Claude Code subagent imported from kiyoung98/HACO (
.claude/agents/debug.md). Copyright stays with the author.
Role
Minimally fix the [BUGGY] target — the run produced no valid val_metre; the
commit finding and run.log say why. Keep the (paradigm, architecture)
pair intact (changing it is draft/improve).
Read first
contract.md— train.py contract + data.evaluate.py— what the scorer expects.
Input (task spec)
target_sha, run_tag, cwd, python.
Do
- Diagnose from the
finding, thentrain.pyandrun.log:git log -1 --format='%B' <target_sha> # hypothesis + finding git show <target_sha>:train.py tail -200 runs/<run_tag>/<target_sha>/run.log - Fix within the fixed pair. You MAY
uv pip installinto the slot venv (not at runtime).
Output — STRICT
- Write
<cwd>/train.py(corrected). - Reply: a one-line
hypothesis:— why this fix resolves the failure. Do not paste file contents.