Refactor Council

Refactor Council (Copilot)

Summon seven sourced refactoring-and-architecture reviewer agents, then produce a safety-first, sequenced refactoring plan that a separate adversary agent has stress-tested. The reviewers are native Copilot custom agents bundled with this plugin; the current session agent acts as the orchestrator. The reviewers argue from the writers' actual published positions and disagree with each other on purpose - the disagreement is the value.

The roster (7 + adversary)

Bundled as custom agents under agents/ (namespaced, e.g. refactor-council:fowler-refactoring-reviewer). Registry and symptom map: [references/refactoring-catalog.md](references/refactoring-catalog.md) for the smell/refactoring vocabulary and safety discipline.

| Reviewer agent | Lens | |---|---| | fowler-refactoring-reviewer | Refactoring catalog + smells; small behavior-preserving steps; two hats; Rule of Three; Strangler Fig | | martin-clean-architecture-reviewer | SOLID; Clean Architecture & the Dependency Rule; small functions; naming; frameworks/DB as details | | feathers-legacy-reviewer | Legacy = code without tests; seams; characterization tests; Sprout/Wrap; get-it-under-test-first | | beck-tidy-first-reviewer | Structural vs behavioral changes; tidyings; coupling/optionality economics; baby steps | | metz-abstraction-reviewer | Duplication over the wrong abstraction; flocking rules; squint test; depend on stable things | | ousterhout-deep-module-reviewer | Deep modules; complexity is the enemy; anti-over-decomposition; comments are load-bearing | | tornhill-hotspot-advisor | Behavioral code analysis; hotspots (complexity x churn); change coupling; where to refactor |

After synthesis, refactor-adversary red-teams the findings and plan.

Parsing the request

1. Read flags: --no-scan, --no-adversary, --since (default 12.month), --personas (subset of the seven; default all seven).
2. The remaining text is the target: a path, a directory, an @file, or diff (use git diff). If none is resolvable, ask which file/dir/diff to review - do not invent one.

Workflow

Phase 1 - Setup

Resolve the target. Read [references/refactoring-catalog.md](references/refactoring-catalog.md) so you and the reviewers name smells and refactorings precisely and enforce the safety discipline in the plan.

Phase 2 - Scan (skip if --no-scan)

Run the bundled scanners via shell and gather a compact evidence bundle:

• Smells: python3 scripts/smell_scan.py --human
• Hotspots (git repos only): python3 scripts/hotspots.py --since --human
• Safety net: detect tests near the target (*_test.*, *.test.*, *.spec.*, test/, tests/, __tests__/). Record present / absent / partial.

The bundle (top located smells, top hotspots + coupling pairs, test-net status) is load-bearing - pass it to every reviewer. If not a git repo or history is shallow, say so and note that prioritization is weakened.

Phase 3 - Reviewer delegation (sequential by default)

Delegate each selected reviewer through the agent tool, giving each: the target context, the scan evidence bundle, the Persona output contract below, and the instruction that the first line must be ## review.

Default to sequential delegation - one reviewer at a time. Copilot parallel fan-out via /fleet over-parallelizes and burns AI credits, and offers no reliability guarantee for an 8-agent council; sequential delegation is the reliable path here. Only fan out in parallel if the user explicitly asks, and then run small waves ( review` and contain the contract sections. Re-delegate an invalid reviewer once with the same bundle; if still invalid, continue and note the missing lens. Never surface raw reviewer envelopes or readiness text.

Phase 4 - Synthesize (critique + plan)

Write one integrated review using the Synthesis output shape below. Surface disagreement (especially Uncle Bob vs Ousterhout on functions/comments; Uncle Bob's DRY reflex vs Metz; decomposition vs Ousterhout); do not flatten it. Include the safety-net status, a sequenced smallest-first refactoring plan (characterization-tests-first when untested; structural vs behavioral separated, two hats), and an explicit "Do NOT refactor" list (cold code, duplication cheaper than abstraction, extraction that adds complexity).

Phase 5 - Adversarial pass (skip only if --no-adversary)

Delegate refactor-adversary with the full synthesized review (findings + plan), the target context, and the scan evidence. It returns SHIP / SHIP WITH CHANGES / HOLD plus challenged recommendations, what survived scrutiny, unsafe-as-sequenced reordering, and what the council missed. Apply the same validate-and-retry rule once.

Phase 6 - Final output

Fold the adversary verdict in before presenting: REJECT/REORDER/WEAKEN revise the plan in place; mark "survived scrutiny" items high-confidence; add a "Risks the council missed" section; if HOLD, lead with it and do not present the contested step without the fix. Keep reviewer framing internal (see Privacy).

Persona output contract

##  review

### What I see

### What concerns me

### What I'd refactor and how
 target, smallest-first>

### Safety net I'd require first

### Where I'd be wrong

Synthesis output shape

# Refactoring review: 

## Convergence (high-confidence signals)
## Disagreement (real tradeoffs you must decide)
## Per-persona top-3
## Safety net status
## Refactoring plan (sequenced, smallest-first)
## Do NOT refactor (and why)
## Risks the council missed (from the adversary)
## Adversary verdict

(Same content as the Claude skill: convergence first, named disagreement, per-persona top-3 in voice, safety-net status with characterization-tests-first when untested, a sequenced behavior-preserving plan with structural vs behavioral marked, an explicit do-not-refactor list, and the adversary's caught risks and verdict.)

Privacy / scope

The reviewer dossiers under agents/references/ are private review aids derived from each thinker's public writing. Do not republish wholesale; keep reviewer framing internal - if a review leaves the team, restate the argument in your own voice. This is a refactoring council: it preserves behavior by definition, so it does not by itself verify correctness, security, or performance. The adversary partially covers those gaps; for a full audit use a staff-level or language-specific review.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: smykla-skalski
• Source: smykla-skalski/sai
• License: MIT

Install and usage instructions live in the source repository linked above.