/dev

Generates source code based on CLAUDE.md.

Triggers

• /dev
• code generation
• code from CLAUDE.md

Arguments

Name	Required	Default	Description
--path	No	.	Target path
--all	No	false	Target all CLAUDE.md files (instead of incremental)
--conflict	No	skip	File conflict handling: skip | overwrite
--dry-run	No	false	Display targets only without generating actual files
--validate	No	false	Automatically run /validate after compilation

Workflow

0. Initialization

CLI_PATH=$("${CLAUDE_PLUGIN_ROOT}/scripts/install-cli.sh")
TMP_DIR=".claude/tmp/${CLAUDE_SESSION_ID:+${CLAUDE_SESSION_ID}/}"
mkdir -p "$TMP_DIR"

1. Determine dev targets

--all mode:

Glob("{path}/**/CLAUDE.md")

Incremental mode (default):

$CLI_PATH diff-compile-targets --root {path}

Result branching:

• Not a git repository → fallback to all targets
• No changes → "All up-to-date. Use --all for full dev." → exit
• Changes found → display target list + reasons

Exit if no targets.

2. Auto-detect language

Analyze file extensions in each target directory to infer the language:

1. Source file extensions in the directory → determine language
2. If no source files, reference parent directory
3. If all fail, ask via AskUserQuestion

3. Check dev-context (optional)

If a dev-context.md exists in the same directory as each CLAUDE.md, use it as reference. Normal operation continues even if absent.

4. Determine dependency order (leaf-first)

Sort by directory depth (deepest first). Independent modules at the same depth can be executed in parallel (up to 3).

5. --dry-run handling

Output only the target list and exit:

Dev targets:
  • src/auth/jwt (depth=3, typescript)
  • src/auth (depth=2, typescript)
  • src/utils (depth=2, typescript)

6. Create session files

For each target, read CLAUDE.md + DEVELOPERS.md + Convention hierarchy and create session files:

0. (--all excluded) Search for spec commits — execute per target directory: a. Find last dev commit:

   LAST_DEV=$(git log -1 --format="%H" --grep="^dev({path}):" 2>/dev/null || echo "")
   

   b. Find spec commits after that:

   if [ -n "$LAST_DEV" ]; then
     SPEC_COMMITS=$(git log --format="%H" --grep="^spec({path}):" ${LAST_DEV}..HEAD 2>/dev/null)
   else
     SPEC_COMMITS=$(git log --format="%H" --grep="^spec({path}):" 2>/dev/null)
   fi
   

   c. If found — extract diff + message for each spec commit:

   # Extract diff (root commit guard)
   PARENT=$(git rev-parse --verify {hash}~1 2>/dev/null || echo "")
   if [ -n "$PARENT" ]; then
     git diff {hash}~1..{hash} -- {path}/CLAUDE.md {path}/DEVELOPERS.md
   else
     git diff --root {hash} -- {path}/CLAUDE.md {path}/DEVELOPERS.md
   fi
   
   # Extract commit message
   git log -1 --format="%B" {hash}
   

   d. If not found: do not include Spec Changes section in the session file
1. Read target CLAUDE.md → extract Requirements, Domain Context
2. Read target DEVELOPERS.md → extract Constraints, Technical Context
3. Resolve Convention hierarchy (module > project > general)
4. Read dev-context.md (optional) → extract Dependencies, approach
5. Write session file → ${TMP_DIR}dev-session-{dir-safe}.md
6. (If spec commits found in sub-step 0) Add Spec Changes section:
   • Extract transition context from commit message body → ### Transition Context
   • Parse commit message Changes section → ### Added, ### Modified, ### Removed
   • If BREAKING flag present → add breaking: true metadata

Session file format: see "Dev Session File Format" in dev-templates.md.

6e. Derive Implementation Tasks (only when Spec Changes present)

When the session file contains ## Spec Changes:

1. Added → [ADD] task: tests+implementation needed for new Constraint/Requirement
2. Modified → [MODIFY] task: modify tests+implementation to match changed Constraint/Requirement
3. Removed → [DELETE] task: remove code+tests related to deleted Constraint/Requirement

Add ## Implementation Tasks section to the session file:

## Implementation Tasks (only when Spec Changes present)
- [ADD] CONST-N: {description}
- [MODIFY] CONST-N: {change details}
- [DELETE] CONST-N: {deletion target}

6f. Execute [DELETE] tasks (only when present)

SKILL handles DELETE directly before the TDD pipeline:

1. Search for imports/references of deletion targets via Grep
2. Collect list of referencing files
3. Delete target files/functions (Bash rm or Edit)
4. Remove imports/calls from referencing files (Edit)
5. Delete related test files
6. Run regression tests → report warning on failure

7. Test Writing Loop (per target, sequential per module)

round = 1, max_safety = 5

loop:
  7a. Create test-writer session file:
      ${TMP_DIR}test-writer-session-{dir-safe}.md
      (format: see "Test Writer Session File Format" mode=write in dev-templates.md)

  7b. Dispatch Task(test-writer):
      Session file: ${TMP_DIR}test-writer-session-{dir-safe}.md
      Save results to ${TMP_DIR} and return only the path

      Extract test_dir, mapping_file from result block.

  7c. Create test-reviewer session file:
      ${TMP_DIR}test-reviewer-session-{dir-safe}-v{round}.md:

      ```markdown
      # Test Review Session
      type: test-review | round: {round} | language: {lang}
      dir_safe: {dir-safe}
      mapping_file: ${TMP_DIR}test-mapping-{dir-safe}.json
      test_dir: ${TMP_DIR}tests/{dir-safe}/
      spec_session_file: ${TMP_DIR}dev-session-{dir-safe}.md
      ```

  7d. Dispatch Task(test-reviewer):
      Session file: ${TMP_DIR}test-reviewer-session-{dir-safe}-v{round}.md
      Save results to ${TMP_DIR} and return only the path

      Extract verdict from result block.

  7e. if verdict == "approved":
        break → Step 7.5

  7f. if round >= max_safety:
        ⚠ Test review loop terminated after {max_safety} iterations.
          Proceeding with best-effort tests.
        break → Step 7.5

  7g. Create Revise session file:
      ${TMP_DIR}test-writer-session-{dir-safe}.md (overwrite):
      Change mode to revise, increment round, add feedback_file
      (format: see "Test Writer Session File Format" mode=revise in dev-templates.md)

  7h. Dispatch Task(test-writer, mode=revise):
      Session file: ${TMP_DIR}test-writer-session-{dir-safe}.md
      Save results to ${TMP_DIR} and return only the path

  7i. round++ → return to 7c

7.5. Copy TMP → target + Verify RED

7.5a. Copy TMP/tests/{dir-safe}/ → target directory
      Copy based on test_files paths in mapping.json

7.5b. Verify RED (SKILL executes directly via Bash):
      Run tests per language:
      | Language | Command |
      | TypeScript | npx jest --passWithNoTests 2>&1 |
      | Rust | cargo test --no-run 2>&1 (compile only) |
      | Python | python -m pytest --collect-only 2>&1 |
      | Go | go test -run "^$" ./... 2>&1 (compile only) |

7.5c. All fail confirmed → proceed to Step 8
7.5d. Some pass → record as existing implementation coverage, proceed to Step 8
7.5e. Compilation itself fails (import errors, etc.) → delegate to green-coder (import fix allowed)

8. Task(green-coder)

Create green-coder session file:
${TMP_DIR}green-session-{dir-safe}.md
(format: see "Green Coder Session File Format" in dev-templates.md)
Include conflict mode from dev arguments in session file header (default: skip)

Dispatch Task(green-coder):
  Session file: ${TMP_DIR}green-session-{dir-safe}.md
  Target directory: {path}
  Detected language: {language}
  Save results to ${TMP_DIR} and return only the path

Check green-result status:
- success: proceed to Step 9
- partial: collect warnings, proceed to Step 9
- failed: report error, move to next module

9. Task(refactorer)

Create refactorer session file:
${TMP_DIR}refactor-session-{dir-safe}.md
(format: see "Refactorer Session File Format" in dev-templates.md)
Implementation Files: implemented_files from green-result

Dispatch Task(refactorer):
  Session file: ${TMP_DIR}refactor-session-{dir-safe}.md
  Target directory: {path}
  Detected language: {language}
  Save results to ${TMP_DIR} and return only the path

refactor-result status:
- success: continue
- rolled_back: record warning (green-coder results preserved)
- skipped: continue

10. Build verification

After all modules complete, run type check based on detected language:

Language	Command
Rust	cargo check --workspace 2>&1
TypeScript/JavaScript	tsc --noEmit 2>&1 (only when tsconfig.json exists)
Python	python -m py_compile $(find src -name "*.py") 2>&1
Other	Skip (warning only)

Success: proceed.

Failure:

1. Extract affected files from error message
2. Report:

   [BUILD FAILED] {error summary}
   Affected files: {file list}
   Recommended action: Review DEVELOPERS.md Constraints for the affected module and re-run /dev
   

3. Return dev status = failed, skip subsequent Steps

Limitation: If new files are not declared in mod.rs/lib.rs, cargo check will not inspect those files. The green-coder agent must always add mod declarations when creating new files.

11. Display changes

git diff --stat

12. Create dev commit

If compilation completed successfully (status != failed), create individual commits per target directory (no consolidated commits):

# Repeat for each dev target
git add {created/modified files in the target directory}
git commit -m "dev({path}): {summary}

{1-2 sentence summary of compiled content}

Changes:
- compiled: {list of generated files}
- tests: {list of generated test files}"

This commit becomes the reference point for git log --grep="^dev({path}):" searches, so individual commits per path are mandatory.

13. Post-dev verification (optional)

If --validate flag is present:

Skill("claude-md-plugin:validate", args: "{path}")

14. Result

---dev-result---
status: success | partial | failed
total: {n}
generated: {n}
skipped: {n}
tests: {passed} passed, {failed} failed
validate: {status} (when --validate is used)
---end-dev-result---

DO / DON'T

DO:

• Follow leaf-first order
• Complete Convention hierarchy resolution when creating session files
• Follow test-writer → test-reviewer → green-coder → refactorer order
• Handle DELETE tasks directly via SKILL before the TDD pipeline

DON'T:

• Modify CLAUDE.md (read-only)
• Pass CLAUDE.md path directly to Agent (pass via session file)
• Enter green-coder without test-reviewer approval
• Use compiler agent (deprecated)

Error Handling

Situation	Response
CLI build failure	install-cli.sh handles automatic build
CLAUDE.md not found	guidance message, exit
test-writer failure	warn, continue with the rest
test-reviewer max_safety reached	best-effort proceed, warn
green-coder failure (single module)	warn, continue with the rest
refactorer regression failure	rollback, warn
Build verification failure (Step 10)	report error, status=failed
Language detection failure	AskUserQuestion