* fix(anthropic): delegate adaptive CLI effort
Strip static --effort args for adaptive runs so Claude Code resolves effort from its environment, settings, and model default. Preserve configured effort for off or absent thinking and replace it only for concrete OpenClaw levels.
Fixes#103245
Co-authored-by: Dan Korotin <korotin.daniil@gmail.com>
* fix(anthropic): make effort dispatch exhaustive
---------
Co-authored-by: Dan Korotin <korotin.daniil@gmail.com>
* feat(agents): derive a provider-declared default utility model when unset
When agents.defaults.utilityModel is not set, utility tasks (titles, progress
narration) now use the primary provider's declared small model
(modelCatalog.providers.<id>.defaultUtilityModel: OpenAI -> gpt-5.6-luna,
Anthropic -> claude-haiku-4-5). Auth is inherent because derivation follows
the agent's primary provider. Setting utilityModel to an empty string
disables utility routing entirely; narration turns on automatically when a
default resolves and stays off otherwise.
Formatting/lint/tests verified on Testbox (oxfmt, oxlint, plugins:inventory,
docs:map, import-cycles, check:test-types, 4 test files).
* fix(ai): drop the temperature parameter for models that reject it
The GPT-5.6 family 400s on temperature via the Responses API (live-verified:
gpt-5.6-luna/-terra reject it; gpt-5.5 and gpt-5.4-mini/nano accept it).
supportsOpenAITemperature gates all three OpenAI payload builders, with a
catalog compat override (compat.supportsTemperature) declared for the 5.6
family in the openai manifest. Without this, utility tasks (titles,
narration) would fail once gpt-5.6-luna becomes the derived default.
Gates on Testbox: oxfmt, oxlint, plugins:inventory, import-cycles,
check:test-types, 6 test files. Live proof on Testbox: gpt-5.6-luna and
claude-haiku-4-5 one-shot completions succeed with temperature requested.
* fix(agents): carry the primary model's auth profile onto the derived utility default
A primary like openai/gpt-5.5@work previously reached utility tasks via the
profiled fallback; the derived default shares the provider, so its trailing
auth profile carries over instead of silently switching to default
credentials (Codex review). Gates on Testbox: oxfmt, check:test-types, tests.
* docs: realign the manifest provider-fields table after adding defaultUtilityModel
Kind-aware tool rows (terminal-style commands with wrapper stripping and
display highlighting, file edits with inline numbered diffs and diffstat,
write previews, key-value args), aggregate group summaries with live run
status, and a new batched chat.toolTitles gateway RPC that titles complex
calls via the configured utilityModel or the OpenAI Luna default (gated to
OpenAI-primary agents, cached in the per-agent SQLite cache_entries).
Also fixes two transcript pairing bugs: result blocks now inherit call
id/name/details at merge time, and results pair with any open call in the
current tool run so parallel calls render as single rows.
Fixes#103554
* fix(agents): gate empty-error-retry on per-attempt side effects, not cumulative (#97877)
Derive currentAttemptReplayMetadata from per-attempt fields via
buildAttemptReplayMetadata so the retry gate can distinguish "this
model call had no tool side effects" from "cumulative session
replayMetadata carries prior-turn history".
Previously raw.replayMetadata was saved as currentAttemptReplayMetadata,
but attempt.ts already merges observedReplayMetadata with accumulated
session state before returning, so both fields carried the same
cumulative value — making the per-attempt check ineffective.
shouldRetrySilentErrorAssistantTurn prefers currentAttemptReplayMetadata
when available, falling back to replayMetadata for legacy harnesses.
Add regression coverage:
- 22→23 empty-error-retry tests (prior-side-effects + clean 5xx scenario)
- 111→114 incomplete-turn tests (currentAttemptReplayMetadata unit tests)
* fix(agents): harden silent retry replay gating
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
The compaction safeguard treated a successful keyless auth resolution
(ok: true with neither apiKey nor headers) as missing credentials and
cancelled compaction. Bedrock providers using auth: "aws-sdk" sign
requests with SigV4 at send time and legitimately return no static key
or header, so every message on a session large enough to need compaction
failed with a false "could not resolve request credentials" error.
Trust the registry's ok:true success signal; it already returns ok:false
when auth genuinely cannot resolve. Add regression coverage for the
keyless SDK-managed auth path.
Co-authored-by: Roy Osherove <575051+royosherove@users.noreply.github.com>
* fix(signal): clean up signal-cli download temp dir on every exit path
installSignalCliFromRelease created a temp dir with fs.mkdtemp for the
download archive but never removed it. Every return path — including the
success return and the extraction-failure / binary-not-found errors — leaked
the directory plus the downloaded archive (signal-cli Linux-native archives
are ~50-130 MB). Repeated installs or retries accumulated multi-hundred-MB
temp dirs under the preferred tmp dir.
Wrap the body after mkdtemp in try/finally that recursively removes the temp
dir. The extracted binary is installed under CONFIG_DIR and is unaffected.
Co-Authored-By: Claude <noreply@anthropic.com>
* style(signal): drop unnecessary String() conversion in test spy
dir is already a string (mkdtemp return); String(dir) triggers the
no-unnecessary-type-conversion lint rule.
Co-Authored-By: Claude <noreply@anthropic.com>
* test(signal): cover success-path temp dir cleanup with real archive
Add a success-path test that mocks the network fetch to return a real tar.gz
archive (containing a signal-cli binary) so installSignalCliFromRelease runs the
full download -> extract -> find-binary -> chmod path against real bytes, then
asserts the temp dir and downloaded archive are removed by the finally clause.
This complements the existing error-path test so both exit paths are covered.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(signal): use owned temp download workspace
* fix(signal): allow bounded native extraction
---------
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(apple): prevent stale model thinking state
* test(apple): complete attachment transport fixture
* fix(apple): scope model patch ordering by route identity
* chore(apple): sync native CI metadata
* fix: normalizeMediaSource should handle uppercase FILE:// URI schemes
URI schemes are case-insensitive per RFC 3986. normalizeMediaSource
only matched lowercase 'file://', so MEDIA:FILE:///tmp/generated.png
was left in assistant text instead of being extracted as a local
media attachment.
Replaced startsWith + replace with a case-insensitive regex to handle
FILE://, file://, and any mixed-case variant.
Fixes#103473
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(media): handle spaced uppercase file URLs
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(discord): keep thread-binding persona UTF-16 safe
resolveThreadBindingPersona truncated the persona label with a raw .slice(0, THREAD_BINDING_PERSONA_MAX_CHARS). When an emoji or other astral code point straddles the 80-unit limit, the raw slice keeps a dangling high surrogate, so the Discord thread-binding persona (webhook display name) can carry a lone surrogate. Route the clamp through the shared truncateUtf16Safe primitive so the boundary code point is dropped whole. Adds a regression test.
* test(discord): assert exact persona boundary
* docs(changelog): credit Discord persona fix
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* feat(gateway): session worktree targeting, branch listing, and group catalog
- sessions.create accepts worktreeBaseRef/worktreeName (write scope) and
execNode (admin); worktree binding persists on the session entry as
worktree { id, branch, repoRoot } and projects onto session rows with
execNode so UIs can show checkout/branch/node state
- new worktrees.branches RPC lists local/remote branches (no fetch) for
base-ref pickers; worktrees.remove now reports snapshotError
- sessions.delete reports preserved dirty checkouts as worktreePreserved
- gateway-owned session group catalog in the shared state DB with
sessions.groups.list/put/rename/delete; rename/delete update member
categories server-side without bumping recency; sessions.patch absorbs
ad-hoc categories into the catalog
- group session display names prefer the human chat title (subject or
space #channel) over stored compact tokens
Part of #103431
* fix(gateway): route group category updates through the session accessor and regen protocol models
- bulk member-category rename/clear uses applySessionEntryReplacements
instead of legacy updateSessionStore call sites (session accessor
boundary guard)
- regenerate Swift GatewayModels for the worktrees.branches schemas,
snapshotError, and new sessions.create params
Part of #103431
* fix(gateway): resolvable remote branch refs and workspace-scoped branch listing
- worktrees.branches returns remote-only branches remote-qualified
(origin/feature-a) so every advertised name works as a worktree base ref
- write-scoped worktrees.branches callers are limited to configured agent
workspaces; other host paths require operator.admin, matching the
sessions.create cwd bar
Part of #103431
* fix(gateway): guard worktree name reuse by owner and env-scope group transactions
- managedWorktrees.create rejects a caller-chosen name whose live or
restorable record belongs to a different owner, so write-scoped
sessions.create cannot bind a session into another session's or a
manual checkout
- session group catalog writes run their SQLite transaction on the same
env-scoped handle as their statements, keeping OPENCLAW_STATE_DIR
overrides atomic and away from the default state DB
Part of #103431
