…and timers (lobehub#815)

* fix(dashboard): prevent memory leaks from unbounded polling, caches, and timers

- Configure QueryClient with gcTime=2min, staleTime=30s, and
  refetchIntervalInBackground=false to stop all 11 polling queries
  from firing in hidden browser tabs
- Cap useInfiniteQuery to maxPages=5 (250 entries) to prevent
  unbounded page accumulation in the logs table
- Replace per-instance setInterval in RelativeTime with a shared
  10s tick emitter (O(1) timers instead of O(n))
- Migrate SessionCache from unbounded Map to TTLMap with LRU
  eviction (activeSessionsCache maxSize=100, sessionDetailsCache
  maxSize=10000)
- Hoist formatResponseTime to module scope to avoid per-render
  closure allocation
- Add refreshWhenHidden=false to big-screen SWR configs

* fix: address PR lobehub#815 review comments

- Move misplaced JSDoc from formatResponseTime to calcPercentageChange
- Log errors in shared-timer catch block instead of silently swallowing
- Use @/ path alias for ttl-map import in session-cache

* fix: address remaining PR lobehub#815 review comments

- Extract QUERY_CLIENT_DEFAULTS from AppProviders so test imports the
  actual config instead of duplicating it (bugbot lobehub#5)
- Store shared-timer state on globalThis for HMR safety, matching the
  pattern already used in session-cache (bugbot lobehub#6)
- Rename clearAllSessionsCache -> clearAllSessionsQueryCache and
  clearAllSessionCache -> clearAllCaches to eliminate single-char
  naming confusion (bugbot lobehub#7)

Groundwork for integrating lobe-vidol as per-agent 3D portraits.
Resolves 6 design decisions (storage, manifest shape, components,
animations, stages, performance) and lays out a bite-sized plan.

No implementation yet — scaffold component follows in a separate
commit; real lobe-vidol integration is blocked on the
tier->runtime.family rename landing and operator approval for the
npm dep.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ial)

Lands the safe-to-ship shell-side portion of sub-project lobehub#5 without npm
dep additions or home-repo edits. The lobe-vidol / @pixiv/three-vrm plug-in
point is marked in src/features/VRMAvatar/components/Scene.tsx — the
outer component shape (size, idle, variant, staticFrame) is final.

Changes:
- types.ts + normalise.ts — accept string/null/object 'vrm' wire shapes
- server/resolver.ts + /api/avatar/[slug]/route.ts — resolver API
  (static-mode, manifest passthrough, MinIO signing via existing SDK)
- useVRMBinding hook — real fetch to /api/pantheon/v1/agents/<slug>/vrm
  with soft fallback to mock catalog when control plane is unreachable
- concurrency.ts — 2-slot cap for animated VRMs (design §6)
- useInViewport + usePrefersReducedMotion — pure-DOM lazy-load hooks
- index.tsx — VRMAvatar + VRMAvatarChip with React.lazy Scene, cap,
  reduced-motion, and forcePlaceholder support
- stages.ts — Matrix-room-alias → background registry
- ChatHeader Tags/index.tsx — VRMAvatarChip injected behind
  NEXT_PUBLIC_VRM_AVATARS feature flag
- Tests: 32 across 7 files (was 3). All green.
- Plan doc — progress ledger and deferred items.

Deferred:
- Task 1: home-repo agent.schema.json object-form upgrade — home repo
  is read-only for this session per exclusion rules.
- Task 5 npm deps (@pixiv/three-vrm, lobe-vidol): blocked on operator
  approval + network; plug-in point documented in Scene.tsx.
- Task 8: AgentGroupAvatar + modal injection; scope limited to header.

Verify:
  pnpm vitest run src/features/VRMAvatar/   -> 32 passed
  pnpm exec tsc --noEmit --skipLibCheck     -> zero new errors

No systemd restart required (shell-only change).

…obehub#5

three@0.181 was already installed alongside @react-three/fiber/drei.
This adds the VRM loader plugin (@pixiv/three-vrm@3.5.2) and animation
runtime (@pixiv/three-vrm-animation@3.5.2) needed to replace the
mountThreeVrmScene stub in src/features/VRMAvatar/components/Scene.tsx.

No lockfile changes committed (workspace lockfile is not tracked).

Injection surface #2 of 2 for Sub-project lobehub#5 VRM avatars. Adds
<VRMAvatar size={128}> to features/AgentHome/AgentInfo (the welcome
card the user first sees when entering an agent session) behind the
same NEXT_PUBLIC_VRM_AVATARS flag as the chat-header chip.

Design spec §3 calls out two injection points: chat-header chip (Tags,
already landed) and agent-select. The functional equivalent of
agent-select in the current UI is AgentHome/AgentInfo — the avatar
card shown before the first message and re-shown when changing agents.
Inbox agents are excluded (no manifest binding).

Tests expanded 32 -> 47:

- Scene.test.tsx (new, 5): placeholder rendered, error status set in
  happy-dom (no WebGL), size propagation, stage background URL,
  gradient fallback for unknown stage.
- VRMAvatar.test.tsx (+4 = 8 total): resolver-API happy path shown
  in placeholder, prefers-reduced-motion skips scene, IntersectionObserver
  gating keeps off-screen avatars as placeholders, 3x concurrency cap.
- injection.test.tsx (new, 6): AgentHome/AgentInfo flag-off and flag-on
  renders, inbox-excluded case; feature-flag predicate isolation tests
  (NEXT_PUBLIC / VITE_PUBLIC parity), TitleTags module parses.

Type-only import of Three.Object3D so Scene.tsx has no new tsc errors.

Verify:
  pnpm vitest run src/features/VRMAvatar/  -> 47 passed
  tsc --noEmit                             -> 0 VRMAvatar errors

Part of LOBE-9434 (#3). Gives the three (currently dormant) self-iteration
builtin agents a declarative tool surface so they no longer run with no tools.

One package `@lobechat/builtin-tool-agent-signal` with internal subdirs:
- `shared/`: the 3 stable identifiers, JSON-schema specs for the full tool
  surface (resource / review / reflection), a result-kind map (read | artifact
  | mutation — the LOBE-9434 #5 discriminator), `createAgentSignalManifest`,
  and one shared `AgentSignalToolExecutionRuntime` that dispatches per api name
  and stamps every result with its kind so `extractFromFinalState` can
  partition outcomes from a persisted snapshot.
- `review/` `reflection/` `feedback-intent/`: per-mode manifests assembled from
  the shared specs + a mode-specific system prompt, exported under their three
  stable identifiers. Review = resource + proposal/idea tools; reflection &
  feedback-intent share the resource + reflection-recorder surface.

Registered all three manifests in `@lobechat/builtin-tools`. `executors` is
omitted on purpose — BuiltinToolManifest defaults to server-only execution.

The server-side execution bridge (wiring the ExecutionRuntime to the existing
createToolSet(adapters) from ToolExecutionContext) lands with the
executeSelfIteration -> execAgent migration (#7); the ExecutionRuntime is
exported (./executionRuntime) and ready for it. No production self-iteration
path is touched — executeSelfIteration still serves all current runs.

Tested: shared ExecutionRuntime (dispatch + kind stamping + error handling) and
review manifest structure. bun run type-check clean for everything touched.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…15364)

Part of LOBE-9434 (#3). Gives the three (currently dormant) self-iteration
builtin agents a declarative tool surface so they no longer run with no tools.

One package `@lobechat/builtin-tool-agent-signal` with internal subdirs:
- `shared/`: the 3 stable identifiers, JSON-schema specs for the full tool
  surface (resource / review / reflection), a result-kind map (read | artifact
  | mutation — the LOBE-9434 #5 discriminator), `createAgentSignalManifest`,
  and one shared `AgentSignalToolExecutionRuntime` that dispatches per api name
  and stamps every result with its kind so `extractFromFinalState` can
  partition outcomes from a persisted snapshot.
- `review/` `reflection/` `feedback-intent/`: per-mode manifests assembled from
  the shared specs + a mode-specific system prompt, exported under their three
  stable identifiers. Review = resource + proposal/idea tools; reflection &
  feedback-intent share the resource + reflection-recorder surface.

Registered all three manifests in `@lobechat/builtin-tools`. `executors` is
omitted on purpose — BuiltinToolManifest defaults to server-only execution.

The server-side execution bridge (wiring the ExecutionRuntime to the existing
createToolSet(adapters) from ToolExecutionContext) lands with the
executeSelfIteration -> execAgent migration (#7); the ExecutionRuntime is
exported (./executionRuntime) and ready for it. No production self-iteration
path is touched — executeSelfIteration still serves all current runs.

Tested: shared ExecutionRuntime (dispatch + kind stamping + error handling) and
review manifest structure. bun run type-check clean for everything touched.

Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Replaced all LOBE-XXX references in comments with descriptive context
instead of internal Linear issue markers. As an open-source project, these
internal references should not be exposed.

Changes:
- LOBE-9834 (empty completion): replaced with inline descriptions of the
  "empty completion" failure mode
- LOBE-6587 (task scheduler): replaced with "task scheduler infra" ref
- LOBE-6634 (getTaskDetail model/provider): updated TODO description
- LOBE-9434 #5/#7 (execAgent migration): removed issue markers

12 files changed

Co-authored-by: Arvin Xu <arvinxx@users.noreply.github.com>