mcp-cli: Skill for Claude Code

Post-Implementation Triage

Measure the actual implementation diff and decide how much review it needs. This replaces upfront estimation — the implementation itself is the best signal.

Philosophy

Predicting effort before implementation is unreliable (validated: best model achieved only 29.7% tier accuracy on 91 historical PRs). Instead:

1. Always implement with opus (the cost is similar across models; quality isn't)
2. Measure what was actually produced
3. Use concrete diff metrics to decide review depth

Usage

After implementation completes, run triage against the PR (works from any directory):

bun .claude/skills/estimate/triage.ts --pr 1152 --json   # PR-based (preferred)
bun .claude/skills/estimate/triage.ts                      # local diff vs main
bun .claude/skills/estimate/triage.ts --base develop       # local diff vs develop

Triage Rules

High scrutiny if ANY of (validated: 92.5% F1, 0% false negatives on 91 PRs):

• src churn ≥ 120 lines (additions + deletions, excluding tests)
• src additions ≥ 100 lines
• 2+ risk areas touched (IPC, auth, workers, server-pool, config, db, transport)
• 4+ source files across 2+ packages

Everything else is low scrutiny.

Pipelines

Low scrutiny: implement → QA → done

High scrutiny: implement → adversarial-review → QA → done

• If adversarial review finds issues → repair → re-review (loop until clean)

Tools

• triage.ts — post-implementation triage (reads git diff, outputs scrutiny level)
• score.ts — AST-level complexity analysis (used by triage and backfill)
• backfill.ts — mines historical PRs into SQLite for validation
• embed.ts — text embeddings for PR similarity (research; not used in triage)
• validate.ts — method comparison (proves triage > prediction)
• validate-triage.ts — validates triage rules against historical data
• db.ts — shared SQLite database helpers (used by backfill, validate, triage)

Data

Historical PR data lives in ~/.mcp-cli/estimates.db. Run bun .claude/skills/estimate/backfill.ts to populate/update.