/superheal

Self-heal target: $ARGUMENTS

You MUST follow the Autoresearch-style protocol. Anything that skips a step is invalid.

Required artifacts (write all of these to .superbuilder/experiments/EXP-NNN.json)

1. problemStatement — what is failing, observed N times.
2. baselineMetric — measured number BEFORE any change. Pick from: storyPassRate, falsePassRate (must be 0), averageRetries, diagnosisQuality, securityBlockSuccess, regressionCount, tokenCost, sourceGroundingRate, contextSufficiency.
3. editableSurface — exactly one of: skills/*/SKILL.md, agents/*.md, commands/*.md, orchestrator config defaults, gate ordering, retry policy, eval rubric.
4. mutation — one change. ONE. Describe diff and rationale.
5. evalTaskSet — fixed list of tasks with deterministic pass/fail. Same set used for baseline and after.
6. scoreBefore / scoreAfter — same metric, same set.
7. safetyRegression — did any blocked operation succeed? Did any false pass appear? Must be none to keep.
8. decision — keep | revert.
9. log — full timeline.

Hard prohibitions

The mutation must NOT touch:

• hooks/hooks.json blocking rules
• approval policy
• production deploy guardrails
• secret-scan rules
• the requirement that stories have evidence to pass

Invoke 10-self-improve skill which holds the full protocol. Hand the resulting EXP-NNN.json to self-improvement-researcher agent for synthesis.

CLI verb (deterministic measurement)

The numeric measurement step runs through the orchestrator's heal verb, not by hand:

# Baseline-only run (no mutation): re-establish current numbers.
bin/superbuilder-heal --baseline-set B
bin/superbuilder-heal --baseline-set A   # uses examples/eval-fixture/

# With a mutation (one change at a time):
bin/superbuilder-heal --mutation /tmp/EXP-XYZ.patch --baseline-set A

The verb writes .superbuilder/experiments/EXP-NNN.json matching the shape in docs/EVALS.md and prints the same JSON to stdout. The agent reads the JSON and applies the keep/revert rule:

Keep iff securityBlockSuccess == 1.0 AND falsePassRate == 0 AND the target metric improved (or held flat in baseline-only runs).

Output

Print a one-screen summary:

• problem
• mutation
• baseline → after (delta + safety regression status)
• decision
• file path of the experiment log