Agent Skills: Deslop

Deslop

Run this only after the change is functionally correct and before commit. PR text should describe already-deslopped code.

Goals

• Keep the smallest clear diff that still solves the task.
• Run focused reviews in parallel instead of one subjective final pass.
• Preserve behavior while improving clarity, type safety, and rule alignment.

Required review vectors (exactly 3 agents, in parallel)

All 3 agents get the same context bundle, but each gets one vector:

1. Rules and docs conformance
   • Check AGENTS.md, nested AGENTS.md, design docs, and core beliefs.
   • Flag drift from documented patterns or ownership boundaries.
2. Type safety and source of truth
   • Preserve canonical types and inference flow.
   • Flag casts, duplicated/redefined types, or unnecessary widening.
   • Prefer compile-time guarantees inside typed internal code.
   • Validate only at untrusted boundaries; do not re-validate trusted internal values.
3. Overengineering and simplification
   • Remove unnecessary code, wrappers, abstractions, factories, or indirection.
   • Prefer direct, local solutions when equivalent.

Required context bundle

Read and pass these paths:

• repo root AGENTS.md
• nested AGENTS.md for changed areas
• docs/index.md
• docs/PLANS.md
• docs/design-docs/index.md
• docs/design-docs/core-beliefs.md
• directly relevant design docs
• active exec plan for this work (if any)
• changed files plus nearby context

If working on an ExecPlan:

• inspect docs/exec-plans/active/
• if one clearly matches, tell reviewers to study it carefully before reviewing

Delegation protocol

1. Read the full context bundle yourself first.
2. Launch all 3 required review agents immediately (in parallel).
   • Do not wait for local lint/slop checks.
3. Give each agent:
   • same context bundle
   • one assigned review vector
   • instruction to return findings first, ordered by severity, with file refs
4. While agents run, start with:
   • pnpm -w lint:slop:delta
5. After all responses arrive, synthesize under these exact headings:
   • How did we do?
   • Feedback to keep
   • Feedback to ignore
   • Plan of attack
6. Use balanced synthesis over any single reviewer’s extreme take.

Auto-apply feedback (unattended flows)

Apply clear, in-scope fixes before commit, prioritizing:

• type drift, casts, duplicate type definitions
• documented boundary/design-belief violations
• dead helpers/code/debug leftovers/placeholders
• removable local indirection

Skip speculative, conflicting, or scope-expanding suggestions; note briefly in synthesis/workpad.

Execution steps

1. Gather context bundle.
2. Launch 3 required review agents in parallel.
3. Run pnpm -w lint:slop:delta and narrow local checks while they run.
4. Wait for responses and synthesize.
5. Apply worthwhile in-scope feedback.
6. Re-run the narrowest affected validation.
7. Update workpad/commit/PR text to describe final post-deslop state.

Stop rules

• No refactors unrelated to the ticket.
• No churn outside changed area for style only.
• Leave subjective/unclear “improvements” alone.
• Do not blindly apply every suggestion.