haabe/mycelium/eval-runner

Eval Runner

Benchmark the agent's performance against defined scenarios. Adapted from n-trax eval system.

Commands

run <category/name>

1. Read YAML from .claude/evals/scenarios/<category>/<name>.yml
2. Parse fields (name, category, task_prompt, success_criteria, budget)
3. Execute setup steps if defined
4. Record start time
5. Execute task via reflexion workflow (read corrections first)
6. Record end time and iteration count
7. Validate ALL success criteria
8. Write result JSON to .claude/evals/results/<timestamp>-<name>.json
9. Report summary

run-all [category]

1. Glob .claude/evals/scenarios/**/*.yml
2. Skip scenarios with status: retired
3. For each: run in isolation (git stash), record result, restore
4. Update .claude/evals/pass-history.json with each result
5. Aggregate and report

run-split <optimization|holdout>

1. Glob .claude/evals/scenarios/**/*.yml
2. Read each YAML, filter by split field matching the requested set
3. Skip scenarios with status: retired
4. For each matching scenario: run in isolation, record result, restore
5. Update .claude/evals/pass-history.json with each result
6. Aggregate and report (label output clearly as "Optimization Set" or "Holdout Set")

report

1. Read all results from .claude/evals/results/
2. Generate summary table:

| Category    | Pass Rate | Avg Iterations | Avg Time | Notes |
|-------------|-----------|----------------|----------|-------|
| discovery   | ...       | ...            | ...      |       |
| delivery    | ...       | ...            | ...      |       |
| integration | ...       | ...            | ...      |       |
| **Overall** | ...       | ...            | ...      |       |

1. List failure patterns and recommendations

prune

1. Read .claude/evals/pass-history.json
2. Flag evals where last_5 is all-pass (saturated) or all-fail (broken)
3. Flag evals with no runs in 30+ days (stale)
4. For saturated evals, suggest: retire or increase difficulty
5. For broken evals, suggest: fix criteria or retire
6. Present recommendations — do NOT auto-retire
7. On user confirmation: set status: retired in scenario YAML, update pass-history.json, log in .claude/harness/decision-log.md

mine

Analyze audit logs to propose new eval scenarios from observed failure patterns.

1. Read .claude/state/change-log.jsonl (last 100 entries)
2. Read .claude/state/diamond-state-audit.jsonl (all entries)
3. Group change-log entries by session_id, identify:

a. Correction clusters: 3+ edits to same file in one session (agent struggled) b. Skill friction: edits to .claude/skills/*/SKILL.md during a session (instructions unclear) c. Missing test coverage: 5+ files changed with no test file edits

1. Count diamond bypass entries from audit log
2. Cross-reference with existing scenarios in pass-history.json to avoid duplicates
3. For each pattern, output a proposed eval scenario as YAML template
4. Tag proposals with source: trace-mining and originating session_id
5. Do NOT auto-create files — present for human review

See .claude/evals/schema.md §Trace Mining Heuristics for pattern-to-eval mappings.

Result Format

See .claude/evals/schema.md for YAML scenario and JSON result formats.

Pass History

After writing each result JSON (step 8 of run), also update .claude/evals/pass-history.json:

• Increment runs and passes (if passed) for the eval
• Append true/false to last_5 (trim to keep only last 5)
• Update last_run timestamp

Creating Scenarios

Place YAML files in .claude/evals/scenarios/<category>/. Define task_prompt, success_criteria, and budget. Set split, status, and source fields per schema.