code

name: code description: Local implementation (`/code 123`). Size auto-detection routes XS/S→patch (direct commit to main), M/L→branch+PR. Override with `--patch`/`--pr`. Does not update CLAUDE.md, run session retrospectives, or manage memory. context: fork allowed-tools: Bash(gh issue view:, gh issue edit:, gh issue list:, gh issue create:, gh api:, ${CLAUDE_PLUGIN_ROOT}/scripts/emit-event.sh:, git checkout:, git pull:, git add:, git status:, git diff:, git commit:, git push:, git merge:, git worktree:, git branch:, gh pr create:, gh pr comment:, ${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-edit.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-comment.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/run-code.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/get-issue-size.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/get-issue-type.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/opportunistic-search.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/gh-label-transition.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/worktree-merge-push.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/test-failure-classify.sh:, ${CLAUDE_PLUGIN_ROOT}/scripts/reconcile-phase-state.sh:, python3:, bats:), Glob, Grep, Read, Write, Edit, TaskCreate, TaskUpdate, TaskList, TaskGet, EnterWorktree, ExitWorktree, ToolSearch

Local Implementation

Receive an Issue number and implement based on the Spec.

If ARGUMENTS contains --help, Read ${CLAUDE_PLUGIN_ROOT}/modules/skill-help.md and follow the "Processing Steps" section to output help, then stop.

Autonomous Mode (--auto)

If ARGUMENTS contains the --auto flag, delegate as follows:

Extract the Issue number from ARGUMENTS (numeric part)
If --patch is present, run ${CLAUDE_PLUGIN_ROOT}/scripts/run-code.sh $NUMBER --patch [--base {branch}] via Bash; otherwise run ${CLAUDE_PLUGIN_ROOT}/scripts/run-code.sh $NUMBER [--base {branch}] (add --base {branch} if --base flag is present)
Exit after the script completes (do not execute subsequent steps)

If --auto is not present, proceed with mode detection below.

Note: Running /code 123 from an interactive session spawns a fork sub-agent via context: fork. Output is not streamed, but a summary is returned upon completion.

Mode Detection

If ARGUMENTS contains the --non-interactive flag, operate in non-interactive mode (set when invoked autonomously via run-code.sh). In this mode, AskUserQuestion cannot be used — see ## Error Handling in Non-Interactive Mode below for the complete policy.

In interactive mode (no flag), use AskUserQuestion to follow the normal steps.

Error Handling in Non-Interactive Mode

In --non-interactive mode (invoked via run-code.sh with claude -p), AskUserQuestion is not available (the process would hang).

Policy: At any step that would call AskUserQuestion, apply auto-resolve + log instead of aborting.

Read ${CLAUDE_PLUGIN_ROOT}/modules/ambiguity-detector.md and follow the "Non-Interactive Mode Handling" section. In summary:

auto-resolve: adopt the most reasonable default choice using model judgment, then record the decision in the Spec's ## Autonomous Auto-Resolve Log subsection
skip: for High-Stakes Decisions (see module for exhaustive list), output a warning and continue without executing the risky action
hard-error abort (exceptions — these still exit with non-zero):
- Size label is not set (cannot determine patch vs. pr route; guide to set Size label, then abort)
- Size is XL (requires sub-issue splitting before coding; guide to split the Issue, then abort)
- patch route: test FAIL persists after 1 repair attempt (cannot push failing tests to main; fix tests manually then re-run /code $NUMBER --patch, then abort)

Steps

Step 0: Route Detection

Determine the route based on Size (Project field preferred → label fallback) and option flags.

Read ${CLAUDE_PLUGIN_ROOT}/modules/phase-banner.md and display the start banner with ENTITY_TYPE="issue", ENTITY_NUMBER=$NUMBER, SKILL_NAME="code".

First, fetch Size (run before route detection):

${CLAUDE_PLUGIN_ROOT}/scripts/get-issue-size.sh "$NUMBER" 2>/dev/null

get-issue-size.sh fetches Size in two stages: Project field preferred → size/* label fallback. Use the output value (e.g., S, M; empty means Size not set) in subsequent steps (e.g., XS/S detection in Step 2).

Fetch base branch (run before route detection):

If ARGUMENTS contains --base {branch}, use that value as BASE_BRANCH. If --base is not specified, default to BASE_BRANCH=main (backward compatibility).

Flag precedence (explicit flag > Size auto-detection):

ARGUMENTS contains --patch → patch route (direct commit to BASE_BRANCH, no PR). Even if Size is XL, --patch takes precedence — skip the XL check and run as patch route
ARGUMENTS contains --pr → pr route (branch + PR flow)

Size auto-detection (when no flags are present):

Follow the Size→workflow mapping table in ${CLAUDE_PLUGIN_ROOT}/modules/size-workflow-table.md:

XS or S → patch route
M or L → pr route (branch + PR flow)
XL → exit with error ("XL requires sub-issue splitting. Split the Issue and run /code on each sub-issue.")
Size not set (empty output) and interactive mode → use AskUserQuestion to let the user choose the route (patch / pr)
Size not set (empty output) and --non-interactive mode → output error message, guide "Set Size via the Project field or size/* label, then run /code $NUMBER interactively", and abort

Record the result (patch / pr) for use in subsequent steps.

Step 1: Fetch Issue Info

gh issue view $NUMBER --json title,body

Generate a short description from the title (e.g., "add-implement-skill").

Consume comments since the last phase (L0 input):

Read ${CLAUDE_PLUGIN_ROOT}/modules/l0-surfaces.md and follow the "Comment Consumption Procedure" section with parameters: ISSUE_NUMBER=$NUMBER, COMMENT_SCOPE=issue, PHASE_NAME=code. The cutoff resolves to the most recent phase/ready label assignment. For resume runs (PR already exists), use COMMENT_SCOPE=issue+pr. Record results in the Spec's ## Consumed Comments section.

Step 2: Worktree Entry

Read ${CLAUDE_PLUGIN_ROOT}/modules/worktree-lifecycle.md and follow the "Entry section" to create a worktree.

Worktree naming convention (by route):

SSoT: This section is the single source of truth for worktree naming in /code. All scripts and modules must derive their naming from this section.

Both patch and pr routes use the same name: code/issue-$NUMBER

EnterWorktree converts / to +, resulting in:

Worktree dir: .claude/worktrees/code+issue-$NUMBER
Branch: worktree-code+issue-$NUMBER

Record the ENTERED_WORKTREE variable for use in subsequent steps.

Edit/Write path conventions inside worktree (CWD-relative):

After entering the worktree, CWD switches to the worktree directory. When editing or creating files with Edit/Write tools, verify CWD first (check with pwd), and use CWD-relative paths rather than absolute paths (e.g., ~/.claude/ or /Users/.../src/...). Using absolute paths would edit the main repository, causing missed commits and conflicts.

Step 3: `phase/ready` Label Check

Skip this check if Size is XS (XS does not require Spec — skip this entire step and proceed to Step 4).

For sizes other than XS: run gh issue view $NUMBER --json labels -q '.labels[].name' to get labels, then check if phase/ready is present.

phase/ready label present: proceed to the next step
phase/ready label absent:
- Confirm via AskUserQuestion (non-interactive mode: auto-resolve — proceed without Spec; read requirements from Issue body directly and continue; record the decision in the Spec's ## Autonomous Auto-Resolve Log subsection)
  - "Continue": proceed with execution
  - "Abort": stop processing and guide "run /spec $NUMBER", then exit

Spec precondition check (run after phase/ready passes, skip if Size is XS):

${CLAUDE_PLUGIN_ROOT}/scripts/reconcile-phase-state.sh --check-precondition code-patch $NUMBER
# pr route: replace code-patch with code-pr

Parse the JSON output. If matches_expected is false (Spec missing and Size is not XS):

Output: "Spec が見つかりません。/spec $NUMBER を実行してください"
In non-interactive mode: warn and continue (consistent with --warn-only default; Step 5 handles missing Spec by reading Issue body)
In interactive mode: abort recommended — guide user to run /spec $NUMBER first

Step 4: Create Branch & Label Transition

For pr route:

The branch was already created by Worktree Entry (EnterWorktree) in Step 2 — no need to run git checkout -b.

For patch route (direct commit to BASE_BRANCH):

The worktree was already created by Worktree Entry (EnterWorktree) in Step 2 — no explicit branch work needed.

Both routes:

${CLAUDE_PLUGIN_ROOT}/scripts/gh-label-transition.sh $NUMBER code

Step 5: Load Spec

Read ${CLAUDE_PLUGIN_ROOT}/modules/detect-config-markers.md and follow the "Processing Steps" section. Retain SPEC_PATH and STEERING_DOCS_PATH for use in subsequent steps.

Search for $SPEC_PATH/issue-$NUMBER-*.md and read the Spec to review the implementation plan.

If no Spec exists, read the requirements from the Issue body and implement accordingly.

Review notes section (only if present in Spec):

If the Spec has a "Notes" section, cross-reference each item against the implementation steps and recognize them as constraints/specifications to consider during implementation. Skip this step if there is no "Notes" section.

External auto-commit area interpretation:

When Spec Notes contains phrases such as the following, interpret as: edit the files to implement the required changes; skip git add / git commit for those specific paths (the external tool, e.g., Obsidian Git or IDE auto-commit, handles committing asynchronously):

"Do not include <path> in git add/git commit — external tool auto-commit area"
"Edit the file as required; skip git add/git commit for this path (external auto-commit)"
"External tool auto-commit area (e.g., Obsidian Git handles vault/)"

These are commit-skip instructions, not skip-implementation instructions. The files must be edited to satisfy ACs.

Phase Handoff read (after loading Spec):

Read ${CLAUDE_PLUGIN_ROOT}/modules/phase-handoff.md and follow the "Read Procedure" section. Parameters: SPEC_PATH, ISSUE_NUMBER=$NUMBER, PHASE_NAME=code.

Read existing retrospective sections (before starting implementation):

If the Spec contains any retrospective sections (e.g., ## Spec Retrospective, ## Code Retrospective) from prior phases, read them before starting implementation. These sections carry forward decisions, pitfalls, and confirmed approaches from earlier phases — reading them avoids repeating known mistakes and aligns implementation with prior design resolutions.

Step 6: Verify Uncertainties (only if present in Spec)

If the Spec has an "Uncertainties" section, before implementing:

Verify each uncertainty using the confirmation method described
If official documentation URLs are listed, review the spec information recorded there; use AskUserQuestion for any unclear points (non-interactive mode: auto-resolve — proceed with best-effort interpretation; record the unclear details and chosen interpretation in the Spec's ## Autonomous Auto-Resolve Log subsection)
If operational verification is needed, confirm with a simple verification script or bats test
If implementation approach is problematic based on verification results, report to user via AskUserQuestion (non-interactive mode: auto-resolve — proceed with the most conservative approach; record the problem details and chosen approach in the Spec's ## Autonomous Auto-Resolve Log subsection)

If uncertainties cannot be verified, or the design premise turns out to be incorrect:

In interactive mode: abort implementation and propose Spec revision
In non-interactive mode: auto-resolve by proceeding with best-effort implementation; note the unresolved uncertainty in the Spec's ## Autonomous Auto-Resolve Log subsection

Step 7: Reference Steering Documents (if present)

Use Glob with path="$STEERING_DOCS_PATH" to check whether the following steering documents exist, then Read only those that exist:

$STEERING_DOCS_PATH/tech.md — Coding conventions, Architecture Decisions (check for convention compliance), Forbidden Expressions (avoid prohibited expressions)
$STEERING_DOCS_PATH/structure.md — Directory structure, Key Files (check file placement)

If not present, skip this step and proceed to the next.

Use information from steering documents for coding convention compliance, file placement, and naming consistency during implementation. If $STEERING_DOCS_PATH/tech.md is Read, reference the ## Forbidden Expressions section and avoid using prohibited expressions in code comments, variable names, commit messages, and new documents.

Project-local Domain files (if present):

Read ${CLAUDE_PLUGIN_ROOT}/modules/domain-loader.md and follow the "Processing Steps" section with SKILL_NAME=code. Domain file content provides additional implementation guidelines and constraints.

Step 8: Implement

Implement the code following the "Implementation Steps" in the Spec.

Implementation scope: all Spec steps are required

Implement every step listed in the Spec's "Implementation Steps" regardless of the verify-type of its associated acceptance condition. A step whose AC is marked  must still be implemented in this PR — "post-merge manual" describes how the AC is verified (by human observation after merge), not whether the step can be skipped. Omitting a step because its AC is post-merge manual is an implementation error.

Use TaskCreate/TaskUpdate to manage tasks while working
Commit after each step completes

Stale Test Assertion Check

If scripts/validate-skill-syntax.py exists, Read ${CLAUDE_PLUGIN_ROOT}/skills/code/stale-test-check.md and follow the "Processing Steps" section.

Follow-up Issue Creation

If scope-out remediations are identified during implementation, create a follow-up Issue using the retro/code label:

gh issue create --title "{remediation title}" --label "retro/code" --body "..."

Follow-up Issue body format:

Background: what was found and why it is out of scope for this Issue
Purpose: what the remediation aims to improve
Acceptance Conditions: specific, verifiable conditions

Do not add the triaged label — it is assigned by /triage afterward. Skip this sub-step if no out-of-scope remediations are identified.

Step 9: Run Tests

Read ${CLAUDE_PLUGIN_ROOT}/modules/test-runner.md and follow the "Processing Steps" section to run tests.

Tier 0: Structured Test-Failure Recovery

On test FAIL, before the generic 1-repair-attempt flow below, run structured recovery:

Write the failing test output to .tmp/test-failure-recovery-$NUMBER.log.
Classify the failure before acting: run ${CLAUDE_PLUGIN_ROOT}/scripts/test-failure-classify.sh --log .tmp/test-failure-recovery-$NUMBER.log and read the category from stdout.
Route by category:
- snapshot / mock / fixture (repairable): perform at most one targeted auto-fix attempt for that class (regenerate snapshot / rebuild mock expectations / update fixture literal). No loop.
- logic / infra (not repairable): skip Tier 0 and escalate immediately to Tier 3 — fall through to the existing Step 9 FAIL handling below (in /auto this path reaches orchestration Tier 3 recovery).
Safety guard — limit changes to the tests/ directory: after the auto-fix, run git status --porcelain and confirm only tests/ paths changed. If any non-tests/ file changed, revert the Tier 0 changes (git checkout -- <files>) and fall through to the existing FAIL handling.
Re-run tests once: PASS → continue to commit; still FAIL → fall through to the existing Step 9 FAIL handling (generic 1-repair-attempt / route-specific abort/continue).
Record the Tier 0 attempt (classification, files changed, outcome) in the Spec Code Retrospective (Step 12), and append attempt details to .tmp/test-failure-recovery-$NUMBER.log for Tier 3 sub-agent reference.

Test FAIL handling (when test-runner.md reports FAIL):

Attempt to fix the failing tests (1 repair attempt)
Re-run tests
If tests still FAIL after the repair attempt:
- patch route (non-interactive mode): hard-error abort — output "Tests still FAIL after one repair attempt. Fix tests manually, then re-run /code $NUMBER --patch." and exit with non-zero
- patch route (interactive mode): use AskUserQuestion to let the user decide (abort / continue)
- pr route: continue — CI will detect the failure; report remaining failures in the completion message

Additional validation (run after tests):

If scripts/validate-skill-syntax.py exists, Read ${CLAUDE_PLUGIN_ROOT}/skills/code/skill-dev-validation.md and follow the "Processing Steps" section.

If scripts/check-forbidden-expressions.sh exists, Read ${CLAUDE_PLUGIN_ROOT}/skills/code/forbidden-expressions-check.md and follow the "Processing Steps" section.

Documentation consistency check (run after validation):

Read ${CLAUDE_PLUGIN_ROOT}/modules/doc-checker.md and follow the "Impact assessment criteria" section to determine whether documentation sync updates are needed for files changed during implementation.

If sync is required, update the target documents (README.md, docs/workflow.md, etc.) before committing.

docs/ja/ sync check (gap detection + sync): If docs/translation-workflow.md exists:

Identify all docs/*.md and docs/guide/*.md files changed in this session (excluding docs/ja/, docs/spec/, docs/stats/, docs/reports/)
For each changed file, check whether the corresponding docs/ja/ counterpart was also updated in this session
For any counterpart that was NOT updated: output "Translation sync gap: [en_file] was updated but docs/ja/[en_file_basename] was not" and then follow docs/translation-workflow.md sync procedure to update the missing counterpart
After sync, optionally run bash scripts/check-translation-sync.sh to confirm IN_SYNC status (runs against committed state — only accurate after commit)

Step 10: Verify Command Consistency

AC consistency check (run before verify-executor):

After completing implementation — especially when a refactor changes the implementation means (e.g., switching from an MCP tool to a CLI command, renaming a flag, removing a dependency) — verify that the related Issue's Acceptance Criteria verify commands still reflect the current implementation:

Re-read the AC section of Issue #$NUMBER (already fetched in Step 1)
For each  command in the AC, check whether the command's target (file path, string literal, tool name, flag name) still matches the implementation after any refactoring
If all verify commands still reflect the implementation accurately: proceed to the next check
If any verify command is stale due to a refactor:
- (a) Align implementation to AC: if the AC's original intent is correct and the refactor deviated from it, update the implementation to satisfy the AC as written
- (b) Update AC to match implementation: if the refactor was intentional and the AC's verify command now tests the wrong thing, update the Issue body verify command to reflect the new implementation means (via gh-issue-edit.sh). This is the typical case when a refactor deliberately changes the implementation approach (e.g., switched from mcp_call to a CLI command, so the allowed-tools entry referencing the old MCP tool is now stale)
- Record the decision in the commit message or PR description

Patch route verify command check:

If patch route (Size is XS/S or --patch flag), before running verify-executor, scan the Issue body's ## Acceptance Criteria > Pre-merge for github_check "gh pr checks" entries.

If found: output "Warning: patch route — github_check "gh pr checks" is incompatible. Auto-fixing to github_check "gh run list" form." and replace each with github_check "gh run list" form (add --workflow=<filename> if there are multiple workflow files under .github/workflows/). Update Issue body via gh-issue-edit.sh. Also update Spec verify commands ($SPEC_PATH/issue-$NUMBER-*.md) with the same fix.

Resolving {{base_url}} to localhost: If verify commands contain {{base_url}}, resolve it before passing to verify-executor:

If environment variable LOCAL_BASE_URL is set, use that value
If LOCAL_BASE_URL is not set, default to http://localhost:3000
Replace {{base_url}} in verify commands with the resolved URL before passing to verify-executor

Read ${CLAUDE_PLUGIN_ROOT}/modules/verify-executor.md and follow the "Processing Steps" section to run verify command consistency verification in full mode. Target: pre-merge verify commands for Issue #$NUMBER. Skip if no hints exist.

Handle results as follows:

If all PASS, complete this step and update checkboxes:
- Pre-create directory with mkdir -p .tmp
- Fetch current Issue body with gh issue view $NUMBER --json body
- For each pre-merge condition line with verify command, replace leading - [ ] with - [x] (preserve the rest of the line, verify command comments , etc.)
- Write updated body to .tmp/issue-body-$NUMBER.md with Write tool
- Update Issue body with ${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-edit.sh $NUMBER .tmp/issue-body-$NUMBER.md
- After update, delete temp file with rm -f .tmp/issue-body-$NUMBER.md
If any hints FAIL:
- Determine the FAIL cause for each failing section_contains/file_contains hint:
  - "Spec-derived literal string absent" — The Spec mandates a keyword that the implementation file should literally contain, but it is missing from the file:
    - First choice: Add the literal string to the implementation file. The verify command is correct — update the implementation to include the required text.
    - Example: section_contains "skills/verify/SKILL.md" "Step 9" "Issue OPEN" FAILs because "Issue OPEN" is absent from the implementation → add that text to the appropriate section in skills/verify/SKILL.md.
  - "Miscalibrated hint" — The keyword was incorrectly specified (wrong file path, wrong string that was never intended to be in the file, etc.):
    - Rewrite the hint with the correct one rather than removing it.
    - Example: file_contains "settings.json" "gh project" FAILs → check actual file content and rewrite to file_contains "settings.json" "Skill(triage)"
- For any hints that were rewritten (miscalibrated hint case):
  - Pre-create directory with mkdir -p .tmp
  - Write updated Issue body to .tmp/issue-body-$NUMBER.md with Write tool
  - Update with ${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-edit.sh $NUMBER .tmp/issue-body-$NUMBER.md
  - After update, delete temp file with rm -f .tmp/issue-body-$NUMBER.md
If any hints are UNCERTAIN (syntax errors, etc.):
- Display a warning and continue
- Do not fix UNCERTAIN hints — they will be re-verified in the /verify phase after merge

Spec sync (when verify commands are modified): If Issue body verify commands () are modified, also apply the same fix to the "Verification > Pre-merge" section in the Spec ($SPEC_PATH/issue-$NUMBER-*.md). Updating only the Issue body without updating the Spec will cause discrepancies flagged in the review retrospective.

Step 11: Commit, Push, or Create PR

Smoke Test (pre-commit behavioral check)

Read ${CLAUDE_PLUGIN_ROOT}/modules/verify-executor.md and follow the "Processing Steps" section to run verify commands in full mode. Target: the ## Smoke Test section of the Spec loaded in Step 5 ($SPEC_PATH/issue-$NUMBER-*.md).

If the Spec has no ## Smoke Test section, skip this subsection entirely (no-op — backward compatible with existing behavior).

When a ## Smoke Test section is present:

Emit a progress line so the watchdog resets its silence counter:
```
echo "progress: Running smoke test for issue #$NUMBER..."
```
Extract all  commands from the ## Smoke Test section and run them via verify-executor in full mode. ToolSearch is in allowed-tools so mcp_call smoke checks can invoke MCP tools.
Handle results:
- All PASS: proceed to the patch/pr commit block below.
- Any FAIL: attempt one repair, re-run smoke. If still FAIL:
  - patch route (non-interactive mode): hard-error abort — output "Smoke test still FAIL after one repair attempt. Fix the issue manually, then re-run /code $NUMBER --patch." and exit with non-zero
  - patch route (interactive mode): use AskUserQuestion to let the user decide (abort / continue)
  - pr route: continue — CI/verify will catch it; report remaining failures in the completion message
- Any UNCERTAIN or SKIPPED (permission/classifier blocked, --when condition unmet, ToolSearch unavailable, etc.): record as SKIPPED in both the Step 12 Code Retrospective "Smoke Test" row and the completion message, and proceed to commit (do not silent-adopt a fallback value; do not hard-block on the environment limitation).

For patch route (commit to BASE_BRANCH):

Type → prefix mapping (exhaustive):

Type	prefix
Bug	`fix:`
Feature	`feat:`
Task	`chore:`
Not set	`patch:`

Determine commit prefix (fetch Type → map to prefix):

Run ${CLAUDE_PLUGIN_ROOT}/scripts/get-issue-type.sh $NUMBER and get the returned Type name (Bug/Feature/Task)
If none are set (empty string): use patch: prefix

Include closes #N only when the base branch is main (GitHub auto-close via closes #N only works when merging to the default branch).

DCO compliance: use git commit -s to add Signed-off-by:. Do NOT use the global HEREDOC pattern from ~/.claude/CLAUDE.md — it omits -s.

git add <changed files>
# If BASE_BRANCH is main: "{prefix} <summary> (closes #$NUMBER)"
# If BASE_BRANCH is not main: "{prefix} <summary>"
git commit -s -m "{prefix} <summary>

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>"

git log -1 --format='%B' | grep -q "^Signed-off-by:" || { echo "ERROR: missing sign-off"; exit 1; }

Push is done in Step 13 Worktree Exit (merge-to-main pattern). Label transition happens after push completes (after Step 13).

For pr route (branch + PR):

Before creating the PR, emit a progress line so the watchdog resets its silence counter:

echo "progress: Creating PR for issue #$NUMBER..."
gh pr create --title "Issue #$NUMBER: {summary of changes}" --base "${BASE_BRANCH}" --body "..."

Include closes #N in PR body only when BASE_BRANCH is main. If BASE_BRANCH is not main, omit closes #N and instead note: "Since the base branch is {BASE_BRANCH}, please close the Issue manually at final merge."

PR title format:

Use the format Issue #N: {summary of changes} (concise English summary)
Example: Issue #48: Fix Mermaid diagram parse error

PR body should include:

Summary: overview of changes
Verification (pre-merge) — bullet list without checkboxes (- only). Informational only; checkbox management is done on the Issue side
Verification (post-merge) — same as above
closes #$NUMBER

PR body template example:

## Summary
{overview of changes}

## Verification (pre-merge)
- {item 1}
- {item 2}

## Verification (post-merge)
- {item 1}

closes #$NUMBER

Auto-append acceptance conditions to Issue:

When creating a PR, compare the Spec verification methods against the Issue acceptance conditions. If the Issue acceptance conditions are missing verification items, fetch the current Issue body with gh issue view $NUMBER --json body, build the updated body with the missing items appended, pre-create the directory with mkdir -p .tmp, write the updated body to .tmp/issue-body-$NUMBER.md with Write tool, update with ${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-edit.sh $NUMBER .tmp/issue-body-$NUMBER.md. After update, delete the temp file with rm -f .tmp/issue-body-$NUMBER.md.

Change tracking comment (post-PR):

After creating the PR, if any of the following changes occurred during Step 10 or Step 11, post a change tracking comment to the Issue:

Step 10 case 2: verify command was rewritten (FAIL → corrected hint)
Step 10 Spec sync: Spec verify commands were updated to match Issue body
Step 11 auto-append: acceptance conditions were appended to the Issue body

If none of the above changes occurred, skip this step.

When changes occurred, write the comment body to .tmp/change-tracking-$NUMBER.md with the Write tool, then post:

mkdir -p .tmp
${CLAUDE_PLUGIN_ROOT}/scripts/gh-issue-comment.sh $NUMBER .tmp/change-tracking-$NUMBER.md
rm -f .tmp/change-tracking-$NUMBER.md

Comment format:

## Change Tracking (by /code)

### Changes Made
- {change summary 1 — what was changed and why}
- {change summary 2}

Step 12: Code Retrospective

Append retrospective information to the Spec and commit.

Content to append:

Deviations from the design (what deviated and why)
Design gaps/ambiguities (problems found during implementation)
Rework (where rework occurred and the cause)

Retrospective writing discipline:

One entry per learning — do not bundle multiple lessons into a single bullet
Record both corrections (what was wrong) and confirmed approaches (what worked and why)
Link related entries across retrospective sections when one finding affects another
Do not duplicate what the repository or git history already records (commit messages, PR descriptions, file diffs) — note only the reasoning and judgment that is not captured elsewhere
Update or delete entries found to be incorrect in subsequent runs; stale or wrong entries degrade memory quality

Template:

## Code Retrospective

### Deviations from Design
- (deviation and reason)

### Design Gaps/Ambiguities
- (problems found during implementation)

### Rework
- (where rework occurred and the cause)

Sync Spec implementation steps (when deviations exist):

If there are items under "Deviations from Design" (reordering of implementation steps, omission/consolidation of steps, adoption of a different approach due to design changes, etc.), in addition to recording in the retrospective, also update the Spec "Implementation Steps" section itself to match the actual implementation. This allows the subsequent /verify phase to verify based on the latest implementation.

No deviations: no need to update Spec implementation steps
Deviations exist: revise Spec implementation steps to match actual implementation and include in the same commit

Steps:

If no retrospective information, write "N/A"
Append ## Code Retrospective section after ## Spec Retrospective in the Spec ($SPEC_PATH/issue-$NUMBER-*.md) using the Edit tool
If scripts/check-forbidden-expressions.sh exists, Read ${CLAUDE_PLUGIN_ROOT}/skills/code/forbidden-expressions-check.md and follow the "Retrospective Guard" section.
If "Deviations from Design" exist, also update the "Implementation Steps" section in the Spec to match the actual implementation
Phase Handoff write (before commit): Read ${CLAUDE_PLUGIN_ROOT}/modules/phase-handoff.md and follow the "Write Procedure" section. Parameters: SPEC_PATH, ISSUE_NUMBER=$NUMBER, PHASE_NAME=code. The handoff is staged with the Spec in the same git add and committed together.

Commit (push is done in Step 13 Worktree Exit):

git add $SPEC_PATH/issue-$NUMBER-*.md
git commit -s -m "Add code retrospective for issue #$NUMBER

Co-Authored-By: Claude Sonnet 4.6 noreply@anthropic.com"

```bash
git log -1 --format='%B' | grep -q "^Signed-off-by:" || { echo "ERROR: missing sign-off"; exit 1; }

For pr route: also push from within the worktree (emit a progress line first so the watchdog resets its silence counter):
```
echo "progress: Pushing branch to origin for issue #$NUMBER..."
git push origin HEAD
```

Step 13: Worktree Exit

Read ${CLAUDE_PLUGIN_ROOT}/modules/worktree-lifecycle.md and follow the Exit section appropriate for the route.

patch route (merge-to-main pattern): Follow "Exit: merge-to-main section". After push completes, transition the label (patch route skips /merge, so label transition happens here).

patch route (XS/S common): After push completes, transition to phase/verify:

${CLAUDE_PLUGIN_ROOT}/scripts/gh-label-transition.sh $NUMBER verify

patch route completes here. Follow the completion report section to inform the user.

pr route (push-and-remove pattern): Follow "Exit: push-and-remove section" (push was done in Step 12, so only delete the worktree).

Step 14: Opportunistic Verification

Only if .wholework.yml in the project has opportunistic-verify: true, Read ${CLAUDE_PLUGIN_ROOT}/modules/opportunistic-verify.md and follow the "Processing Steps" section to run opportunistic verification. The skill name is /code. Skip this step if not configured.

Completion Report

Output the route-specific prefix, then read ${CLAUDE_PLUGIN_ROOT}/modules/next-action-guide.md and follow the "Processing Steps" section.

patch route prefix: "Direct commit and push to main complete."
pr route prefix: "PR creation complete."

Parameters to pass to next-action-guide:

SKILL_NAME=code
ISSUE_NUMBER=$NUMBER
PR_NUMBER={PR number if pr route}
ROUTE={patch|pr}
RESULT=success

Notes

Direct work on the main branch is only allowed for the patch route (size/XS·S or --patch flag). For the pr route, always create a branch
Prioritize the Spec for reference
If tests fail, fix them before continuing
Always create and write temp files with the Write tool. Creating or writing temp files via Bash cat/echo/redirect (>) is prohibited (causes confirmation prompts)
bats test @test names must be in English (ASCII). Multibyte characters (Japanese, etc.) cause test name parse failures and result in 0 tests executed. See: #226
Brace expansion ({1,2,3}) is prohibited. Use globs for deleting multiple files: rm -f .tmp/issue-*.md (brace expansion triggers Claude Code security warnings)
Files under .claude/ are treated as sensitive files by Claude Code — Edit and Write tools are rejected for these paths. When implementation requires editing .claude/ files (e.g., settings.json.template), use Bash (e.g., python3 -c "..." or sed -i) instead
External auto-commit areas (directories committed asynchronously by external tools such as Obsidian Git or IDE auto-commit): edit the files as required to satisfy ACs, but skip git add / git commit for those specific paths. The external tool handles committing. A Spec Notes instruction like "do not include in git commit procedure" is a commit-skip instruction only — not skip-implementation.