The panel from #125199 rendered raw wire enums in code tags, used
form-grid/field markup whose styles the agents page never loads (bare
unstyled inputs outside the card), referenced non-existent avatar
classes, and put callouts inside the settings group. Rebuild it on the
canonical settings primitives: status rows with dot status and friendly
source/evidence labels, a System/This Agent segmented scope control,
the settings secret input (gains a disabled prop), a danger status row
for errors, and a quiet-inherit + primary-save action row. Retitle to
"GitHub Identity" and move it below Tool Access and Available Right
Now. Add a .settings-account primitive for the 20px round avatar.
Live-tested on an isolated dev gateway: native-credential verify via
the GitHub API, segmented scope switching, and the empty-token error
row. Controller behavior unchanged.
* feat(agents): unify agent status into a durable progress_card
Replace the write-only update_plan to-do tool and the fragmented plan
rendering with one durable status artifact per session: progress_card
({plan?, markdown?}, replace-on-write, 8 KiB markdown / 50-step caps).
Cards persist in a lazy-additive session_progress_cards table in the
per-agent DB (no schema-version bump), broadcast progressCard.changed,
and render from the store with exactly one live placement per view
(session rail when visible, else the composer-adjacent bar); transcripts
collapse to one-line receipts, and the sidebar hovercard shows other
sessions' cards inline (markdown + <progress>, DOMPurify allowlist, no
iframes). The three stream-derived plan renderers and their dedup
heuristics are deleted.
Codex runs disable the native plan tool per thread
(tools.update_plan.enabled=false) and receive progress_card via the
dynamic-tool bridge; compaction restore now reinjects the card (steps +
bounded markdown). Card writes still emit the legacy plan stream event so
native apps and channels keep working until their per-platform
migrations. Policy names map update_plan -> progress_card; the shipped
tools.updatePlan=false kill switch is honored.
Net -277 production LOC; -480 test LOC.
* test(agents): regenerate Codex prompt snapshots for update_plan thread-config disable
* chore(protocol): allowlist progressCard.changed for native apps pending card migration
* fix(ci): repair progress card integration checks
* fix(codex): canonicalize native progress cards
* test(gateway): reconcile progress card method order
* test(codex): stabilize native approval fixture
* docs(gateway): document portal tool gating and tighten its description
Portals need no dedicated config key, but nothing said so: add an
availability section covering the coding-profile/group:ui default,
sandbox and owner-only gating, tools.deny recipes, and the fact that
portal listeners inherit the Gateway bind interfaces. The group:ui
roster in config-tools was also missing portal and dashboard.
The tool description named the exec tool, which repo policy forbids
because gating can remove it and leave a dangling reference; it now
states the capability instead, and is ~18% shorter.
* docs(gateway): use the global tools policy for the fleet-wide portal deny
The canonical fleet-wide knob is top-level tools.deny, not the agent
defaults layer; match the json5 style the tools docs already use.
* 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
* feat(gateway): task-suggestion acceptance modes with compact split-button card
- taskSuggestions.accept gains additive mode (worktree default | local | cloud | session) + cloudProfileId; operator.admin unchanged; worktree requests stay {taskId}-only for older gateways
- local reuses the plain-cwd create; cloud mirrors the Control UI create-empty -> sessions.dispatch -> send sequence with pre-claim profile validation and full rollback; session delivers the prompt into the source session (steer single active run, plain turn when idle, ambiguous refused) and can only cancel the chip on failure, never touch the source session
- Control UI card redesigned per owner reference: repo-basename eyebrow, clamped summary, instructions expander, split button with per-profile cloud items or a disabled "No cloud environment configured" hint
- TUI gains the same modes as selector rows with capability gating and double-Enter arming
Release-note context: operators now choose where an accepted follow-up runs — isolated worktree (default), the project checkout, a configured cloud worker, or as the next instruction in the current session — from a compact suggestion card.
* fix(gateway): negotiate task-suggestion acceptance modes and enforce steerable-run invariant
- Gate non-worktree task-suggestion actions on the negotiated Gateway capability.
- Require exactly one steerable run ID before delivering into an active source session.
* fix(ui): hide inert task-suggestion chips and strip bidi controls
- render no suggestion card for operators who can neither accept nor dismiss (matches TUI behavior)
- strip RTL/bidi control characters from chip title, summary, project path, and instructions before rendering
- document the git-checkout requirement for spawn_task project directories (gateway enforcement landed in #120940)
* fix(ui): gate task-suggestion actions on negotiated gateway methods
Mirror the TUI capability rule: accept requires the taskSuggestions.accept method plus operator.admin, dismiss requires taskSuggestions.dismiss plus operator.write. List-only gateways render no chip; e2e fixtures now advertise the mutation methods they exercise, and a new regression proves a pending suggestion renders no chip when only listing is advertised. Addresses the ClawSweeper P2 on this PR.
* docs(gateway): document llamacpp toolSchemaProfile for custom llama.cpp endpoints
Custom openai-completions providers pointed at a llama.cpp / llama-server
endpoint bypass the GBNF-safe schema normalization that built-in llama-cpp,
ollama, and lmstudio providers apply automatically. The toolSchemaProfile:
"llamacpp" switch activates it, but the supported value was never named in
the docs, so users hit llama.cpp 400 'failed to parse grammar' on tools with
large minLength/maxLength (e.g. cron trigger.script, maxLength 65536).
- Expand the toolSchemaProfile row to name the values recognized on a model's
compat (llamacpp, gemini) and when llamacpp must be set explicitly;
cross-reference the unsupportedToolSchemaKeywords escape hatch.
- Add a 'Local models (llama.cpp / llama-server)' provider example showing
the required compat.toolSchemaProfile: "llamacpp" on a custom endpoint.
Related to #117070
Co-Authored-By: iCodeMate <noreply@anthropic.com>
* docs(gateway): describe llamacpp profile as stripping, not clamping
Address review feedback on #117120: the fallback sentence claimed the
llamacpp profile "also clamps" constraints, but cleanSchemaForLlamacppGbnf
only removes `pattern` outright and drops `maxLength` at or above the
2000-repetition threshold (LLAMACPP_GBNF_MAX_REPETITION_THRESHOLD). It
never clamps a value, and small bounds like `minLength: 1` are preserved
(confirmed by the projection test at src/plugin-sdk/provider-tools.test.ts).
Reword the escape-hatch sentence to describe the implemented stripping
behavior and contrast it with the unconditional keyword-list drops, so
operators do not expect a preserved bounded constraint.
No code or behavior changes; documentation only.
* docs(gateway): scope llamacpp profile description to its actual transformations
Address follow-up review on #117120: the toolSchemaProfile table row
described `llamacpp` as "rewriting tool schemas into the JSON Schema
subset llama.cpp can compile to GBNF," which overstates what
cleanSchemaForLlamacppGbnf does. The cleaner only removes `pattern`
outright and drops `maxLength` at or above the 2000-repetition threshold
(LLAMACPP_GBNF_MAX_REPETITION_THRESHOLD); it leaves every other keyword
and `minLength` untouched, and those are not guaranteed GBNF-safe.
Reword the row to state the exact transformation contract and explicitly
note it is a targeted compatibility cleaner, not a blanket GBNF-safety
guarantee. No code or behavior changes; documentation only.
Co-Authored-By: iCodeMate <noreply@anthropic.com>
* docs(gateway): tighten llama.cpp profile contract
* docs(gateway): order local model examples
---------
Co-authored-by: iCodeMate <noreply@anthropic.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* feat(tools): per-model code-mode capability flags and auto master-switch tier
* feat(anthropic): flag claude-sonnet-4-6 as code-mode preferred
* revert(anthropic): drop claude-sonnet-4-6 from the code-mode preferred set
* test(anthropic): type the manifest compat field in the catalog contract test
* fix(anthropic): carry catalog compat onto hand-built forward-compat model rows
The update_plan checklist tool has been default-on with no model gating for a
while, so keeping it under tools.experimental misrepresented it and kept a
container key alive for a single boolean. Rename it to a plain boolean
tools.updatePlan (camel-cased tool id, matching tools.agentToAgent and
tools.toolSearch), keeping the !== false default-on semantics.
The strict schema rejects the retired key, so openclaw doctor --fix now lifts
tools.experimental.planTool to tools.updatePlan and deletes the emptied
container. Registration also drops five parameters that were never read
(agentSessionKey, agentId, modelProvider, modelId, pluginToolAllowlist).
Docs: config-tools.md dropped a stale claim that the tool defaults off behind a
strict-agentic GPT-5 rule that no longer exists, experimental-features.md no
longer lists the tool and now cites the real Codex app-server floor (0.143.0,
per MIN_CODEX_APP_SERVER_VERSION).
Config baseline core count drops 2307 -> 2306.
* feat(sdk): always persist media facts and ship facts-first replacements for legacy Media* surfaces
PR 1 of the media legacy retirement program (audit-frozen, 4 PRs).
- Every media-bearing user turn now persists normalized __openclaw.media
facts unconditionally while continuing to emit the legacy top-level
Media* projection byte-identically (dual-write bridge; the conditional
shouldPersistStructuredMediaEntries gate now always includes media).
- New replacement APIs, shipped before any removal: typed hook media
facts (media[], originalMedia[], mediaStagingPending) on message
events; {{AttachmentPath}}/{{AttachmentUrl}}/{{AttachmentContentType}}/
{{AttachmentDir}}/{{AttachmentIndex}} template variables; focused
openclaw/plugin-sdk/media-local-roots subpath split out of the
deprecated agent-media-payload facade.
- Every legacy surface carries @deprecated naming its replacement, under
one named compatibility record media-legacy-projection with the
operator-approved removeAfter 2026-10-01 (two release trains; deletion
additionally gates on a clean published-plugin artifact sweep).
- Generic transcript append invariant documented; SDK migration, hooks,
and configuration docs updated to the facts-first path.
Writer golden matrix proves legacy bytes and model prompt bytes are
unchanged while nested facts become unconditional. 2,189 broad media
tests green; SDK api-baseline regenerated on fresh-env Testbox.
* feat(sdk): register media-local-roots subpath exports and deprecation metadata
Completes PR 1: package export map for openclaw/plugin-sdk/media-local-roots
plus the deprecated-subpath inventory and doc metadata entries for the
media-legacy-projection record.
* chore(sdk): track media-local-roots entrypoint and deprecated-export budgets
* fix(sdk): keep deprecated MSTeams buildMediaPayload re-export through the compat window
Deleting shipped runtime-api re-exports belongs to retirement PR 4 after
the media-legacy-projection window; PR 1 only deprecates. Also formats
the migration-guide schedule table.
* docs: regenerate docs map for media migration additions
Fix MiniMax-M3 Anthropic-compatible requests so OpenClaw no longer sends the disabled-thinking payload that makes M3 return empty content. M3 defaults now stay on MiniMax's omitted/adaptive thinking path, explicit `/think off` is still respected, and MiniMax-M2.x keeps the disabled-thinking default that prevents reasoning_content leaks.
Also wires the MiniMax thinking policy through bundled provider-policy loading so pre-runtime and configless embedded-agent paths resolve the same defaults.
Thanks @IamVNIE for the live MiniMax API repro and initial patch.
Forward initial image/file attachments when spawning ACP subagents through the existing sessions_spawn attachment opt-in. Remove the PR-only acpEnabled config split so ACP uses the same attachment gate as other runtimes.
Also fix the PR branch CI fallout: type the browser element CLI request mock and use Vitest env stubs in the Azure speech test to satisfy the changed-path security scan.
Verification:
- GitHub CI passed on f6ca26b160.
- Autoreview clean.
- Crabbox AWS live OpenAI proof passed: cbx_a576d49493fe / run_081dcc6c6a1b.
Thanks @zhangguiping-xydt.
Move vLLM Qwen thinking control onto configured model compat metadata and carry it through catalog/model-selection/runtime thinking contexts.
Also migrate legacy provider/default request params in doctor and keep Pi/runtime model rows buildable with explicit reasoning defaults.
Thanks @rendrag-git.
Co-authored-by: rendrag-git <253747599+rendrag-git@users.noreply.github.com>
Peter Steinberger