harness-transformer

name	harness-transformer
description	Convert a deterministic HarnessDiagnosis into a bounded MutationPlan JSON. Use when mutation pressure is high enough that the harness should propose candidate changes, but do not apply those changes yet.

harness-transformer Skill

Purpose

Turn diagnosis findings into a small, reviewable MutationPlan using only the allowed mutation classes defined by the plugin invariants.

Trigger

Use this skill after harness-analyzer when:

• mutation_pressure >= 0.30
• any critical or high finding is marked mutation_candidate: true
• the user asks for a bounded harness mutation plan

Steps

1. Load the latest HarnessDiagnosis JSON.
2. Run:

   python ${CLAUDE_PLUGIN_ROOT}/scripts/harness_core/transformer.py \
     --diagnosis ${CLAUDE_PLUGIN_ROOT}/state/harness-report.json \
     --json-output ${CLAUDE_PLUGIN_ROOT}/state/mutation-plan.json
   

3. Validate the emitted plan:

   python ${CLAUDE_PLUGIN_ROOT}/scripts/harness_core/validate_invariants.py \
     --harness-dir $CLAUDE_PROJECT_DIR \
     --mutation-plan ${CLAUDE_PLUGIN_ROOT}/state/mutation-plan.json
   

4. Summarize:
   • patch count
   • permission widening flag
   • top target paths
   • rollback path
   • generation mode (rule_based vs llm_assisted)

Constraints

• Do not apply the patch plan directly.
• Do not emit mutation types outside the schema allowlist.
• If permission_widening: true, surface the rationale explicitly.
• LLM assistance is advisory only and must be post-validated by deterministic schema and invariant checks.
• If LLM refinement is unavailable or invalid, keep the rule-based candidate plan.

Next Skill

Route to harness-candidate-pr-loop only after invariant validation passes.