Refactor

Client dispatch: In Codex, read codex-runtime and refactor/codex.md, execute that native procedure, and do not continue into the Claude procedure below.

Interaction rule: Every question, confirmation, or approval directed at the user — anywhere in this skill, including error recovery — MUST be asked with the AskUserQuestion tool. Never ask in plain text. If an instruction says "ask the user" or "confirm", that means AskUserQuestion.

Read the subagent-safety reference skill before delegating work to subagents.

Phase 1: Context Loading

Read project-core, resolve neutral configuration, and use the resulting config. Use root and applicable project AGENTS.md as guidance.

Progressive disclosure: Do NOT eagerly read reference docs in this Context section. Read relevant docs/<topic>.md files (and any legacy .claude/rules/lessons-learned.md if present) when you reach the analysis step that needs them.

Monorepo Context Loading

If isMonorepo is true in the resolved config:

1. Determine affected project(s): From the scope and file paths, match against the projects array in config to identify which project(s) the analysis affects.
2. Read per-project AGENTS.md for stack details and conventions.
3. Pass project context to subagents, including applicable AGENTS.md content and on-demand docs.

Shell rules: Read the shell-rules skill before running any gh commands (covers heredoc temp-file pattern).

Config Existence

Before any other checks, verify neutral-first config resolution succeeded. If not, stop immediately and tell the user: "cenci is not configured for this project. Run /cenci:configure first to set up."

Parse Scope from $ARGUMENTS

• If arguments provided → use as explicit scope (file paths, directories, globs)
• If no arguments → determine scope from recent changes:

  git diff --name-only HEAD~10..HEAD
  

  If this returns files, use them as the scope.
• If no git changes → ask the user for scope via AskUserQuestion:

  "No scope provided and no recent git changes found. Which files or directories should I analyze?"

Tip for large or unfamiliar codebases: Phase 3 fans out 3 parallel analyzer subagents per run. On a first run against an unfamiliar or large codebase, prefer scoping to a single directory or a handful of files before analyzing the whole repo — it keeps that fan-out cheap to validate before committing to a full-repo pass.

Phase 2: Scope Analysis

1. Gather target file list from the determined scope:
   • If scope is a glob pattern → expand with Glob tool
   • If scope is a directory → list all source files recursively (exclude node_modules, .git, dist, build, bin, obj, lock files, and auto-generated files)
   • If scope is specific files → use directly
2. Collect file metadata for each file:
   • Full path
   • Line count (use wc -l via Bash)
   • Language (infer from extension)
   • Whether it's a test file (matches common patterns: *.test.*, *.spec.*, *_test.*, *_spec.*, files in __tests__/, tests/, test/ directories)
3. Read project stack info from the resolved config — framework, testing library, language
4. Prepare shared context string for subagents containing:
   • File list with line counts and languages
   • Stack info (framework, testing, language)
   • Project conventions from CLAUDE.md
   • Lessons learned

Phase 3: Parallel Analysis

Launch all 3 analyzers as parallel Task tool calls in a SINGLE message.

1. Gather shared context ONCE before launching analyzers — by path, not by paste:

   • The file list with metadata (paths, line counts, languages, test-file flags) from Phase 2
   • Stack info and CLAUDE.md conventions (short summaries, not full file dumps)

   Do not read full file contents in the main agent. All three analyzers have Read/Grep/Glob and read the files they need themselves — pasting contents would duplicate the scope into the main context plus each of the three Task prompts.

2. Launch ALL THREE analyzers as three parallel Task tool calls in a SINGLE message:

   • Task 1: duplication-analyzer agent — pass the file path list; it reads and compares code blocks itself
   • Task 2: security-analyzer agent — pass the file path list + stack info + CLAUDE.md conventions
   • Task 3: structure-analyzer agent — pass the file path list + line counts + test file identification

3. Wait for all 3 to complete, then proceed to Phase 4.

Note on large scopes (20+ files): tell each analyzer which files to prioritize rather than reading them yourself:

• Files over 200 lines
• Files modified recently (from git log)
• Test files over 300 lines
• Files with common vulnerability patterns (controllers, routes, auth, API handlers)

Phase 4: Results Consolidation

Parse each agent's structured output and merge findings into a unified list.

Deduplication

If multiple analyzers flag the same file/location:

• Security + Duplication both flag same code → keep as separate findings but note the overlap
• Structure + Duplication both flag same large file → merge into a single "Split and deduplicate" finding
• Identical findings from different analyzers → keep the one with more detail, discard the other

Sorting and Grouping

1. Sort by severity: Critical → High → Medium → Low
2. Group by category: Security → Duplication → Structure
3. Assign effort estimates to each finding:
   • S (< 1hr): Simple extractions, constant consolidation, renaming
   • M (half day): File splits, shared module creation, test reorganization
   • L (1+ day): Architectural changes, security overhauls, major refactors

Numbering

Assign each finding a sequential number (F1, F2, F3...) for easy reference in Phase 5.

Phase 5: Present Findings to User

Display the consolidated report using markdown formatting.

Report Format

## Refactoring Analysis Report

**Scope**: <files/directories analyzed>
**Files analyzed**: <count>
**Total findings**: <count> (Critical: N, High: N, Medium: N, Low: N)

---

### Security Findings

#### F1. [CRITICAL] <title>
- **Location**: `path/to/file:line`
- **Issue**: <description>
- **Risk**: <what could happen>
- **Suggested fix**: <approach>
- **Effort**: S | M | L

<details>
<summary>Details</summary>
<verbose details, code examples, references>
</details>

#### F2. [HIGH] <title>
...

---

### Duplication Findings

#### F3. [HIGH] <title>
- **Pattern**: <what is duplicated>
- **Locations**: `file1:lines`, `file2:lines`
- **Occurrences**: <count>
- **Suggestion**: <extract to shared function/module/constant>
- **Effort**: S | M | L

---

### Structure Findings

#### F4. [MEDIUM] <title>
- **File**: `path/to/file` (<N> lines)
- **Issue**: <why it needs attention>
- **Suggestion**: <split/reorganize approach>
- **Effort**: S | M | L

---

### Summary
| Category | Critical | High | Medium | Low | Total |
|----------|----------|------|--------|-----|-------|
| Security | N | N | N | N | N |
| Duplication | N | N | N | N | N |
| Structure | N | N | N | N | N |
| **Total** | **N** | **N** | **N** | **N** | **N** |

Estimated total effort: <sum of effort estimates>

No Findings

If all 3 analyzers returned no findings:

"No refactoring issues found in the analyzed scope. The code looks clean!"

STOP HERE if no findings.

User Decision

After presenting the report, ask the user via AskUserQuestion:

"Which findings should I create tickets for?"

Options:

• All findings — Create tickets for everything
• Critical and High only — Skip Medium and Low
• Let me pick individually — I'll ask about each one
• None (keep report only) — Don't create any tickets

If the user selects "None" → skip to Phase 7 (Summary) with no tickets created.

Phase 6: Ticket Creation

Main-agent-only — requires auth for gh commands.

Prerequisites

Read the shell-rules skill before running any gh commands.

Extract owner/repo from git remote get-url origin.

Individual Selection Mode

If the user selected "Let me pick individually":

• For each finding, present it briefly and ask via AskUserQuestion:

  "Create a ticket for F: ? (, )" Options: Yes / No

• Collect all approved findings, then create tickets for those.

Ticket Body Template

For each approved finding, use the Write tool to author the raw title and body as plain text (never a heredoc, and never a hand-escaped JSON literal — the title is free text and must never be interpolated directly into a command line). Refactor has no ticket/PR identifier in scope, so F<N> alone would collide across concurrent runs — scope both filenames with a <run-id> (a run-unique key generated once at skill start, e.g. a timestamp literal, which is inherently safe under the shell-rules scope-key format rule) in addition to the per-finding F<N> suffix:

• ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-title.txt — content: refactor: <title>
• ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-body.md — content:

  ## Refactoring: <title>
  
  ### Category
  <Security | Duplication | Structure>
  
  ### Severity
  <Critical | High | Medium | Low>
  
  ### Description
  <finding description>
  
  ### Affected Files
  - `path/to/file1`
  - `path/to/file2`
  
  ### Suggested Approach
  <how to fix, with code examples if relevant>
  
  ### Effort Estimate
  <S | M | L> — <reasoning>
  
  ---
  *Generated by cenci refactor analysis*
  

Build the payload with the shell-rules skill's canonical jq -n --rawfile snippet, as its own standalone Bash call:

jq -n --rawfile title ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-title.txt --rawfile body ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-body.md '{title: ($title | rtrimstr("\n")), body: $body, labels: ["Refined"]}' > ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-payload.json

Then create the ticket with a separate standalone Bash call:

gh api repos/<owner>/<repo>/issues -X POST --input ${TMPDIR:-/tmp}/cenci/cenci-refactor-<run-id>-F<N>-payload.json --jq .number

The --jq .number output is the new ticket's issue number — proof the API accepted valid JSON (refine's documented create-time proof; no title re-fetch here, to avoid a single-call-site gh issue view grant).

Error Recovery

• If gh command fails with auth error → stop and tell the user to check authentication
• If ticket creation fails for one finding → report the error, continue with remaining findings
• If all ticket creations fail → report the error pattern and suggest manual creation

Phase 7: Summary

Present a final summary:

## Refactor Analysis Complete

### Tickets Created
| # | Finding | Ticket | Severity | Effort |
|---|---------|--------|----------|--------|
| F1 | <title> | #<id> | Critical | M |
| F2 | <title> | #<id> | High | S |

### Findings Not Ticketed
| # | Finding | Severity | Reason |
|---|---------|----------|--------|
| F3 | <title> | Medium | User skipped |
| F4 | <title> | Low | User skipped |

If no tickets were created (user selected "None"):

## Refactor Analysis Complete

Report generated with <N> findings. No tickets were created.
Review the findings above and run `/cenci:refactor` again when ready to create tickets.

STOP HERE. Do not offer to implement any of the refactors.