shipnote logoshipnote

Pipeline

How it works

A nightly export runs on GitHub. Drafting and review run on your machine. Each stage either succeeds with evidence or stops and says why.

  1. 01

    Fetch

    Working

    The nightly GitHub Action finds every public repo where your account made commits in the last 30 days, exports each commit with its full diff, scans the export for secrets, and uploads it as an artifact that expires after 10 days.

  2. 02

    Select

    Working

    Commits already used are skipped, as are denylisted repos and trivial commits such as lockfile-only changes. The rest are ranked by change size, new files, and message quality.

  3. 03

    Decide the mode

    Planned

    If at least one story clears the threshold, shipnote writes a story post. Otherwise it writes one theme post and no tweets.

  4. 04

    Draft

    Planned

    The story goes to the agent you configure, with voice samples and recent posts from agent-brain. The agent returns structured JSON, so code can check every claim.

  5. 05

    Validate

    Planned

    Every cited SHA must exist in the dataset, and every named file, function, or number must appear in that commit. Tweets must fit 280 characters. Claims that fail are dropped and logged.

  6. 06

    Review and approve

    Planned

    You see each draft beside its evidence and approve, edit, or reject it. Nothing is marked used until you run the command that does so, and nothing posts automatically.

When something fails

If the model is down, the dataset is partial, or every draft fails validation, shipnote writes a review file that explains what happened and produces no drafts. It does not retry forever or post a weaker substitute.