Gemini CLI headless stream-json only emits assistant text/tool events, while the
actual thought trace lives in the local ~/.gemini session JSON. That left Gemini
responses with no structured thinking payload for the frontend, so there was
nothing for ThinkingContent to wrap even though the CLI had recorded thoughts.
The service now looks up the matching local session after a successful turn and
emits a normal thinking system_info block when the stored response content matches
the streamed assistant text.

Constraint: Gemini headless stream-json does not currently expose model thoughts as first-class events
Rejected: Guess thinking from assistant prose prefixes | too brittle and would wrap normal response text as fake reasoning
Confidence: medium
Scope-risk: narrow
Directive: If Gemini CLI eventually exposes thought chunks in stream-json directly, prefer those over local snapshot recovery and keep the content-match guard to avoid attaching stale thoughts
Tested: cd packages/api && pnpm run build && node --test test/gemini-agent-service.test.js
Not-tested: Live end-to-end Gemini invocation through the running Cat Café UI

…e data

Kimi's 7-day consumption was being reported as zero because the daily
usage aggregator only counted normalized input/output token fields,
while Kimi often contributes last-turn or context snapshot tokens
instead. Session-chain drill-downs also hardcoded assistant rows to a
purple palette, which made Kimi sessions visually inherit the wrong
family style.

This change teaches usage aggregation to fall back to Kimi-compatible
token fields and carries session cat identity into the session viewer so
Kimi rows render with the gray Kimi theme instead of a ragdoll-like
purple.

Constraint: Kimi invocation usage may omit normalized input/output token fields
Constraint: Search-opened session viewers still may not know cat identity and must fall back neutrally
Rejected: Recompute daily usage from official quota pools | different metric from invocation consumption
Rejected: Keep assistant session rows universally purple | clashes with provider-specific family themes
Confidence: high
Scope-risk: narrow
Reversibility: clean
Directive: Preserve provider-aware token fallbacks in daily usage aggregation; do not assume all cats emit normalized input/output fields
Tested: pnpm --filter @cat-cafe/api run build; CAT_CAFE_DISABLE_SHARED_STATE_PREFLIGHT=1 node --test packages/api/test/usage-aggregator.test.js packages/api/test/usage-route-cache.test.js; pnpm exec vitest run src/components/__tests__/session-chain-panel.test.ts src/components/__tests__/SessionChainPanel-viewSession.test.ts src/components/audit/__tests__/SessionEventsViewer.test.ts
Not-tested: Live runtime UI after deployment to the shared 3003/3004 instance

Kimi stores session state by canonical workdir path, but this branch still matched resume state with bare resolve() equality. In common git-worktree and symlinked workspace setups that dropped last_session_id lookup, so Kimi resumed as a fresh session even when kimi.json already had the correct state.

This change canonicalizes both the live working directory and persisted kimi.json entries through resolve + realpathSync + normalize, and locks the behavior with a regression test that exercises a symlink alias against the real stored path.

Constraint: Kimi persists work_dirs using canonical filesystem paths while Clowder often invokes cats from symlinked or worktree aliases
Rejected: resolve()-only comparison | misses symlink aliases and recreates the reported blocker
Confidence: high
Scope-risk: narrow
Directive: Keep Kimi session-state path matching aligned with any other Kimi workdir canonicalization logic
Tested: pnpm --dir packages/api run build; cd packages/api && node --test test/kimi-agent-service.test.js; pnpm --filter @cat-cafe/api run lint; pnpm biome check packages/api/src/domains/cats/services/agents/providers/KimiAgentService.ts packages/api/test/kimi-agent-service.test.js --diagnostic-level=error
Not-tested: End-to-end kimi-cli resume on Windows junction paths

Locks down the guarantee that explicit `bound account "..." not found`
errors bubble up intact instead of being wrapped in a generic
`failed to resolve bound account` message. This verifies that
invokeSingleCat never silently falls through to service.invoke when
the explicitly bound account is missing.

[宪宪/Opus-46🐾] (test authored by 砚砚/GPT-5.4)

… [梵花猫/kimi🐾]

Project-level Kimi skills were discovered but their SKILL.md metadata
(description + triggers) was not extracted because the metadata scan
only checked user-level ~/.kimi/skills, not project-level .kimi/skills.

- Add projectRoot/.kimi/skills to skillDirCandidates for metadata extraction
- Add regression test verifying project-level Kimi skill metadata
  (description + triggers) appears in capability board response

Fixes review comment zts212653#7 from PR zts212653#361

…[烁烁/Gemini-2.5🐾]

Implements @gemini25's design suggestions for PR zts212653#361:

1. Brand colors: Each cat now uses its own brand color when selected
   - Claude (anthropic): #9B7EBD (purple)
   - Codex (openai): #5B8C5A (green)
   - Gemini (google): #5B9BD5 (blue)
   - Kimi: #4B5563 (gray)
   - Dare: #D4A76A (brown)

2. Cat avatars: Added avatar images to ChoiceButton and PillChoiceButton
   - Uses existing /avatars/{client}.png assets
   - 40x40px for ChoiceButton, 20x20px for PillChoiceButton

3. Subtitles: Added subtitle descriptions for each cat in CLIENT_OPTIONS
   - Kimi: '擅长处理中文长文本、总结归纳和资料整理'
   - Claude: '擅长深度思考与代码审查'
   - Codex: 'OpenAI 官方 Codex CLI'
   - Gemini: '多模态长上下文专家'

Files changed:
- hub-add-member-wizard.parts.tsx: brand colors, avatars, subtitle support
- HubAddMemberWizard.tsx: pass client/subtitle props
- hub-cat-editor.model.ts: add subtitle field to CLIENT_OPTIONS

…ules

Extract standalone helpers into two focused modules to bring
KimiAgentService.ts under the 350-line hard limit:

- kimi-config.ts (275 lines): config reading, path resolution,
  session/context utilities, MCP config writing
- kimi-event-parser.ts (114 lines): stream-json event parsing,
  text/thinking extraction, usage stats, prompt building
- KimiAgentService.ts (335 lines): class only, imports from above

All 14 kimi-agent-service tests pass. No public API change.

Addresses P2 from Round 2 review (zts212653).

Auto-format files flagged by `pnpm biome check --diagnostic-level=error`:
- KimiAgentService.ts: import ordering
- kimi-config.ts: formatting
- GeminiAgentService.ts: line length formatting
- gemini-agent-service.test.js: import ordering
- RightStatusPanel.tsx: formatting
- session-chain-panel.test.ts: formatting
- .kimi/mcp.json: formatting

Remove unrelated files flagged in Round 3 review:
- standup-aggregator.ts, standup.ts, and their tests (independent feature, will be separate PR)
- omx.png avatar (OMX residual after scope convergence)
- standup route registration from index.ts and routes/index.ts