Review Against SOLID Principles — Ensemble (Arm B)

Fixed paths (relative to this arm directory, the claude -p working directory):

• Ruleset: ../ruleset/solid-rules.json
• Targets: every *.py under ../corpus/cases/
• Worker reports: ./findings/workers/ (absolute path when calling the planner/workers)
• Final output: ./findings/findings.md
• Ensemble scripts: plan_ensemble.py and reduce.py from the ensemble-rule-review skill. From this arm directory they are at ../../../plugins/plugin-creator/skills/ensemble-rule-review/scripts/; the runner may pass an absolute path instead — use whichever it provides.

1. Plan (deterministic). Run the planner over the shared ruleset, with ./findings/workers/ (as an ABSOLUTE path) as the report dir:

   uv run <scripts>/plan_ensemble.py ../ruleset/solid-rules.json \
     --report-dir "$(pwd)/findings/workers" --window 2 --json

   Use the worker assignment and recommended keep-threshold from the planner's output.

2. Map (parallel). Spawn the principles-reviewer agent once per worker in the plan. Give each its groups + per-group rules as YOUR RULE SLICE, the identical targets (all files under ../corpus/cases/), and its planner outfile. Run all workers in one parallel batch. Workers write location as corpus/cases/<file>:<line>.

3. Reduce (deterministic). Merge by corroboration on (group, location) and write the ranked result to ./findings/findings.md:

   uv run <scripts>/reduce.py "$(pwd)/findings/workers" --glob 'worker-*.md' \
     --keep-threshold 2 > ./findings/findings.md

The reduced, ranked findings in ./findings/findings.md are the arm's output — the scorer parses them against ../corpus/gold.json.

Each principles-reviewer worker emits the fixed candidate schema.