daily_report

Auto-generates a concise daily work report from three local sources and delivers it to Feishu:

1. git commits — your commits that day across all repos under ~/code/*
2. Claude Code sessions — that day's session titles, prompts, and edited files
3. Feishu calendar — that day's meetings/interviews (via CalDAV)

Sources 1 & 2 are harvested for the current user plus any "extra users" — service/bot accounts whose work should roll up into the same report. None are folded in by default; name them via the DAILY_REPORT_EXTRA_USERS env var (space-separated) or --extra-users a b. Their $HOME is 0700, so gather_context.py reads them via passwordless sudo -n -u <user> (git repos under /home/<user>/code/* and sessions under /home/<user>/.claude/projects). Commits shared between clones are de-duplicated by hash; an unreachable extra user is skipped with a note (never a crash). Disable extras with a bare --extra-users. This is how work done under a service account (e.g. report_hub) rolls into the same report.

Remote machines (sessions only). Claude sessions on other hosts — e.g. Claude Desktop on a Windows box reached over msys2 bash — fold in via DAILY_REPORT_SSH_HOSTS="host1 host2" (ssh aliases) or --ssh-hosts. Each host is one ssh -o BatchMode=yes call that streams back ~/.claude/projects/*/*.jsonl modified since the window start; no git is read there. Sessions carry a [<host>] tag; an unreachable host is skipped with a note. Desktop session names (custom-title) are used as titles, ignoring the New session placeholder.

Automated pipeline sessions are dropped. Queued automation (e.g. tq_ai library/alphas mining) runs claude in ephemeral per-job scratch checkouts and can spawn hundreds of sessions a day — under the primary user too — swamping the report. Sessions whose cwd matches DAILY_REPORT_EXCLUDE_CWD_GLOBS (or --exclude-cwd, e.g. /tmp/tqlib_ro_* /tmp/aha_ro_*) are skipped, keeping only interactive Claude-Max work; the dropped jobs' git commits still count. Match only the scratch dirs, not the pipeline's own code dirs. Empty / bare --exclude-cwd = no filtering.

A small LLM pass (Claude headless) synthesizes the material into themed bullets.

Layout

Path	What
~/my/daily_report/	this repo — all code + config
~/my/daily_report/holidays/	cached China holiday/workday JSON (one file per year), auto-refreshed
~/.claude/skills/daily-report	symlink → this repo, so /daily-report works as a skill
/tq/scratch/<user>/daily_report_log/	generated reports + run_daily.log + last_covered cron cursor (NOT in repo)

Files

• gather_context.py — git + Claude-session material for a date or date range
• gather_calendar.py — Feishu calendar for a date/range (CalDAV). See FEISHU_SETUP.md
• workday.py — is a date a China workday? (调休-aware) + computes each cron run's report span
• fetch_holidays.py — refresh the cached China holiday calendar (holiday-cn JSON) into holidays/
• send_feishu.py — deliver a report to Feishu (group webhook, or 1:1 DM)
• run_daily_report.sh — gather → synthesize (claude -p) → archive → deliver (workday-gated)
• SKILL.md — the /daily-report skill definition (interactive use)
• FEISHU_SETUP.md — calendar (CalDAV) + delivery setup
• .feishu_*.json — local secrets, gitignored (see the .example files)
• daily_report.local — local site config, gitignored (extra service accounts, automated-session EXCLUDE_CWD_GLOBS, OUT_DIR override; see the .example)

Usage

Interactive (in Claude Code): /daily-report 2026-06-01

Manual / scheduled:

~/my/daily_report/run_daily_report.sh                 # cron mode: workday-gated, span since last report
~/my/daily_report/run_daily_report.sh 2026-06-01       # one specific date (never touches cron state)
~/my/daily_report/run_daily_report.sh 2026-06-01 --no-send
~/my/daily_report/run_daily_report.sh --no-send        # cron-mode dry-run (generate only; no DM, no state)

Scheduling & workday gating

No systemd/cron natively on this box, so cron was installed and a job added:

0 7 * * *  ~/my/daily_report/run_daily_report.sh >> /tq/scratch/<user>/daily_report_log/cron.log 2>&1

Runs every morning at 07:00. The cron run is China-workday-aware (incl. 调休 makeup workdays — driven by workday.py + the cached calendar in holidays/):

• On a workday it reports every day since the last report (cursor file last_covered), so the first workday after a weekend/holiday folds the whole break into ONE report (e.g. the Thursday back from National Day covers 2026-09-30 → 2026-10-07).
• On a non-workday it doesn't synthesize — it DMs a short rest-day placeholder that points at the next workday (e.g. 📅 Holiday: 国庆节 (2026-10-01) — … on the next workday (2026-10-08)).
• last_covered advances only after a successful send, so a failed send or a missed run is retried next time — no dropped or double-sent days, and it self-heals after downtime.

The calendar self-updates: each run calls fetch_holidays.py --quiet (throttled to ~monthly), which pulls next year's calendar once the State Council publishes it; otherwise everything is offline. After a host reboot the cron daemon isn't auto-started (no systemd) — re-arm with sudo service cron start.

Cursor (last_covered). Lives at /tq/scratch/<user>/daily_report_log/last_covered and holds the last date already reported. The next cron run covers (last_covered, yesterday]. Seed/reset it with echo 2026-06-14 > .../last_covered; delete it to re-bootstrap (next run then covers just yesterday). A corrupt/blank cursor self-heals to "cover yesterday". Archived span reports are named YYYY-MM-DD_YYYY-MM-DD.md (single days stay YYYY-MM-DD.md).

Output format

English; a single markdown bullet list (no section headers, no commit count). 3–5 themed work bullets (- **Topic**: …, max 8), then a final calendar bullet when there are events: - 📅 2 interviews; Quant Team Weekly; 1-on-1 with Chang Zhang (interviews collapsed to a count; meetings prefer their own name, falling back to "Meeting with " — names romanized to pinyin — only when the title is generic). See SKILL.md for the exact rules.

Delivered to Feishu as a blue-header interactive card (Daily Report — <date>) with the markdown rendered — bold Topics, inline code, one bullet per line. To paste it into a Feishu Doc with formatting intact, copy from the Feishu desktop client (the web client drops bold on paste); for a 100%-faithful copy, import the archived .md into Feishu Drive (drag it into 云空间). See FEISHU_SETUP.md for delivery setup + paste tips.

Setup on a fresh checkout

pip install requests icalendar recurring_ical_events
python3 fetch_holidays.py                               # seed holidays/ (this + next year)
cp .feishu_caldav.json.example .feishu_caldav.json     # then fill in (or use --set-caldav)
cp .feishu_webhook.json.example .feishu_webhook.json   # then fill in (or use --set-webhook)
cp daily_report.local.example daily_report.local       # optional: extra service accounts / OUT_DIR
ln -s ~/my/daily_report ~/.claude/skills/daily-report  # register the skill

holidays/*.json is committed, so a fresh checkout already works offline; the fetch_holidays.py step just refreshes it. The calendar uses China's statutory holidays + 调休 makeup workdays via holiday-cn.