Understand an unfamiliar repo
Measured evidence for the understand-repo case. Fired in 2 of 4 runs -- the prompt is underdetermined, which is a finding about the prompt, not a plugin failure.
| mode | model | run | outcome | skills fired | hook denials | mode denials | diff | cost | time |
|---|---|---|---|---|---|---|---|---|---|
plan | haiku | #1 | PASS | befund:befund-stage-map | 0 | 0 | 0 B | $0.26 | 61s |
plan | haiku | #2 | FAIL | -- | 0 | 0 | 0 B | $0.24 | 65s |
plan | sonnet | #1 | FAIL | -- | 0 | 0 | 0 B | $0.35 | 97s |
plan | sonnet | #2 | PASS | befund:befund-stage-map | 0 | 0 | 0 B | $0.40 | 95s |
Why the failing runs failed
- expected skill(s) never fired: befund:befund-stage-map
What the run said (plan, haiku, run #1)
Full picture confirmed — that's every import in the repo. Here's the real module map, in a repo that turns out to be a small, deliberately booby-trapped teaching example. ## Modules (5 files, one package) | Module | Role (per docstring) | Imports out | |---|---|---| |
pipeline/__init__.py| package marker | none | |pipeline/ingest.py| Stage 1: read raw records | none | |pipeline/normalize.py| Stage 2: normalize records |pipeline.score| |pipeline/score.py| Stage 3: score records
Recorded 2026-09-14 from wf-2026-09-14-acceptedits, wf-2026-09-14-plan, wf-2026-09-14-plan-sonnet with 2.1.236 (Claude Code). Every number above comes from a cell record; regenerate with test/workflows/curate.py.