obsidian-capture

One capture = one scripts/capture.py call. The script handles everything deterministic — resolves today's daily / this week's weekly note (creates it from the vault template if missing), routes to the right section, fills empty placeholder slots before appending, formats Tasks-plugin dates — so never edit the notes by hand. Needs $OBSIDIAN_VAULT and python3 (stdlib only).

Commands

scripts/capture.py todo "回 offer 邮件"                    # daily ✅ 其他, due today
scripts/capture.py todo "准备 design review" --must        # daily 🎯 今天必须完成, due today
scripts/capture.py todo "写 freshrss theme" --weekly       # weekly ## Todos, due = today+3d, capped at Sunday
scripts/capture.py todo "读论文" --due friday -p high      # any todo: explicit due + ⏫ priority
scripts/capture.py log "和 recruiter 通话"                 # daily 🪵 Log, timestamped now (--time HH:MM to backfill)
scripts/capture.py learned "TIL 内容"                      # daily 📝 今日记 → 新知:
scripts/capture.py diary "发生的事"                        # daily 📝 今日记 → 新事:
scripts/capture.py fun "有趣的事"                          # weekly ## Something Fun This Week
scripts/capture.py reflect "杠杆" "答案全文"               # weekly Reflections, under the Q: matching the substring
scripts/capture.py done "package 邮件"                     # tick first matching open #agent-todo (today's notes first, then all), appends ✅ date
scripts/capture.py annotate "加 CI" "blocked: 缺权限"      # append a sub-bullet under first matching open #agent-todo

# agent-owned todos (#agent-todo): captured for a later, possibly weaker, agent to execute cold
scripts/capture.py todo "给 skills repo 加 CI" --agent \
  --ctx "context: repo ~/Local/wilbeibi-skills，GitHub Actions 跑 skill lint" \
  --ctx "done-when: push 后 Actions 全绿"
scripts/capture.py agenda                                  # list open #agent-todo due ≤ today + context, file:line (--days 90 default)

--due accepts today, tomorrow, +Nd, mon..sunday, this-week, next-week, YYYY-MM-DD, none — pass the user's phrase through instead of computing dates yourself. Run --help for the rest.

Judgment (the only part the model decides)

• Plain todos are the human's; --agent ones are yours. agenda/done/annotate only ever match #agent-todo lines, so they can't touch a human todo — but never mark one done without checking it against its own done-when:.
• Keep the capture text in the language the user used; don't translate or rephrase.
• "必须 / must / today's deadline / blocking" → --must; otherwise a plain todo.
• Daily vs weekly: default daily. Use --weekly only when the user says weekly / 本周, or the task is week-scoped with no fixed day. A todo due far in the future still goes in today's daily — the Tasks plugin finds it by its 📅 date, not by file location.
• "让 agent 做 / 给 agent 记" → --agent. The executor may be a weak model with no conversation history, so the todo must be cold-start executable: --ctx "context: ..." with absolute paths/repos/URLs and --ctx "done-when: ..." acceptance criteria. Pull specifics from the current conversation; if you can't make it self-contained, ask the user rather than capture a vague one. (The script refuses --agent without --ctx.)
• The script prints the file, section, and exact line it wrote — relay that as confirmation, and stop. If it errors (unmatched section/question/todo), fix the arguments; don't fall back to hand-editing.