Cleanup Cycles

Detect circular import dependencies and break them where it's mechanically safe. Cycles between leaf utilities can be fixed by extraction; cycles between core modules need architectural decisions and are reported, not auto-fixed.

Preflight

1. Language detect: package.json (TS/JS), pyproject.toml (Py), go.mod (Go), Cargo.toml (Rust). Note: Go and Rust prevent cycles at compile time, so this skill is mostly TS/Py work.
2. Git state: refuse auto-apply on dirty tree.
3. Report dir: ensure .claude/cleanup-reports/ exists.
4. Existing scripts: check package.json for a cycle:check (or similar) script — if the repo already wires up madge with custom config, prefer that over defaults.

Detect

TypeScript / JavaScript

# Madge is the standard. Skott is faster for large repos.
bunx madge --circular --extensions ts,tsx,js,jsx --json src/ apps/ packages/ > /tmp/madge.json 2>/dev/null \
  || npx madge --circular --extensions ts,tsx --json . > /tmp/madge.json

Each entry is an array describing one cycle: ["a.ts", "b.ts", "a.ts"].

Python

pipx run pycycle --here --verbose > /tmp/pycycle.txt 2>&1 || true
# Fallback: import-linter with auto-config
pipx run import-linter > /tmp/import-linter.txt 2>&1 || true

Go

# Go enforces acyclic at compile; this just confirms build is clean.
go build ./... 2>&1
# Optional: visualize with goda for ergonomics
go install github.com/loov/goda@latest 2>/dev/null
goda graph ./... > /tmp/goda.dot 2>/dev/null || true

If go build fails with import cycle, that's the report.

Rust

cargo build 2>&1   # rustc rejects cycles
# For visibility into module graph: cargo-modules
cargo install cargo-modules 2>/dev/null
cargo modules generate tree 2>/dev/null || true

Assess

Write .claude/cleanup-reports/cleanup-cycles-{YYYY-MM-DD}.md:

# Circular Dependencies Assessment — YYYY-MM-DD

## Summary
- Total cycles: N
- HIGH confidence (auto-fixable by leaf extraction): X
- MEDIUM confidence (refactor needed): Y
- LOW (architectural redesign): Z

## Cycles

### Cycle 1 — HIGH
- Path: `a/util.ts → b/helper.ts → a/util.ts`
- Shared piece: `formatCurrency` defined in `b/helper.ts`, called by `a/util.ts`. `b/helper.ts` imports a single constant `LOCALE` from `a/util.ts`.
- Plan: Extract `LOCALE` to new `a/constants.ts`. `b/helper.ts` imports from there. Cycle broken.

### Cycle 2 — MEDIUM
- Path: `domains/user/index.ts → domains/account/index.ts → domains/user/index.ts`
- Both modules export and consume each other's primary types. No leaf to extract.
- Recommendation: introduce a `domains/shared/types.ts` for cross-domain types, OR invert one direction with dependency injection.

## Critical Assessment
[2-3 paragraphs: what does the cycle pattern reveal about the architecture? Are cycles concentrated in one area? Is there a missing layer?]

Apply

Auto-fix HIGH-confidence leaf-extraction cycles only.

Confidence rubric

HIGH (auto-apply):

• Cycle has exactly 2 modules.
• One direction of the cycle is a single small thing: a constant, a type, a pure utility function ≤20 lines, with no further dependencies inside the cycle.
• Extracting that thing to a new module will provably break the cycle.

MEDIUM (report only):

• Cycle has 3+ modules.
• Both directions consume non-trivial APIs of the other.
• Extracting would require moving classes/functions with their own dependency tails.

LOW (note for human):

• Cycle is structural (e.g., bidirectional ORM relations, parent/child component refs) — may be intentional.
• Cycle disappears under conditional imports — leave alone, document.

Execution (HIGH only)

1. Identify the leaf piece (constant/type/util).
2. Create new file at the appropriate location: src//.ts. Prefer placing inside the consumer that has fewer outside imports.
3. Move the leaf there.
4. Update both old modules' imports.
5. Re-run madge/pycycle to confirm the cycle is gone.
6. Commit: chore(cleanup): cleanup-cycles — N cycles broken via leaf extraction.

Verify

# Re-run cycle detection — should report 0 for HIGH-applied cycles
bunx madge --circular . 2>&1
pycycle --here 2>&1
go build ./... 2>&1
cargo build 2>&1

# Then standard typecheck/test/lint (see cleanup-unused for full list)
bun run check 2>&1 || npx tsc --noEmit && npx eslint .
pytest 2>&1

If verify fails or new cycles appear: git revert HEAD, mark fixes as MEDIUM, escalate.

Output

• "Broke N circular dependencies. M cycles deferred for architectural review."
• Path to report.
• Verify status.

NEVER

• Auto-apply on cycles with 3+ modules — these always need human judgment.
• Use barrel files (index.ts re-exports) as the cycle-breaking solution — they often hide cycles instead of fixing them.
• Touch framework-imposed cycles (e.g., React component file importing its own types from a sibling) — those are conventions, not bugs.
• Move types into a types.ts god-file — prefer co-location with the smallest scope that breaks the cycle.
• Suppress cycle warnings via tooling config — fix or report, never silence.

Source & license

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

• Author: raintree-technology
• Source: raintree-technology/agent-starter
• License: MIT
• Homepage: https://claude.raintree.technology

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