Run Eval Harness
Own the whole run-monitor-parse-diff loop and return once with a complete summary. The user should never have to ask “check the current state”.
1. Locate the harness
Search in order; stop at the first hit:
- Project docs: README, AGENTS.md, docs/ mentioning “eval”.
- Task runners:
justfile,Makefile,package.jsonscripts,pyprojectscripts with eval targets. - Convention paths:
evals/,eval/,benchmarks/, files matching*eval*.py|ts|go.
If multiple harnesses exist, pick the one the user named; otherwise list
them and pick the default documented in the repo. Note the dataset in use
(golden set path) and where reports land (for example eval_report*,
results/, run_log.jsonl).
2. Launch in the background
- Run via the documented entry point with the repo’s own defaults; do not invent flags.
- Use a background shell so the session stays free; capture stdout to a log file in the scratchpad.
- Record start time, git commit, model or provider config in effect.
3. Monitor without spamming
Poll the log at an interval matched to expected runtime (a 10 minute run gets checks every 2 to 3 minutes, not every 15 seconds). Detect and report early: crash, auth failure, rate limiting (429 or backoff messages), or a stall with no new output for 3 poll cycles. On transient provider errors, retry the run once before reporting failure.
4. Parse into the normalized summary
Extract whatever subset the report provides:
| Metric | Notes |
|---|---|
| Pass / total, pass rate | Per strictness level if the harness has them |
| Latency P50 and P95 | Report both; never substitute mean |
| Tokens in / out per case | And totals |
| Cost per run and per case | From real configured prices, never hardcoded guesses |
| Failures | Case id, expected vs actual, one-line cause each |
5. Diff against the previous run
Find the most recent prior report or run log. Report: metric deltas, newly failing cases, newly passing cases, and whether config changed between runs (model, dataset, prompt version, commit). If no prior run exists, say so and record this one as the baseline.
6. Report once
Return a single summary: headline result, the metric table, the diff, and failure explanations. State the exact command used, the commit, and the report file path so every number is reproducible. If the harness itself is broken, report the root cause and stop; do not silently fix eval logic, and never edit the golden set or scoring code to make a run pass.