daily-retro

Experimental. Daily retro from your agents' own logs: reads today's Claude Code and Codex sessions on this machine, groups them by project, adds git state (commits, uncommitted files, unpushed commits

SKILL.md

GitHub

Daily retro (experimental)

Part of the Personal Corp framework. Weekly retro answers "what did the week give". Daily retro answers a smaller question every evening: what did my agents do today, and what did they leave half-done. You do not remember twenty sessions in five repos; the logs do.

Experimental means: the shape may change, it is not in the plugin's active skill list, you copy it by hand (see README).

What it reads

Only local files, nothing leaves the machine:

Harness Logs
Claude Code (CLI and desktop) ~/.claude/projects/<project>/<session>.jsonl
Codex ~/.codex/sessions/YYYY/MM/DD/rollout-*.jsonl

A harness that is not installed is skipped. Another harness: add a reader to scripts/collect.py, same output shape.

Steps

  1. Collect. python3 <skill>/scripts/collect.py (today) or --date YYYY-MM-DD. It prints a markdown digest: per project the sessions (time, harness, title or first prompt, last reply), commits of the day and tails: uncommitted files, unpushed commits, a session that ended on your message without a reply, a session that ended with a question to you. Secrets that look like tokens are masked.
  2. Read, do not retell. Turn the digest into the retro file retro/daily/YYYY-MM-DD.md in your HQ (create the folder once):
    • Done — one line per result that exists outside the chat: a commit, a merged PR, a file, a sent message. No result outside the chat, no line.
    • Tails — every tail from the digest plus the promises you find in last replies ("will do after…", "waits for your word"). Each tail gets one owner action: commit, push, answer the session, issue comment, drop (with one reason).
    • Tomorrow — at most three tails that block something else.
  3. Close tails like the manager. For a tail that belongs to a task: find its issue and leave a work-record comment (what was done, what is left, link to commit). If the manager skill is installed, use it: it owns the issue contract. Commit and push only your own paths; somebody else's uncommitted edits stay as a tail with the owner's name, you do not revert them.
  4. Report. Three lines to the owner: how many sessions and projects, what got done, the three tails for tomorrow. In the automation run the report goes where the owner reads (Telegram, a note, the HQ dashboard) — whatever the HQ already uses.

Automation

The point is that nobody has to remember to run it. Pick one, templates in automation/:

  • Claude Code desktop: Scheduled tasks → new task, every day at 21:30, prompt /daily-retro, folder = your HQ.
  • macOS launchd: automation/daily-retro.plist runs claude -p "/daily-retro" in the HQ folder at 21:30. Load it once: launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/corp.daily-retro.plist.
  • Linux systemd: automation/daily-retro.service + .timer.

A run that finds zero sessions writes one line "no agent sessions today" and stops.

Boundaries

  • Reads logs, does not edit them. Does not send the raw digest anywhere: last replies may hold private context; the retro file is written in your own words.
  • Does not merge, deploy, pay or post on its own. Such tails go to "Tomorrow" as the owner's decisions.
  • The digest is a pointer, not proof. Before writing "done" check the commit, the PR or the file.