code-refinement

Code Refinement Workflow

Analyze and improve living code quality across six dimensions.

Quick Start

/refine-code
/refine-code --level 2 --focus duplication
/refine-code --level 3 --report refinement-plan.md

When To Use

• After rapid AI-assisted development sprints
• Before major releases (quality gate)
• When code "works but smells"
• Refactoring existing modules for clarity
• Reducing technical debt in living code

When NOT To Use

• Removing dead/unused code (use conserve:bloat-detector)

Analysis Dimensions

Plugin-Specific Patterns

Detection patterns for plugin and skill codebases where standard code quality heuristics miss structural issues.

Delegation Stub Bodies

A skill that declares "delegates to X" but still carries the full template body is doing double duty. The delegating skill should be a thin wrapper (under 30 lines) that routes to the target. Flag any delegating skill whose body exceeds 50 lines.

Module Explosion

Flag skills with 10+ module files where 40% or more of content overlaps. Signal: two modules covering the same API surface from different angles (e.g., both describing the same config options or the same CLI flags).

Oversized Single Modules

Flag individual module files exceeding 500 lines as candidates for splitting or trimming. Large modules defeat progressive loading by forcing full-file reads for partial information.

Dead Python References

Skills referencing Python commands (python -m module.name or python -c "from module import ...") where the referenced module does not exist in the plugin's src/ directory. These are stale references to renamed or removed code.

Progressive Loading

Load modules based on refinement focus:

• modules/duplication-analysis.md (~400 tokens): Duplication detection and consolidation
• modules/algorithm-efficiency.md (~400 tokens): Complexity analysis and optimization
• modules/clean-code-checks.md (~450 tokens): Clean code, anti-slop, error handling
• modules/architectural-fit.md (~400 tokens): Paradigm alignment and coupling

Load all for thorough refinement. For focused work, load only relevant modules.

Required TodoWrite Items

1. refine:context-established: Scope, language, framework detection
2. refine:scan-complete: Findings across all dimensions
3. refine:prioritized: Findings ranked by impact and effort
4. refine:plan-generated: Concrete refactoring plan with before/after
5. refine:evidence-captured: Evidence appendix per imbue:proof-of-work
6. refine:findings-verified: Citations confirmed by citation_verifier.py
7. refine:execution-complete: All wave-listed candidates closed-or-rationale'd (only required when invocation includes "execute findings" or stronger; see Step 6)

Workflow

Step 1: Establish Context (refine:context-established)

Detect project characteristics:

Step 2: Dimensional Scan (refine:scan-complete)

Load relevant modules and execute analysis per tier level. For dimension 7 (Additive Bias), run Skill(imbue:justify) to compute the bias score, check Iron Law compliance, and flag unnecessary additions or workarounds.

Step 3: Prioritize (refine:prioritized)

Rank findings by:

• Impact: How much quality improves (HIGH/MEDIUM/LOW)
• Effort: Lines changed, files touched (SMALL/MEDIUM/LARGE)
• Risk: Likelihood of introducing bugs (LOW/MEDIUM/HIGH)

Priority = HIGH impact + SMALL effort + LOW risk first.

Step 4: Generate Plan (refine:plan-generated)

For each finding, produce:

• File path and line range
• Anchor: verbatim source text at the cited line
• Current code snippet
• Proposed improvement
• Rationale (which principle/dimension)
• Estimated effort

Step 5: Evidence Capture (refine:evidence-captured)

Document with imbue:proof-of-work (if available):

• [E1], [E2] references for each finding
• Metrics before/after where measurable
• Principle violations cited

Fallback: If imbue is not installed, capture evidence inline in the report using the same [E1] reference format without TodoWrite integration.

Step 6: Execute Findings (refine:execution-complete)

Steps 1-5 produce a plan. Steps 6 produces closures. Both are part of the skill. Execution does not stop at planning unless the user explicitly says "plan only".

Execution mode detection

Match the user's invocation phrasing against this table to determine execution scope:

The triggers compose: --tier 3 --execute all findings --ignore-scope-guard means run every Wave 2 and Wave 3 candidate to closure regardless of branch size.

Completion gate (when execution mode is active)

The task is not complete until ALL of the following hold:

1. Wave 2 candidates (medium-effort, listed in synthesis "Wave 2 Candidates" section): every entry has either a closure commit or an explicit per-item line in the synthesis stating why it is not viable.
2. Wave 3 candidates (large-effort, listed in synthesis "Wave 3 Candidates" section): same gate. Do not pre-emptively defer LARGE-effort items with generic "needs dedicated PR" rationale when the user said "execute all". Execute the mechanical ones (split-by-class, mixin-package, module-merge) and reserve "deferred" only for items requiring architecture-level decisions (schema changes, new dependency declarations, new venv layouts).
3. Synthesis updated: docs/refinement/<date>/00-synthesis.md records every closure with its commit SHA and every deferral with one-sentence rationale.

Anti-pattern detector for the agent itself

If the model finds itself doing any of the following during execution, this is a stop-hook leak. Go back to executing findings:

What decides the gate

The gate reads records, not the agent's account of its own work: .review/findings.json and the citation verifier's exit code (see the next section) say which candidates are closed, and the synthesis lists the rest. Execution is bounded by max_waves (default 3, one wave per pass over the open candidates). When the budget runs out, or when the harness fires a stop signal, the run stops: record every remaining candidate in the synthesis as a gap with what would close it, and end the turn. Do not resume past a stop signal. A harness stop is a signal from outside the run, and a run that overrides it has made itself the judge of its own completeness, which is the arrangement measured to drift (arXiv 2607.17641).

Verify Findings Are Grounded (refine:findings-verified)

Write findings to .review/findings.json, run the citation verifier (Skill(imbue:review-core) Step 5), and drop or label UNVERIFIED any the verifier rejects.

Exit Criteria

• All six analysis dimensions have a verdict (finding or "no issue detected") for the target scope.
• Each finding includes a file path, line range, and verbatim Anchor (the exact source text at that line).
• Every reported finding carries a Location + verbatim Anchor confirmed by citation_verifier.py (exit 0), or unverified findings were dropped or labeled UNVERIFIED.

Tiered Analysis

Cross-Plugin Dependencies

Supporting Modules

• Code quality analysis - duplication detection commands and consolidation strategies

When optional plugins are not installed, the skill degrades gracefully:

• Without imbue: Evidence captured inline, no TodoWrite proof-of-work
• Without conserve: Uses built-in clean code checks (subset)
• Without archetypes: Skips paradigm-specific alignment, uses coupling/cohesion principles only