learningslisted

<objective> Learnings are what a project knows about working on itself that is not written in its code: which helper already exists, which pattern the reviewer always rejects, which test is flaky, which assumption turned out wrong. They are only valuable while they stay small. The index is loaded at the start of every run, so an unpruned collection dilutes the signal and eventually crowds out the ticket. Writing is half the job; **pruning is the other half.** </objective> Two modes: **write** (end of a run) and **prune** (`/gauntlet:prune-learnings`). --- # Mode 1 — Write Called by `gauntlet:runner` when a run ends, success or failure, and **before** the work directory is deleted — the worklog is the source material. ## What to extract Read `.gauntlet/work/<id>/worklog.md` and pull out only what will still be true and useful on a *different* ticket: - **Codebase facts the agent had to discover the hard way** — a helper that already existed, a convention not written down, a module that owns something surprising. - **Findings the reviewer raised** — especially any raised more than once in a run, which means the author's default was wrong. - **Corrected assumptions** — something the plan assumed and the code disproved. - **Toolchain reality** — a test that is flaky, a suite that needs a flag, a typecheck that is slow enough to change how you work. - **What actually went wrong**, on a failure. This is the highest-value entry there is; the next attempt reads it fir