lesson-harvestlisted

# lesson-harvest — mechanize finding the lessons a report already wrote ## Why this exists A pilot or executor's closure report already carries a "## Lessons for the memory" section by convention (see the `pilot` agent definition's "Final report contract"). Nobody automatically reads it. The manual alternative — a human re-reading every report, every card, to find what's worth keeping — doesn't scale and gets skipped under time pressure, which is exactly how a documented gotcha gets rediscovered the hard way a second time. This tool closes only that one gap: given a report file, tell the session **whether there is anything to harvest, and what it is** — nothing more. It is deliberately NOT a memory writer: the project's memory store may have exactly one writer (the session doing the integration), and deciding a fiche's name, type, and wording, or whether it duplicates an existing fiche, is a judgment call this tool cannot make. ## The two-sided contract this enforces - A report whose section explicitly says **"None."** produces **no output artifact and no noise** — `hasLessons: false`. Do nothing. - A report whose section carries content produces the **candidate items, verbatim** — `hasLessons: true`, `items: [...]`. Read each one and decide, per the project's own memory discipline (one lesson → one operative fiche home, a short index line, never restating detail in the index), whether and how to persist it. - A report with **no such section at all** is reported