codebase-improver

Codebase Improver

Take a codebase from "I think there are problems" to "the problems are mapped, prioritized, and fixed" — end to end, human in control at every gate.

The improvement engine is built on deep modules (John Ousterhout, A Philosophy of Software Design): a deep module has a small interface hiding a large implementation. Deep modules are more testable, more AI-navigable, and let you test at the boundary instead of inside. The goal of the Deepen stage is to turn shallow, tightly-coupled clusters into deep modules.

This skill is self-contained (Python/TS); worker roles run inline by default and may use explicitly authorized delegation.

Map ──► Audit ──► Deepen (recursive) ──► (next round)
 │        │          │
scan    audit      explore → candidates → design N interfaces
        roles      → RFC → optionally implement → recurse

Host portability

Without delegation authority, run every worker role serially inline; this requires zero AgentTool calls. Apply the operating defaults before selecting distinct workers. Do not claim inline roles are independent or parallel; report unavailable isolation when it is essential to the requested result.

Core principles

Self-contained. Mapping is inline; checks and contracts live in bundled references/. No other skill required.
Blueprint is a bonus, not a requirement. Use a BLUEPRINT.md as stronger anchors if one exists; otherwise rely on the lightweight scan + bundled checks.
Deep modules are the target. Improvement means deepening: shrinking interfaces, hiding complexity, testing at the boundary. Friction encountered while reading the code IS the signal.
Compose worker roles; delegate only with authority. The main agent sequences the workflow, aggregates results, runs checkpoints, and owns stateful decisions. Requested workers stay within the user's delegation scope; otherwise the same roles run inline.
Human-in-the-loop at every gate. Never decide scope, which candidate to pursue, which interface to adopt, whether to recurse, or whether to commit/push. Pause and ask. (This skill is HITL by design — unlike an AFK autopilot, it does not fabricate human-gate decisions.)
Evidence over assertions. Every finding cites a real path:line.
Recursion is bounded and gated. Never auto-recurse. Each deeper level needs an explicit user yes, and there is a default depth cap (see Recursion control).
Respect conscious decisions. A choice documented in CLAUDE.md, a decision spec, or a blueprint is not a finding.
Simplicity over brevity. Clarity is the goal — never line count. When a transformation makes code simpler to read (explicit over clever, named intermediate steps, flat over nested), prefer it even if it adds lines. A longer, obvious implementation beats a short, dense one. This applies to the implementation inside a module; the module's interface still stays small. Behavior must be preserved and proven by tests, and code is never padded for its own sake. See references/simplify-playbook.md.
Implementation is test-driven and quality-looped. Code changes go through a TDD + review + QA loop (tests first → implement → review → fix → QA-simulate → fix → repeat until clean, capped). See references/quality-loop.md.

Stage 1 — Map

Understand the repo's shape through a read-only map role, using the selected execution mode.

Detect the stack. Targets Python and TypeScript (manifests: pyproject.toml/requirements.txt/setup.py; package.json/tsconfig.json). Neither → say so and stop. Both → treat each as its own project.

Lightweight structural scan:

find . -maxdepth 2 -type f \( -name 'pyproject.toml' -o -name 'requirements.txt' \
  -o -name 'setup.py' -o -name 'package.json' -o -name 'tsconfig.json' \) \
  -not -path '*/node_modules/*' -not -path '*/.venv/*'
tree -L 3 -I 'node_modules|.venv|venv|dist|build|.next|__pycache__' 2>/dev/null || \
  find . -maxdepth 3 -type d -not -path '*/node_modules/*' -not -path '*/.venv/*' | sort

Note project roots, apparent layers, and repeated folder patterns.

Opportunistic blueprint. If a BLUEPRINT.md exists (.claude/…, docs/…), read it for stronger anchors. If not, proceed — do not block.
(Optional) Pack context for an LLM via repomix / tree -L 3 if the user wants a single-file snapshot.

Stage 2 — Audit (whole repo)

The main agent composes read-only audit roles using the selected execution mode. Catalogs:

references/universal-checks.md — language-agnostic anti-patterns.
references/audit-catalog.md — Python/TS health signals + detection commands.
references/audit-worker.md — the contract every audit role follows.

Partition the repo into scopes from the Stage 1 scan: one per-subtree worker per significant module (local checks), plus one cross-cutting worker (duplication, manifest dependency direction, stale deps, repo-wide secrets). ~1 worker per module; 2–4 for a small repo.
Run every worker. Inline roles run serially; explicitly authorized delegations may run concurrently. Each gets its scope, catalogs, blueprint path, and scratch output path. Workers are read-only.
Aggregate (main agent): merge fragments, dedupe repeated path:line, assign IDs, sort by severity.
Execution mode: use the same partition and output shape in either mode; record actual isolation.

Present the findings inventory:

ID	Finding	Evidence (`path:line`)	Category	Anchor	Severity

⏸ Checkpoint A (always):

"Audit found N findings: X blockers, Y should-fix, Z nits. Review before I look for deepening opportunities? Any you consider intentional / out of scope?"

The inventory feeds the Deepen stage: it points exploration at the friction-heavy areas. Ephemeral by default; save to docs/audit/findings-<YYYY-MM-DD>.md only if asked.

Stage 3 — Deepen (recursive)

This stage replaces a flat "plan then execute". It runs the deep-module flow, then recurses in the three forms below. Each candidate handled is one deepening pass; passes nest.

3.1 Explore for deepening opportunities

Run an exploration role in the selected execution mode to navigate the codebase the way an AI would — organically, not by rigid heuristics — seeded by the audit findings. Note where you hit friction:

Understanding one concept requires bouncing between many small files.
A module is so shallow its interface is nearly as complex as its implementation.
Pure functions were extracted only for testability, but the real bugs hide in how they're called.
Tightly-coupled modules create integration risk in the seams between them.
Parts that are untested or hard to test.

The friction IS the signal.

3.2 Present candidates

A numbered list of deepening opportunities. For each:

Cluster — which modules/concepts are involved.
Why coupled — shared types, call patterns, co-ownership of a concept.
Dependency category — one of the four in references/deep-module-reference.md.
Test impact — which existing tests would be replaced by boundary tests.

Do NOT propose interfaces yet.

⏸ Checkpoint B: "Which of these would you like to explore?"

3.3 Frame the problem space

For the chosen candidate, write a user-facing explanation: the constraints any new interface must satisfy, the dependencies it must rely on, and a rough illustrative code sketch to ground the constraints (a sketch, not a proposal). Show it, then immediately proceed to 3.4 — the user thinks while the design passes run.

3.4 Design multiple interfaces

Run 3+ design roles (see references/interface-designer.md) serially inline by default; explicitly authorized distinct workers may run concurrently. Each produces a radically different interface under a different constraint:

Agent 1: minimize the interface — 1–3 entry points max.
Agent 2: maximize flexibility — many use cases, extension points.
Agent 3: optimize for the most common caller — make the default case trivial.
Agent 4 (if cross-boundary deps): ports & adapters.

Each returns: interface signature, a usage example, what complexity it hides, dependency strategy, trade-offs.

Present the designs sequentially, compare them in prose, then give your own opinionated recommendation — strongest design and why, or a hybrid if elements combine well. The user wants a strong read, not a menu.

⏸ Checkpoint C: user picks an interface or accepts the recommendation.

3.5 Capture as an RFC

Create a refactor RFC as a markdown file in docs/ using the template in references/deep-module-reference.md. Don't ask for pre-review — create it and share the path.

3.6 (Optional) Implement — TDD + QA quality loop

The deep-module flow ends at the RFC by design. If the user wants it built now, run the quality loop in references/quality-loop.md. It is test-driven, composes worker roles, and iterates review→fix→QA→fix until clean (capped). The main agent orchestrates and keeps the git/HITL gates.

⏸ Checkpoint D (before any code change): "Build this RFC now on a branch, or leave it as an RFC for later? (If building: I'll go TDD — tests first — then implement, review, and simulate QA, looping until clean.)"

On a yes, in order:

Branch git switch -c improve/<rfc-slug> (confirm; never push/PR without explicit go-ahead).
Tests first (red). A test-writer role writes boundary tests from the RFC's acceptance criteria and preserved behavior; run them to define the target.
Implement (green). An implementer role makes the tests pass in small steps, applying references/simplify-playbook.md.
Review. Run correctness and maintainability review roles following references/review-rubric.md; merge and dedupe findings.
QA plan + simulate. Run QA-plan and QA-simulation roles; retain pass/fail/skip evidence.
Fix & loop. Blocking findings or QA failures → a fix role addresses them → step 4. Cap at MAX_QUALITY_ITERATIONS (default 3), then report "needs human".
Report what changed, tests, review + QA status, path:line.

The loop runs to completion then reports — it does not prompt mid-loop, but it also never fabricates a decision: anything genuinely needing a human (a blocked dependency, an ambiguous behavior change) stops that branch and is flagged, per principle 5.

Configuration (defaults, override in plain language)

MAX_QUALITY_ITERATIONS = 3 — review→fix→QA cycles before flagging "needs human".
REVIEW_BLOCKING_SEVERITY = high — only findings at this severity or above block; lower ones are recorded, not blocking.
GIT_STRATEGY = branch — create branch + commit; PR/push only on explicit go-ahead; never auto-merge.

3.7 Recurse — three forms, each gated

After a pass, offer the user the next move. Never recurse automatically.

Decompose (downward). If the chosen candidate is too big to deepen as one unit, split it into sub-parts and run 3.1–3.6 on each sub-part. The sub-parts inherit the parent's framing.
Re-explore (deeper). After implementing a refactor, re-run 3.1 scoped to the changed area — deepening one module often exposes the next layer of friction underneath. A simplify pass (apply references/simplify-playbook.md to the changed code) is a valid lightweight form of this.
Re-audit (outward). When the chosen candidates are done, optionally re-run Stage 2 to surface the next round across the repo, then return here.

⏸ Checkpoint E (after each pass):

"Pass complete. Recurse? (a) decompose this candidate further, (b) re-explore the area I just changed, (c) re-audit the whole repo for the next round, or (d) stop here."

Recursion control

Recursion is powerful and easy to run away with, so it is bounded:

Default depth cap = 3 nested levels (a pass inside a pass inside a pass). To go deeper, ask explicitly: "We're 3 levels deep. Keep going?"
Always gated. Every deeper level passes through Checkpoint E first. No silent continuation.
Track the stack. Keep a short visible trail of where you are, e.g. payments → refund flow → idempotency key, so the user never loses the thread.
Diminishing returns stop. If a level surfaces only nits, say so and recommend stopping rather than recursing for its own sake.

Where files go

Artifact	Location	When
Findings inventory	chat by default; `docs/audit/findings-<YYYY-MM-DD>.md` if asked	Stage 2
Refactor RFC	`docs/<rfc-slug>.md`	Stage 3.5
QA test plan	`docs/qa/test-plan-<branch>.md`	Stage 3.6, when implementing
Code changes	edited in place on branch `improve/<rfc-slug>`	Stage 3.6
Repomix snapshot (optional)	`repomix-output.xml` at repo root	Stage 1, if requested
`BLUEPRINT.md`	only read if it exists — this skill never writes one	Stage 1

Audit/exploration/design roles write intermediate fragments to a scratch workspace, not the repo. Only RFCs and (if asked) the saved inventory land in docs/. Never write outside the repo; never overwrite a same-day file without asking.

Stop-and-ask checkpoints summary

#	Checkpoint	When
A	Review findings inventory	End of Stage 2, always
B	Pick a candidate to explore	Stage 3.2, always
C	Pick an interface design	Stage 3.4, always
D	Implement now vs RFC-only	Stage 3.6, before any code change
E	Recurse vs stop	After each pass, always
—	Branch / commit / push	Whenever a git side-effect is about to happen

Add pauses anytime something is genuinely ambiguous. Never invent a decision the human should make.

What this skill does NOT do

It does not depend on any other skill — mapping, checks, and design contracts are bundled.
It does not propose interfaces before the user picks a candidate, nor implement before the user okays it.
It does not recurse on its own, exceed the depth cap silently, or do big-bang rewrites.
It does not push branches, open PRs, or commit without explicit per-action approval.

About this skill