paulrberg/agent-skills/code-polish

Code Polish

Resolve scope once. Make only high-confidence simplifications. Fix defects supported by evidence according to their risk. Verify the final state.

Modes

• --simplify: simplify only.
• --review: review and fix only.
• Neither or both: simplify, then review the simplified result.
• --with-profile <name> / --skip-profile <name>: add or suppress review profiles. Skip wins.

Fixed Scope

1. Require a Git repository.
2. Use explicit paths, patterns, ranges, natural-language targets, or a supplied resolved-scope block when present.

Otherwise, use only files modified in this session. If session history is unavailable, use all uncommitted tracked and untracked files.

1. Exclude lockfiles, generated outputs, vendored code, minified bundles, and large data snapshots from manual review

unless explicitly requested. Validate relevant excluded outputs through their generator, schema, or invariants.

1. Resolve and retain one authoritative scope set and optional exclusions for execution. Do not broaden or recompute

scope later. Stop if it is empty.

Simplify

Preserve public contracts, inputs, outputs, side effects, error behavior, performance-sensitive characteristics, telemetry, and operational guards. Apply only changes with a concrete comprehension or defect-risk benefit:

• flatten avoidable control-flow nesting.
• clarify misleading names or dense transforms.
• remove real duplication when the abstraction reduces total complexity.
• tighten local types and contracts without broad churn.
• remove only dead code caused by this session's edits.

Do not split by line count, perform architecture cleanup, convert sync/async APIs, add speculative configurability, or replace readable duplication with a one-use abstraction. A no-op is a valid result.

Review and Fix

Judge the diff against the user's request. Prioritize CRITICAL → HIGH → MEDIUM → LOW:

• CRITICAL: exploitable security, data loss, or critical outage path.
• HIGH: behavior, error-path, boundary, or performance defect affecting core behavior.
• MEDIUM: resource leak, complexity hotspot, test gap, over-scoped change, speculative complexity, or weak success

criterion likely to cause defects.

• LOW: localized clarity or style issue with a real maintenance cost.

Every finding must cite a verified location, triggering input/state, failure mode, blast radius, and evidence in the changed code. Merge duplicates and apply the smallest defensible fix. Resolve ambiguous intent from the request and repository first. State routine assumptions. Ask only when the answer changes behavior or scope. While that item waits, continue independent in-scope fixes.

Select every applicable profile and read it once. Read all selected profiles together in one parallel turn:

Profiles live at references/profiles/<name>.md. Missing selected profiles are a stop condition.

Verification and Report

Run the narrowest formatter/lint, targeted tests, typecheck, and invariant checks that prove the final touched behavior. Broaden only for shared contracts. Name skipped checks and why.

Summarize scope with the file count and smallest useful repository-relative roots, globs, ranges, or user-supplied targets. Do not enumerate every file merely to prove scope. Name individual paths only for a small explicit scope or to clarify exceptions and findings.

Findings include severity, location, impact, evidence, fix, and confidence. A residual risk states the assumption, consequence if wrong, and how to check it.

Under Issues and caveats, group verified fixes with evidence as Resolved, and remaining problems, limitations, or unverified assumptions as Open, with their impact and next step. Omit empty groups. Report each item once. Put neutral context and agreed decisions under changes or scope.

Reserve blocker for something preventing required work and risk for a specific potential adverse outcome. A workaround leaves an item open when the underlying issue still affects the result. Completion requires fixed scope, traceable edits/findings, and validation evidence.

Lead a successful report with ### ✨ Code polish — ✅ complete, then summarize scope, meaningful changes, and verification. For small results, use short prose. Add tables or sections only when they clarify multiple changes or findings. Honor the user's requested format.

When review ran and found no defects, state ✅ No verified review findings. If a stop condition below prevents completion, lead with ### ✨ Code polish — ⛔ blocked and report the evidence and required decision. Keep severity tokens, profile IDs, commands, locations, reproduction inputs, and security evidence exact and undecorated.

Stop when behavior parity or required high-risk validation cannot be established, or a fix requires an unrequested public-contract change or larger redesign.