# rehearsal-runs/ — curated per-run evidence (2026-09-27)

One directory per sub-agent run. Generated by `tools/rehearse/run-rehearsal.sh`;
full raw sessions (pi jsonl transcripts, complete workspaces) live in
`/tmp/rehearsal/runs/<run>/` and are intentionally not committed.

Per run:

| File | Content |
|------|---------|
| `agents-md.txt` | The exact AGENTS.md the agent worked against (shared design instructions + variant tool rules) |
| `prompt.txt` | The user prompt given to `pi -p` |
| `measure.txt` | Tokens (sum + per-usage breakdown), turn/tool-call counts, session span, wall clock, tool-name histogram (from `measure_run.sh`) |
| `verify.txt` | Deterministic gate result: diff size vs pristine, `mvn test` with the solution's test + approved file, structural checks (see `verify_outcome.sh`) |
| `diff.patch` | Final diff of the run workspace vs the pristine exercise (excludes test artifacts) |
| `pi-stdout.txt` | The agent's final report (summary + final diff as it saw it) |
| `time.txt` | `/usr/bin/time -v` output (wall clock, max RSS, exit status) |
| `mvn-verify-tail.log` | Tail of the mvn output for the gate run — only present for failures |

Run → variant/model map:

| Dir | Variant | Model |
|-----|---------|-------|
| `A`   | CLI mandated        | openrouter z-ai/glm-5.3-flash |
| `B`   | plain pi tools      | openrouter z-ai/glm-5.3-flash |
| `C`   | CLI optional + intro| openrouter z-ai/glm-5.3-flash |
| `A+Q` | CLI mandated        | entrim Qwen/Qwen3.6-35B-A3B |
| `B+Q` | plain pi tools      | entrim Qwen/Qwen3.6-35B-A3B |
| `C+Q` | CLI optional + intro| entrim Qwen/Qwen3.6-35B-A3B |

Narrative, drift documentation, and conclusions: see
[`../rehearsal-findings.md`](../rehearsal-findings.md).