launchd 最小 PATH 不含 /opt/homebrew/bin,keepalive 每小时触发但 4 次全部
exit 127(node: command not found)。对齐 run-web.sh 的 PATH 导出后,
launchctl kickstart 实测:访问成功、无墙、实例正常回收。

- 导航「站点登录」后新增入口 /latest:每源一张可折叠卡片,
  摘要行=源名+运行时间+✓n/✗m+已取尽标记,展开为文章标题(链接原文)+成功/失败徽标
- 数据取自各源最新运行 JSON 的业务结果(.result.articles),按运行时间倒序

- 抽屉改为分行布局:当前定时(chips)/采集频率/第N次逐行+实际时间列,
  不再一行平铺;加收起按钮
- 回显:schedule_board 将实际定时按时段归类(与规划器同口径)对应到各次意图;
  未配置时段意图的源按实际时间归类生成回显(所见即所排)

- 操作列由平铺的频率+6时段编辑器改为 ✎排程/▶采集/⏸停用(⏵启用)/✕ 一行小按钮,
  意图编辑器移到点击后行下方展开的抽屉,表格不再被撑宽
- ▶采集=立即采集一次(/run),⏸/⏵=调度与技能总开关(/toggle)
- 时间轴上的 06:00 交付线竖线与图例说明移除(规划仍按交付线排,仅不再展示)

- 队列目录 .queue → .queue.port-<port>:同站(同浏览器同账号)互斥,
  异站并行(如 nikkei 9315 与 reuters 9408 同时跑,日常即两条流水线)
- 站点端口在排队锁之前解析,锁粒度与 runlock 一致
- reuters-world 09:27 实跑验证重试链:attempt1 交付17篇+结构化交付失败
  → attempt2 续会话确认 all_seen → 合并17篇并推送成功

昨夜 nikkei-markets(GLM 无工具会话)与 reuters×3(DataDome 墙)失败时,
模型都通过 structured_output 如实上报 status=FAILED/0交付——但 attempt_needs_retry
只认运行错误(is_error/结构化缺失),业务失败被当'质量失败'直接放弃,attempt2/3 一次没跑。

- attempt_needs_retry 新增 business 判据:status=FAILED 或 0 交付;
  取尽豁免优先(all_seen/全非文章页虽自报 FAILED 但业务上是成功,不重跑)
- business 失败不走 --resume 续会话(会话确实没干成活,续接只会复述失败),
  attempt2 同 provider 整场重跑,attempt3 仍切备用 provider
- 实测矩阵:四失败信封→重试;成功交付→不重试;取尽→不重试

- bin/keepalive.sh <site>:解析站点端口/profile,采集进行中自动跳过,
  状态写 logs/<site>.keepalive.json
- LaunchAgent com.newsharvester.keepalive-reuters(StartInterval 3600,装于本机);
  面板调度列表(platform_layer.list_schedules / nh_schedule_list)排除 keepalive*,
  不显示为数据源
- 背景:DataDome 按风险评分拦 reuters(出口为数据中心 IP),保温访问维持
  cookie 活性降低凌晨采集撞墙概率

freq dropdown now spawns 第1次..第6次 window selectors (凌晨批次/上午/
下午/晚间), each run placed independently — e.g. reuters-world 2×/day as
overnight+afternoon → 02:00 + 12:00. Sources without explicit windows keep
auto semantics (first run in the deadline batch + extras evenly across
day_window); legacy string windows normalized. JS toggles run selectors
with the freq value; intent persists as a runs list in schedule-plan.json.

schedule-intent/schedule-gap concatenated str / str (default_conf_dir
returns str) → 500 on save; schedule_board already wrapped its copy.

- freq>=2 with a preferred window now means: first run in the overnight
  deadline batch + remaining runs spread in the window (reuters-world
  2×/day+morning → 04:00 overnight + 07:00 morning)
- global min gap is a dropdown on the schedule board (10..120 min),
  saved via /schedule-gap and immediately re-planned
- applied: gap 30min, reuters-world 04:00/07:00

调度中心 rows now maintain 每日次数 + 时段偏好(自动晨间批次/上午/下午/
晚间) via dropdowns; POST /schedule-intent persists to schedule-plan.json
and immediately re-plans all times. Planner gains windows support: a
source with a preferred window gets all its runs evenly placed inside
06:30-11:30 / 12:00-17:30 / 18:00-21:30, outside the morning batch;
global gap/deadline math stays automatic. Applied: reuters-technology →
morning (06:30), freeing an overnight slot (everything shifts 40min
later, more slack before the 06:00 deadline).

- harvest.sh: global queue lock (runlocks/.queue, mkdir-atomic, stale-PID
  steal, 180min wait cap) — a source whose time arrives while another is
  harvesting now waits (phase 排队中) and auto-starts when the previous
  run exits; runlock is created before queueing so the panel shows queued
  sources; manual panel trigger no longer refused (it queues)
- bin/plan_schedules.py + schedule-plan.json: frequency-driven planner —
  deadline/safety/min-gap/day-window/avoid all configurable; morning slots
  laid backward from deadline-safety with exact gaps (wrapping to previous
  evening when the roster exceeds the night), freq>1 sources spread evenly
  across the day window, global gap resolution; --dry-run supported
- 调度中心: ⚙ 按计划重排全部排程 button (POST /schedule-plan-apply)
- applied: 12 sources — 3×4/day (wsj/ft/digitimes-feed latest morning
  slots) + 9×1/day, 40-min spacing, last start 05:30 vs 06:00 deadline

A fresh attempt2 re-fetched every article (minutes + dollars) just to
re-deliver a result whose data was already archived and pushed. attempt2
now extracts session_id from the failed envelope and runs
claude --resume with a short re-deliver prompt (no re-collection, no
browser); heartbeat is not reset in resume mode. attempt3 (fallback
provider) stays a fresh full run — cross-provider resume not trusted.

When attempt1 delivers data but fails structured reporting, the final
attempt fast-exits all_seen with 0 articles — the record then showed
nothing delivered, leaving the operator to guess. After the attempt loop,
merge earlier attempts' delivered articles (URL-deduped, archive_path
required) into the final envelope, recompute counts, and record each
attempt's contribution in warnings. Single verdict, full visibility.
Unit-tested: 0+3+2(1 dup) → 4 articles, per-attempt notes.

digitimes updates registry/index incrementally while still fetching, so a
mid-fetch registry write was misread as 推送/收尾 and overrode the agent's
own correct 采集中 heartbeat. Disk inference has now been wrong in both
directions (stale when the agent forgets to flip,超前 on incremental
writers) — trust the agent's heartbeat whenever present; disk evidence
and neutral 运行中 remain fallbacks only.

On auto-retry, attempt2 resets the heartbeat to 采集中 and refetches, but
attempt1's completion files (registry etc.) are newer than the run start —
they overrode the fresh heartbeat and showed 推送/收尾 while the browser
was actively fetching. Disk evidence must beat the heartbeat's own mtime
to override; attempt resets naturally re-anchor the baseline.

edit_source checked news-harvester@<name>.timer existence (Linux layout);
on macOS schedules are LaunchAgents plists so the edit page always 404ed.
Use pl.schedule_state/schedule_times (platform layer) for the guard and
current time instead.

- attempt_needs_retry checked .result type, but the envelope's result is a
  plain text message even on success — the structured business result lives
  in .structured_output. Every healthy run was therefore misjudged and
  silently re-run (attempt2 all_seen fast-exits made it cheap and invisible;
  reuters-technology went to attempt3 and was then killed by a schedule
  reload). Criterion now: rc!=0 / empty / invalid JSON / structured_output
  not an object / is_error.
- modify-source.sh refuses to reload a schedule while the source (site)
  has a live harvest runlock — launchctl bootout terminates the running
  job (exactly what killed reuters-technology attempt3); --force overrides.
  Panel schedule-time route already had this guard; CLI is now covered.
- reuters-technology record rebuilt from its valid attempt1 envelope
  (SUCCESS, 10 articles, matching push syncstate).

The agent occasionally forgets to flip the .phase heartbeat on transition;
when the objective disk evidence (registry/.summaries/.manifest mtime) is
newer than the heartbeat file, trust the disk. Seen live: heartbeat stuck
at 采集中 while manifest had landed 3 minutes earlier.

Two validation refinements so 'nothing new but junk candidates' stops
alarming as FAILED (reuters-markets case: 18 candidates, 16 deduped, the
only 2 new ones rejected as non-article pages):
- majority-failure rule downgrades to a warning when source_exhausted and
  every failure is a non-article rejection (og:type/body), same standing
  as all_seen
- preloaded per-URL INVALID_OUTPUT failures no longer set task-failed
  (task-level no-URL failures and RUNTIME/PERMISSION still fatal); the
  per-article signal already lives in counts and warnings

attempt_needs_retry triggers on runtime error / empty output / non-object
result (GLM StructuredOutput text-return case) / is_error (1301 etc);
content-quality failures do NOT trigger (a provider switch won't fix them).
Flow: attempt1 primary → attempt2 same provider (10s backoff) → attempt3
FALLBACK_PROVIDER via resolve_continue_provider with its budget override
(ccswitch-gpt=30). Idempotent end-to-end (7-day registry dedup + slug-keyed
push), so re-runs only fill gaps; per-attempt raw outputs kept as *.attemptN.
Verified decision matrix: success/runtime-error/invalid-output/api-error/
business-fail.

Disabled/fully-off sources (skill off + schedule not active) no longer
appear; transitional state (skill off but schedule still active) stays
visible until reconciled.

systemd is-active semantics leaked to macOS (merged rows hardcoded
active=unknown; scheduled rows showed systemd 'inactive'). Map every row
to schedule_state: 调度中 / 已禁用 / 未排程.

Replace guessed phases with ground truth: harvest.sh seeds
logs/<source>.phase=采集中 and the prompt requires the agent to overwrite
it the moment it enters each stage (采集中/摘要生成/推送/收尾); trap deletes
it on exit. Panel reads the phase file first, falls back to factual disk
signals (registry/.summaries/.manifest mtimes), and shows neutral 运行中
when no signal exists — no more inference, no more (推定) labels.

Redundant runs (all articles deduped by the 7-day registry) burned ~19
turns / $0.6 verifying nothing new. Prompt now instructs: when every
candidate is already in the registry, return immediately with empty
articles + source_exhausted=true + exhaustion_reason=all_seen — which the
result contract already accepts as SUCCESS (exhaustion is not failure),
also dodging the GLM StructuredOutput misreport for this case.

A redundant run (all articles deduped) never rewrites .manifest.json, so
the disk-signal inference stayed at 采集中 during the whole summary/finalize
phase while the browser sat idle. After 10 minutes with no fetch signals,
report 摘要生成(推定) instead.

active_harvests() only read launchd/systemd service state, so runs started
via the panel button or a terminal (not launchd) were invisible on the
homepage. Union with runlock-based detection: each port-<port>.<pid>.lock
created by harvest.sh itself, PID liveness-checked, source name recovered
from the PID's command line — launcher-independent.

Lock files are port-<port>.<pid>.lock (two dots); ${f##*.} yielded the
literal 'lock' instead of the PID, so the liveness check always failed
and ANY exiting harvest deleted sibling LIVE locks — with all locks gone
the last exiter then quit the site-shared browser out from under a
still-running sibling harvest (nikkei-world 09:24 exit killed the
browser mid nikkei-markets run; earlier 05:00 run likely degraded the
same way until the 60min watchdog TERM'd it at exit=143).

Parse correctly: basename minus .lock suffix, then last dot-segment.
Self-test: live lock survives, dead-PID lock cleaned.

上次 trigger time duplicated 最近运行's log mtime — remove it (column now
just 下次执行) along with its log-mtime backfill in index().

launchd exposes no last-trigger either (systemd-only) — fill it from the
run log mtime (MM-DD HH:MM); merged unscheduled/disabled rows now carry
their parse_last_log too, so only genuinely never-run sources show
从未触发 (verified: 10 ran-today sources show real times, 4 never-run
reuters sources remain 从未触发).

- homepage: drop 成本/Session columns (detail-page info); running badge now
  shows inferred phase (采集中/摘要生成/推送收尾) via _harvest_phase():
  runlock mtime = start, vault-day .manifest/.summaries/.registry mtimes
  mark phase transitions — zero intrusion into the harvest process
- detail page: timeline-style progress bar (采集中→摘要生成→推送/收尾→完成)
  with pulsing current-stage dot, green done segments, elapsed minutes,
  10s auto-refresh while running

栏目列表 now surfaces running harvests: top banner (⟳ 正在采集: <sources>,
browsers stay open until finish), per-row 运行中 badge above 最近运行,
and 30s auto-refresh while any harvest is active — previously an
in-flight run was invisible until its .last.log landed.

launchd exposes no next-trigger (systemd-only), so the 栏目列表 下次 column
was always '-' on macOS — compute it from schedule times (same _next_fire
as the schedule board) and highlight it amber.

Blue dashed deadline marker at 06:00 on every timeline (交付线 label) +
intro copy; multi-time sources show 每日N次 badge. Editorial schedule
applied: 02:30 nikkei-world / 03:00 reuters-business / 03:30 bloomberg /
04:00 ft / 04:30 wsj / 05:00+17:00 nikkei-markets (12h interval), all
finishing ~35min before the 06:00 deadline.

app.py honors WEB_HOST/WEB_PORT env (default 127.0.0.1:8080); run-web.sh
template exports them from web.conf — set WEB_HOST=0.0.0.0 to reach the
panel from other systems on the same network (HTTP + Basic Auth, intranet
use intended).

- new /schedules 调度中心: per-source row with a 24h timeline (green dots
  at fire times, next fire highlighted amber, red now-line), 定时/下次触发
  (countdown)/上次运行 columns, sorted by next fire, 60s auto-refresh
- create/update: POST /schedule-time (HH:MM list, validated) →
  nh_schedule_write + enable (+skill enable for new rows)
- delete: POST /schedule-delete/<name> → new platform.sh nh_schedule_remove
  (bootout + rm plist incl. disabled + entry script + per-source conf;
  skill source and history kept; digest/web protected)
- guards: active harvests block time changes/deletes
- nav link 调度中心; schedule_times() string parsed to list

Clearer purpose naming: the table's function is per-source login
verification (test button + last result), not state management.

站点凭据表新增 人工登录 按钮:拉起站点共享实例窗口→用户在里面登录→
后台 login-capture.js 每15秒自动检测(最长15分钟)→成功后记录登录态、
缺配方站点自动生成检测配方并回收浏览器实例;页面自动刷新显示
向导状态(运行中/登录已记录/未检测到)。复用 _login_test 轮询通道,
键 manual:<site>。

- notify-mail.sh: bash 3.2 (stock macOS) treats expanding an empty array
  under set -u as unbound variable — drop the MSMTP_ARGS array for direct
  branch invocation (works identically on Linux bash 5)
- daily-digest.sh: empty result .json files (run killed / watchdog timeout,
  claude never wrote output) pass 'jq empty' but emit nothing on extraction,
  corrupting the summary array with consecutive commas; treat non-empty-size
  files as parse errors so killed runs show up honestly as interrupted