Refactoring audit

Produce local-only refactoring plans for this monorepo. Output is planning material for humans — not code changes unless the user explicitly asks to implement something.

Language: All audit outputs MUST be written in English.

Stability rule — no pinned source paths in skill docs

This skill, its templates, and repo-rules-checklist.md describe patterns, thresholds, and grep commands — not specific source file paths. File paths move, split, and get deleted; hardcoded paths in instructions go stale.

• Skill artifacts → patterns + metrics only.
• Audit output docs (refactoring*.md) → list paths discovered during the current run (from collect-metrics.sh + reads). Regenerate tables each audit; do not copy example paths from templates.
• Phase milestones → verify with grep/counts (see Step 4), not memorized filenames.

Outputs (never commit)

Git rule: These files MUST stay out of git. After writing, ensure each path is listed in .git/info/exclude. Run git check-ignore -v <file> and confirm git ls-files does not list them. Never git add refactoring docs. Do not add them to .gitignore unless the user explicitly wants team-wide ignore rules in the repo.

Excluded areas

Leave celebration code out of every audit: the celebrations Nx project (@epam/ai-dial-celebrations), the chat-api celebration domain and anything it owns (celebration-named services, types and config), and the UI_EVENT* env vars. The feature is still under development, so do not collect metrics, smells, dead-code candidates or phase items for it, and do not propose moving, splitting or renaming it, even when it trips a threshold. Mention the exclusion once in refactoring.md.

When to use

• User asks for refactoring plan, tech-debt review, or codebase analysis
• User asks to refresh refactoring docs after major merges/archives
• User wants backend vs frontend debt split
• Periodic audit (compare with previous refactoring*.md date and Δ line counts)

Prerequisites

Read before analyzing:

• openspec/config.yaml — stack, architecture, lib isolation
• AGENTS.md — skill routing, RTL, Nx conventions
• Previous local docs if they exist (refactoring.md, refactoring-backend.md, refactoring-frontend.md)
• Archived OpenSpec for completed refactors: openspec/changes/archive/*split-*, *dial-core*, *dedupe*

Invoke ./.agents/skills/nx-workspace/SKILL.md if unsure about project names or targets.

Workflow

Copy this checklist and track progress:

Refactoring audit:
- [ ] 1. Collect metrics
- [ ] 2. Read prior docs + OpenSpec archives
- [ ] 3. Deep dive (top entries from metrics — not a fixed file list)
- [ ] 3b. Structural smells pass (else-if ladders, key dispatch, large switches, nested ternaries)
- [ ] 3c. Convention violations pass (AGENTS.md / RTL / lib isolation / imports)
- [ ] 3d. Dead-code pass (unused files, exports/types, dependencies, orphan projects)
- [ ] 4. Verify completed phases from code (grep/counts, not memory)
- [ ] 5. Write/update three docs (English) — paths from this run only
- [ ] 6. Ensure .git/info/exclude
- [ ] 7. Optional: next OpenSpec prompt
- [ ] 8. Summarize for user

Step 1 — Collect metrics

Run:

bash .claude/skills/refactoring-audit/scripts/collect-metrics.sh

This is the primary source of truth for which files to inspect. Treat every ranked list, smell, violation, and dead-code section as the candidate set for Steps 3–3d.

Run the dedicated dead-code collector separately because it may be slower and requires a local Knip installation for full coverage:

bash .claude/skills/refactoring-audit/scripts/collect-dead-code.sh

Optionally run verification (note result in docs if WIP branch):

npm exec nx test chat-api
npm exec nx test chat

Supplement with targeted greps (see repo-rules-checklist.md):

npm exec nx show projects --type=lib
ls openspec/changes | rg -v '^archive$'
rg "extends AppService|MUST stay in sync" apps/chat-api apps/chat --glob "*.{ts,tsx}"

Step 2 — Compare with prior audit

If previous docs exist:

• Carry forward completed items with OpenSpec archive evidence
• Compute Δ line counts for god modules that appear in both audits (match by path from prior doc + current metrics)
• Close items that landed (e.g. archived split-files-service, split-use-dial-file-manager)
• Do not revert completed checkboxes without code proof
• Drop prior-doc rows for files that no longer appear in metrics (refactored away)

If no prior docs: establish baseline; skip Δ column.

Step 3 — Deep dive (metrics-driven)

Do not use a fixed checklist of filenames. Instead, for each area, read the top N from Step 1 plus any smell/violation hits:

While reading, classify patterns (god service, facade already split, dispatch ladder, etc.) — not whether a file matches a historical name.

OpenSpec: active changes + recent archives since last audit date.

Step 3b — Structural smells pass (mandatory)

Uses metrics script output (see repo-rules-checklist.md for thresholds). For each hit, read the file and classify by pattern type:

Document in Structural smells sections. Columns: path (from this run), pattern type, branch/count, suggested fix, priority (usually P2 unless actively growing). Do not skip small files when metrics flagged them.

Step 3c — Convention violations pass (mandatory)

Cross-check against AGENTS.md, openspec/config.yaml, eslint.config.mjs, RTL rules. Full grep mapping: repo-rules-checklist.md.

Use metrics script violation sections + anti-patterns grep. For each non-zero category, sample-read hits and document in Convention violations (path from this run, rule violated, detail, fix, priority).

Do not treat every grep hit as debt — confirm context (tests, generated code, documented exceptions). Do flag patterns that contradict documented architecture.

Step 3d — Dead-code pass (mandatory)

Use three complementary signals; no single signal is sufficient:

1. Run TypeScript checks across every configured project to catch unused local declarations and imports (noUnusedLocals is enabled in the workspace):

   npm exec -- nx run-many --target=typecheck

2. Run scripts/collect-dead-code.sh. It performs both comprehensive and production Knip passes when a local Knip binary is available. The production pass identifies code kept alive only by tests/tooling; the comprehensive pass also covers unused test and tooling code.

3. Inspect the Nx project graph for projects with no in-repo dependents:

   npm exec -- nx graph --print \
     | jq -r '.graph as $g | ($g.nodes | keys[]) as $name | select([ $g.dependencies[]?[]? | select(.target == $name) ] | length == 0) | $name'

Classify every production Knip hit, and sample-read default-mode-only hits, as one of:

• Confirmed dead — no runtime, configuration, public API, or external-consumer reachability.
• Reachability/config gap — Knip is missing an entry, plugin, generated artifact, dynamic import, or path mapping.
• Intentional public API — exported for consumers outside this repository.
• False positive / framework-managed — loaded by NestJS metadata/DI, Vite/Webpack, Nx targets, scripts, or another convention that static imports do not expose.
• Needs owner confirmation — evidence is insufficient for safe deletion.

Review candidates in cascade order: unused production files first, then their exports/types/members, then dependencies. One unreachable file can cause all downstream symbols and dependencies to appear unused.

Coverage rules:

• Mark Full only when typecheck and both Knip passes complete, configuration hints and unresolved imports are investigated, and production hits are manually classified.
• Mark Partial when Knip is unavailable/fails, typecheck cannot complete, or reachability/configuration gaps remain. List the missing signal explicitly; never report “no dead code” from partial coverage.
• If Knip is not installed, do not download or add it automatically. Continue with TypeScript/Nx signals, mark coverage Partial, and ask before changing package.json/the lockfile when full analysis is required.
• An Nx project with zero inbound edges is only a candidate: applications are graph roots, and publishable libraries may have external consumers.
• Exclude generated chat-api-client findings. Do not hand-edit or recommend deleting generated client code.
• Never run knip --fix, delete files, or remove dependencies during an audit. Produce planning findings only unless the user separately authorizes implementation.

Document dead-code findings in a dedicated Dead code section. Include: path/package, candidate kind, source signal and mode, classification, evidence/consumer check, suggested action, and priority. Keep confirmed findings separate from unverified candidates.

Step 4 — Score phases honestly

Use the phase checklist in templates.md. Mark ✅ only with grep/count evidence from this run:

Step 5 — Write documents

Follow section structure in templates.md. English only. Populate tables from Step 1 output only — templates show column shapes, not real paths.

Quality bar:

• Every god-module row includes line count from metrics
• Every audit states dead-code coverage (Full or Partial) and the signals that ran
• Every confirmed dead-code row includes manual reachability evidence; tool output alone is not confirmation
• Priorities P0–P3 with concrete next OpenSpec change names
• Separate backend vs frontend debt
• Name archived OpenSpec changes with dates when marking complete
• Include verification note if tests were not green

Step 6 — Git exclude maintenance

Append to .git/info/exclude if missing:

refactoring.md
refactoring-backend.md
refactoring-frontend.md
docs/split-*-openspec-prompt.md
docs/*-openspec-prompt.md

Do not commit .git/info/exclude changes (it is local by design).

Step 7 — Optional OpenSpec prompt

When user wants the next refactoring step or P0 item needs OpenSpec:

• Create docs/{kebab-case-change}-openspec-prompt.md (English)
• Use the prompt template in templates.md
• Add path to .git/info/exclude

Step 8 — User summary

Reply in English with:

• Audit date
• Top 3 backend + top 3 frontend debts (from current metrics)
• Top structural smell + convention categories (with counts; name paths only from this run)
• Dead-code coverage, confirmed counts, and unverified candidate counts by kind
• What closed since last audit
• Recommended next 2–3 actions (OpenSpec names)
• Confirm docs are git-excluded

Analysis heuristics

God module candidates: service/hook/component >400 lines, or spec >1000 lines (from metrics).

Structural smell candidates: metrics smell sections — else-if ≥8, def.key === dispatch ≥3, switch ≥10 cases, nested ternary on same line.

Convention violation candidates: metrics violation sections + repo-rules-checklist greps.

Dead-code candidates: TypeScript unused-local diagnostics; Knip unused files, exports, types/members, and dependencies; Nx projects with zero in-repo dependents. Treat all as candidates until reachability and external-consumer checks are complete.

Backend smells: Express types in services, duplicate DTO enums, route sync comments, monolithic controller specs, config registry dispatch in service instead of handler map.

Frontend smells: hooks with many server-api imports, god contexts, hardcoded user-facing strings in utils, duplicated inline state when a hook exists, large switch/else-if in components.

Lib smells: libs importing app/server-api/i18n; lib total LOC >8000 without split plan; spec larger than implementation; mega-switch (≥10 cases).

Do not recommend: drive-by refactors unrelated to ranked debt; rewriting generated chat-api-client; deleting code directly from unverified static-analysis output; committing planning docs.

Incremental vs full audit

Additional resources

• Document templates: templates.md
• Repo rules → grep mapping: repo-rules-checklist.md
• Metrics script: scripts/collect-metrics.sh
• Dead-code collector: scripts/collect-dead-code.sh
• Knip audit configuration: knip-audit.json