- classify enrich sources: encyclopedia, platform pages (hongguo, kuaikaw, video
platforms), douban (search snippet only) and other sites' search snippets
(snippet mode, pages never opened); one extractor per source, fields carry
source kind and confidence, merged by noisy-OR across hosts
- new cast field (applied as MediaCredit with placeholder ids)
- robots.txt check (cached per host 1 day) and login-wall/captcha give-up
- per-enrichment source attempts + GET /identity/enrich-stats, 30-day hit-rate
table under settings -> metadata -> identity
- setting identity.enrichSnippetAllSites (default on); default domains add
baike.com and kuaikaw.cn
- replay fixtures from the real run on the 4 user-library short dramas
(sanitized Brave responses, hongguo page, robots.txt)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
camofox reaps a tab after TAB_INACTIVITY_MS (5 min) without API tool calls;
desktop-mode input goes straight to Selkies and screenshots don't count, so
tabs disappeared under the user during takeover.
- BrowserKeepalive (in-process): sessions held by an open desktop WebSocket
(ref-counted), a human takeover lease, or a running/queued agent run that
used browser.* get a keepalive every BROWSER_KEEPALIVE_MS (default 120s,
0 = off). Holds are re-checked each round; release/lease expiry/agent end/
session close stop it and camofox reaps as usual.
- Keepalive call: camofox 1.17 has no keepalive/touch endpoint and /stats does
not bump toolCalls; use GET /tabs/{id}/downloads without includeData/consume
(read-only metadata). Goes through BrowserService.run so restarts or reaped
tabs self-heal via BR-3 reconnect; no lastAutomationAt, no control needed.
- Runbook browser.md §12, deploy env table, compose env; tests against the
testkit mock camofox.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- adapters: overviewLanguages() limits overview/tagline chain to Chinese
variants (zh-CN→zh-SG→zh-TW→zh-HK); titles keep full chain; season
supplement fetch skips languages useless for overview
- setting metadata.overviewNonChineseFallback (default off) + UI switch
- media detail shows "暂无简介" for empty entity/episode overview
- scripts/overview-recompute.ts: recompute stored non-Chinese overviews
from kv_cache only (no TMDB requests), dry-run by default, --apply
- tests: chain unit tests, Heart Signal (103147) fixture, server
recompute test, MP differential for episode overview
- runbook + CHANGELOG
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Remove the "换掉的标识" mapping list under humanized reports; the raw
report text is kept only as "原文" in the result line's 技术信息 (no mapping).
- Session stats line above the overview bar, always shown: tool calls, MCP
calls (only when any), tokens used and left to the budget (K/M + percent),
cost; live while running, fixed after the run ends.
- Live turn status: "思考中 / 调用工具中 / 回答中 · 已调用 N 次工具、M 次 MCP";
after the turn the count merges into each progress line ("调用 N 次工具").
- agent.spawn calls answered with an existing same-goal child (reused=true)
are not counted in "派出了 N 个子会话".
Screenshots (desktop + mobile): apps/web/test/screenshots/ag4/ (gitignored).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- Drop "site rate limit counted per process; cross-process token bucket in
progress on c2a": migration 0041 shipped the PG token bucket (61f585a).
- "原站视图是 1.5 秒截图轮询,没有在真实 camofox 上核验": desktop mode (WT-1)
exists and was verified on real camofox (v0.2.4 notes); keep only the
"HTTP failure does not fall back to the browser" part, still true.
- HHClub was checked against the real site in VER-3 / HH-2; keep 观众、馒头、CHD.
- Other entries re-checked against 11-requirements-coverage and code: still hold.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A spawn whose goal (NFKC, case, whitespace and punctuation folded) matches a
child of the same parent that is still queued/running, or succeeded/waiting
within the last 10 minutes, returns that child instead of creating a new one:
output gets reused=true, a message ("已有同目标子会话…") and the child's
current result summary. Failed/cancelled/over-budget children are not reused.
Spawn decisions are serialized per parent in-process so parallel duplicate
calls in one batch create a single child. Replays keyed by idempotency still
hit the original child as before. The subagent constraint in the system prompt
asks for Chinese sub-goals and no duplicate spawns.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Subagent / vision subflow prompts (audience=agent) get a dedicated section:
report goes to the parent session, and thinking, step notes and reports stay
in Chinese even when the goal, tool results or web pages are in English.
Prompt regression test covers both the shared hard rule and the new section.
Child reports reaching the user are already humanized: the child session page
uses the same FIX-7 report-humanize path, and the parent's final report resolves
child run IDs to their goal titles; agent.spawn/await tool rows show no raw summary.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
root