the field description, and code mode deliberately defers
descriptions - the model sees `timeout?: number` and nothing else. A caller that
guesses milliseconds passes 900000 intending 15 minutes and gets ~10 days, so
the command never times out.
Add `timeoutSeconds` as the canonical field, matching the convention already
used by the config default (`tools.exec.timeoutSeconds`) and the sibling
(`yieldMs`). Keep `timeout` as a deprecated alias; `timeoutSeconds` wins when
both are supplied. Project the new field onto `nodeExecSchema`, which
hand-picks its properties.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix(ci): satisfy test-type and dependency checks
Two CI failures from the test approach, not the change itself.
- check-test-types: TypeBox's TOptional does not surface `description` on its
static type, so the schema assertions failed under tsgo. Read it through a
narrow accessor instead.
- check-dependencies: knip flagged resolveExecTimeoutSeconds as an unused
export because only the test imported it. Move it beside the schema it
resolves, where the exec runtime imports it, so it is production-reachable.
Verified with the same commands CI runs: `run-tsgo.mjs -p
test/tsconfig/tsconfig.core.test.json` reports no errors in these files, and
knip --production reports no unused exports.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix(ci): fit the exec description budget and satisfy oxlint
- `agent-tools.schema.test.ts` enforces a 550-char budget on the exec tool's
model-facing descriptions; the two new ones pushed it to 640. Shortened both.
The rename is what makes this possible: the unit now lives in the field name,
so the description no longer has to carry it. Extended that budget test to
pin `timeoutSeconds` alongside `timeout`.
- oxlint `no-base-to-string`: the test's description accessor typed the field as
`unknown` and stringified it. Typed as `string | undefined` instead.
Verified: budget test passes (137 tests green across both suites), and all six
behavior tests still fail against origin/main.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix: remove the unit-ambiguous exec timeout field instead of deprecating it
Owner decision on the review's "ambiguous model-visible alias" finding: drop
`timeout` entirely rather than keep it as a deprecated alias.
Keeping it visible defeated the purpose. Code mode renders property names and
types and defers descriptions, so a model would still be offered a bare
`timeout` sitting next to a millisecond-based `yieldMs` and could still pick it.
A deprecation marker in a description the model never sees changes nothing.
Backward compatibility is not required here: every tool call is constructed
fresh by the model from the current schema, and a repository search found no
internal caller passing `timeout` to the exec tool. `timeoutSeconds` is now the
only exec timeout field, on both the main and node-only surfaces.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix: migrate the internal exec callers to timeoutSeconds
Removing `timeout` from the exec schema broke three real internal callers that
the earlier survey missed, plus two test files and the runtime's own reads.
`check-test-types` caught it; my local search had filtered `timeout:` matches to
lines mentioning "exec" and discarded these.
- src/auto-reply/reply/bash-command.ts
- src/auto-reply/reply/commands-diagnostics.ts
- src/auto-reply/reply/commands-export-trajectory.ts
all passed `timeout: timeoutSec` and now pass `timeoutSeconds: timeoutSec`.
- bash-tools.exec-run.ts reads the field through the same typeof narrowing the
third call site already used; `params` is loosely typed, so a raw read is
`unknown`.
- Test fixtures updated. A vitest `waitFor({ timeout })` option in
background-abort.test.ts is NOT an exec param and was left alone.
`run-tsgo.mjs -p test/tsconfig/tsconfig.core.test.json` now reports zero errors
in changed files; the 5 remaining `rejectSymlinks` errors are pre-existing on
main. 148 tests pass across the affected suites.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* docs: synchronize the exec config table with timeoutSeconds
Addresses the P2 finding. The tools.exec.timeoutSeconds row still described the
per-call override as `timeout` and `timeout: 0`, which no longer exist. The
only remaining bare `timeout` reference in this page is the deliberate
contrast with the process tool's millisecond field.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix: migrate the QA lifecycle caller and the shared ExecToolArgs contract
Two findings from the re-review of 6e8c3560e7, both real.
[P1] test/e2e/qa-lab/runtime/openclaw-exec-process-lifecycle.e2e.test.ts:174
still passed `timeout: 0.05`. After the removal that field is accepted and
ignored, so the call silently used the default timeout and could not produce
the asserted 50 ms `overall-timeout` result. Migrated.
[P2] ExecToolArgs in bash-tools.exec-request-preparation.ts:33 still declared
`timeout?: number` while the runtime reads only `timeoutSeconds`, offering
request-preparation users an accepted-but-inert argument. Renamed.
Why the typecheck missed both: ExecToolArgs is
`Record<string, unknown> & {...}`, and that index signature defeats excess
property checking, so a stale member produces no error at any call site. The
e2e file also sits outside the core test tsconfig I had been running locally.
Left alone deliberately: `processTool.execute({ action: "poll", timeout })` is
milliseconds and correct, and `createBashTool` in sessions/tools/bash.ts is a
separate tool surface with its own timeout.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix(test): update the node exec tool property assertion to timeoutSeconds
src/gateway/tool-resolution.exclude.test.ts:413 asserts the exact property list
of the node-forced exec tool and still expected "timeout". Renamed to match
nodeExecSchema.
Found by checks-node-compact-large-5; my local core-test tsconfig does not cover
src/gateway, so it did not surface there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix: accept the legacy exec timeout as an undocumented alias
Removing the field from the schema does not reject it at runtime: the exec
schema accepts unknown properties, so a call built against an older schema was
accepted and then silently ignored, taking the default deadline instead of the
one it asked for.
Route every read through resolveExecTimeoutSeconds, which prefers the canonical
timeoutSeconds and falls back to a numeric timeout. The alias stays out of
execSchema, nodeExecSchema and ExecToolArgs, so no new caller can adopt it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* docs: update the background-process guide to timeoutSeconds
The exec parameter table and the inheritance note still taught `timeout` and
`timeout: 0`, which kept the retired spelling publicly discoverable and
contradicted the exec guide.
`poll`'s own millisecond `timeout` is left alone; it is a different tool and a
different unit, which is the collision this rename removes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XQojYRCUdXhcKFeu4QW5Gh
* fix(exec): reject removed timeout field
* chore: regenerate plugin SDK API baseline
---------
Co-authored-by: Marvinthebored <262704729+Marvinthebored@users.noreply.github.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Co-authored-by: Patrick Erichsen <patrick.a.erichsen@gmail.com>
Queue position becomes an explicit, operator-owned fact on ChatQueueItem
instead of an accident of arrival time and storage array order. One
comparator now serves the visible projection, the drain head selection,
the steer rebuild, and the alias merge, so what the composer shows is
what the Gateway receives.
Reordering lives entirely on one handle at the left of a queued row:
drag it, or focus it and press the arrow keys. Keyed rows keep focus on
the handle as it moves. Rows already handed to a run keep their place.
Also routes .chat-image-action through var(--cursor-action); its
hardcoded pointer was failing the cursor-policy test on main.
* feat(ui): mark GitHub links in Control UI chat
Chat markdown anchors pointing at github.com now carry a decorative GitHub mark before their text, so a bare URL and a [#3434] shorthand read as the same destination. The mark is painted from the bundled brand path through a CSS mask on currentColor: it stays out of the accessibility tree and out of copied text, tracks the chat text size, and follows the link color in every theme. Image-only links such as badges and code spans stay undecorated.
* feat(ui): break long autolinked URLs at any character in chat
An autolinked URL is one unbreakable word, so the inherited overflow-wrap only splits it after every other option is exhausted: a long URL was pushed whole onto the next line, and with the new GitHub mark ahead of it the mark was left stranded alone at the end of the previous line. Autolinked anchors now carry markdown-bare-url and get unconditional character break opportunities, so the URL fills the line it starts on. Authored link labels keep word-boundary wrapping.
* fix(ui): drop resting underline on chat GitHub links
* test(ui): split markdown link tests into a sibling file
markdown.test.ts crossed the 1000-line oxlint budget once the GitHub-link
cases landed. Move the link-rendering describes (autolinks, explicit
protocol links, file links, bare-url marks, GitHub marks) into
markdown-links.test.ts. Pure move; no assertion changes.
* test(ui): assert GitHub links rest without an underline
Covers the resting/hover decoration contract through the real rendered
page, and pins the rule's scope with a non-GitHub link that keeps the
base underline.
* feat(agents): rename spawn_task tool to suggest_task
Models across vendors read the name spawn_task as 'starts background
work' and refuse to call it when the user has not asked to start
anything (Claude Sonnet 5, verbatim: 'spawn_task (which actively starts
background work)... Since you didn't ask me to start work, I'm flagging
this as a follow-up instead'). The tool only records a suggestion card,
so the name defeated the feature.
Rename the model-facing tool to suggest_task everywhere (catalog,
descriptors, presets, display config, OpenClawKit tool-display, docs)
and migrate persisted operator state via doctor: config tool policies
and cron SQLite toolsAllow entries that reference spawn_task are
rewritten to suggest_task. No runtime alias remains; runtime exposes
only the canonical name.
Live clean-room matrix on a dev gateway (virgin session per cell):
suggested-task-card phrasing now yields exactly one card on
anthropic/claude-sonnet-5, claude-haiku-4-5, openai/gpt-5.6-sol,
gpt-5.6-luna, and google/gemini-3.1-pro; pre-rename the same bait
produced prose-only follow-ups.
* chore(doctor): keep legacy tool-name predicate module-local
isLegacyTaskSuggestionToolName has no external consumers; the exported
form tripped the knip unused-export gate (deadcode:exports).
* fix(doctor): keep tool-name migration out of plugin-owned config
The rename migration traversed the whole raw config, so any object
carrying a toolsAllow key was rewritten - including opaque
plugins.entries.*.config, which core must not mutate. Scope the apply
pass to the typed core roots (tools, agents, channels, gateway) and pin
plugin config preservation in the test. Plugin-owned tool lists migrate
via the owning plugin's doctor contract instead.
* fix(ci): heal main deadcode and stale Swift protocol mirror
Unrelated main breakage blocking this PR's merge gate, healed here per
landing policy:
- #121653 left dead exports in src/skills/workshop: make
removeSkillCollectionDirectory and SkillCollectionRestoreResult
module-local; register collection-review.ts under the documented knip
ignoreIssues test-only-export convention (production runs it via the
scheduled maintenance loop).
- #121673 added ApprovalResolveParams.reviewer without regenerating the
Swift mirror; regenerate GatewayModels.swift (additive only).
* fix(ci): satisfy generic approval-runtime request signature in resolver test
Third main-heal carried by this PR: #121673 (9935ca3b30) left the
scoped-request mock's concrete inferred type unassignable to the generic
GatewayNativeApprovalRuntime request signature, breaking check-test-types
on main. Cast the mock at the runtime literal; assertions keep the Mock.
* chore(ci): drop approval-test cast superseded by main's typed mock rewrite
* chore(ci): drop duplicate knip entry superseded by main's heal
Peter Steinberger