ship

name	ship
allowed-tools	Bash(git:*), Bash(gh:*), Bash(tea:*), Bash(mkdir:*), Bash(grep:*), Bash(awk:*), Read, Edit, Glob, Grep, Write
description	Create branch, commit, push, open PR, auto-review, fix issues, and merge — the full ship workflow. Pushes and can merge, so it runs only when invoked explicitly.
disable-model-invocation	true

You are automating the complete git workflow to ship code changes. After creating the PR, you will automatically review it, fix any issues, and merge it. The user may provide a branch name and description as arguments: $ARGUMENTS

Pre-loaded Context

The following git state is injected before Claude processes this prompt. Injection is a load-time precondition — a non-zero exit aborts prompt expansion and the skill never reaches Claude (ADR-0011 F3), so every command below is guarded to exit 0 everywhere and reports the sentinel (not a git repository) instead of failing.

Working tree status: !git status -s 2>/dev/null || echo "(not a git repository)"

Diff summary (unstaged): !git diff --stat 2>/dev/null || echo "(not a git repository)"

Staged diff summary: !git diff --cached --stat 2>/dev/null || echo "(not a git repository)"

Current branch: !git branch --show-current 2>/dev/null || echo "(not a git repository)"

Remote: !git remote -v 2>/dev/null || echo "(not a git repository)"

Diff size (lines changed): !git diff HEAD --shortstat 2>/dev/null | grep -oE '[0-9]+ (insertions?|deletions?)' | awk '{s+=$1} END {print s+0}'

The diff-size value is always a bare integer. It counts insertions plus deletions, and git diff HEAD covers staged and unstaged changes, so the number is correct in the normal pre-ship state where everything is already git add-ed. Outside a repository it is 0.

Destructive Action Warning

This skill modifies git state (creates branches, commits, pushes, and merges). Before proceeding:

• Confirm the user intends to ship the current changes
• Never force-push or push to main/master directly
• Always create a feature branch for changes
• If the working directory has uncommitted changes that look unrelated, ask before staging everything

Input Validation

Optional Arguments:

• <branch-name> - Custom branch name (default: auto-generated from changes)
• draft - Create PR as draft
• --dry-run - Preview all operations without making any changes
• --audit - Log all git and PR operations to .claude-plugin/audit.log (see common-patterns.md)

Dry-Run Mode: When --dry-run is specified:

• Show what branch name would be created
• Show what files would be staged and committed
• Show the proposed commit message
• Show the PR title and body that would be created
• Prefix all output with [DRY-RUN] to clearly indicate preview mode
• Do NOT execute any git commands that modify state (checkout, add, commit, push, pr create)

Audit Mode: When --audit is specified:

• Log every git and PR operation to .claude-plugin/audit.log
• Each log entry is a JSON line with: timestamp, command, action, details, success
• Create .claude-plugin/ directory if it doesn't exist
• Append to log file (never overwrite existing entries)

Example log entries:

{"timestamp": "2026-01-14T10:30:00Z", "command": "ship", "action": "git_checkout", "details": {"branch": "feat/new-feature"}, "success": true}
{"timestamp": "2026-01-14T10:30:01Z", "command": "ship", "action": "git_commit", "details": {"sha": "abc1234", "message": "feat: add new feature"}, "success": true}
{"timestamp":            
            
            

Pre-flight Checks

Use the pre-loaded context injected above — do NOT re-run these git commands:

1. Verify git repository — if any injected value above is the literal sentinel (not a git repository), abort with: ship must be run inside a git repository. Do NOT infer this from empty output: empty is a legitimate result for several of these commands inside a valid repo.
2. Verify a remote is configured — if the injected Remote block is empty (and not the sentinel), the repository is valid but has no remote. Abort with: No git remote configured. Run: git remote add origin <url> — there is nothing to push to and Phase 0 cannot detect the platform.
3. Confirm uncommitted changes — check the injected git status -s and diff summaries; if both are empty, abort with a clear message
4. Confirm current branch — check the injected branch name; if not main, ask the user if they want to proceed from the current branch or abort
5. Diff size gate — the injected "Diff size (lines changed)" value is an integer. If it is greater than 500, note this for Phase 6 (will suggest /code-review ultra instead of standard review)

Phase 0: Platform Detection

Detect the git hosting platform and select the appropriate CLI:

1. Parse the injected git remote -v output (pre-loaded above) for the push remote URL — do NOT re-run the command
2. If URL contains github.com → set PLATFORM=github
   • Verify gh auth status succeeds
   • If gh is not installed or not authenticated, abort with install/auth instructions
3. Otherwise → set PLATFORM=gitea
   • Verify tea login list shows a login matching the remote URL's host
   • If tea is not installed, abort with: Install tea CLI: https://gitea.com/gitea/tea
   • If no matching login exists, abort with: Run: tea login add --url <host-url> --name <name> --token <token>
4. Store the platform choice for all subsequent steps
5. Display: Platform detected: [GitHub|Gitea] (using [gh|tea] CLI)

Draft PR limitation: The tea CLI does not support --draft for PR creation. When the draft argument is passed on a Gitea repo, warn the user that draft PRs are not supported on Gitea and create a normal PR instead.

Execution Steps

Phase 1: Determine Branch Name

• If the user provided a branch name in arguments, use it
• If not, analyze the staged/unstaged changes and generate a descriptive kebab-case branch name (e.g., fix-login-validation, add-user-export-feature)
• Confirm the branch name with the user before proceeding

Phase 2: Create and Switch to New Branch

git checkout -b <branch-name>

Phase 3: Documentation Gate, Stage, and Commit

3.1: Documentation Gate

Before staging changes, verify project documentation is current. LAB_NOTEBOOK.md is a hard gate — do not proceed to staging until notebook requirements are met.

LAB_NOTEBOOK.md (Hard Gate):

1. Check if LAB_NOTEBOOK.md exists in the project root
2. If it does NOT exist — skip this gate entirely
3. If it exists: a. Read the project's CLAUDE.md AND the system-level ~/.claude/CLAUDE.md for lab notebook rules — the project's CLAUDE.md contains the "Lab Notebook — MANDATORY Logging Protocol" section with the project's logging rules, entry template, and project-specific tags; the system CLAUDE.md may contain additional constraints b. Read the latest entries in LAB_NOTEBOOK.md (the Experiment Log section) c. Analyze the current changes using the pre-loaded diff summaries injected at the top of this prompt (do NOT re-run git diff or git diff --cached — use the injected output) d. Determine if there's a current entry covering the changes being shipped:
   • An entry dated today whose objective relates to the changes
   • An IN PROGRESS entry that covers the current work e. If no current entry exists — create one:
   • Add a new entry following the notebook's template (next sequential Entry number)
   • Include: Date, Environment (from recent entries or system context), Objective (derived from the changes), Status: COMPLETE
   • Summarize what was done in Actions & Results
   • For Hypothesis: reconstruct from the changes — what was the expected outcome?
   • For Rollback Plan: reference the git SHA before the changes (git rev-parse HEAD)
   • Update the Decision Log and Action Items tables at the top if applicable f. If a current entry exists but is IN PROGRESS — close it out:
   • Update the entry: add final results, set Status to COMPLETE
   • Fill in Duration if empty
   • Update What Worked / What Failed based on outcomes g. Notebook updates will be included in the commit at step 3.2

Other Documentation (Soft Warnings — non-blocking):

Scan for documentation that may need attention. Report but do NOT halt:

• CHANGELOG.md — warn if it exists and changes include new features or breaking changes
• README.md — warn if changes add new user-facing features, commands, or endpoints

Display results before proceeding:

Documentation Gate:
  LAB_NOTEBOOK.md: [Updated entry E{NNN} | Created entry E{NNN} | Not found (skipped) | Current (no update needed)]
  Warnings: [list, or "None"]

3.2: Stage and Commit

• Stage all changes (including any notebook updates from 3.1): git add -A
• Analyze the diff to generate a clear, conventional commit message
• Format: <type>: <concise description> (e.g., feat: add CSV export for user data)
• Include a body if changes are complex enough to warrant explanation
• Show the user the proposed commit message and proceed unless they object

Phase 4: Push to Remote

git push -u origin <branch-name>

Phase 5: Create Pull Request

• Set the PR title to match or expand on the commit message
• Generate a PR body that includes:
  • Summary of what changed and why
  • Key files modified
  • Any testing notes if apparent from the changes
• Target branch: main

GitHub:

gh pr create --title "..." --body "..." [--draft]

Gitea:

tea pr create --title "..." --description "..."
# Note: --draft is not supported by tea CLI
# If user requested draft, warn and create normal PR

• Open the PR as a draft if the user included "draft" in their arguments (GitHub only)

Output

After completion, display:

• Branch name created
• Commit SHA
• Direct link to the PR

Error Handling

• If any git command fails, stop immediately and show the error
• If push fails due to remote rejection, suggest possible causes
• If PR creation fails, provide the manual gh pr create or tea pr create command the user can run

---

Phase 6: Auto-Review

After the PR is created, automatically analyze it for issues.

6.1 Fetch PR Information

GitHub:

# Get the PR number from the just-created PR
PR_NUMBER=$(gh pr view --json number -q '.number')

# Fetch the diff for analysis
gh pr diff $PR_NUMBER

Gitea:

# Get the PR number (parse from tea pr create output, or list open PRs)
PR_NUMBER=$(tea pr list --output json --fields index --state open | jq '.[0].index')

# Fetch the diff for analysis
tea pr view $PR_NUMBER --fields diff --output simple

6.2 Analyze the PR

Perform a comprehensive review across these dimensions:

Security Analysis (CRITICAL/WARNING)

• Hardcoded secrets, API keys, or credentials
• SQL injection vulnerabilities
• XSS vulnerabilities
• Insecure dependencies
• Missing input validation
• Authentication/authorization issues
• Sensitive data exposure in logs

Performance Analysis (WARNING)

• N+1 query patterns
• Unbounded loops or recursion
• Large file reads into memory
• Missing pagination
• Inefficient algorithms (O(n²) when O(n) possible)
• Blocking operations in async contexts

Code Quality Analysis (WARNING)

• DRY violations (copy-paste code)
• Functions/methods over 50 lines
• High cyclomatic complexity
• Inconsistent naming conventions
• Missing error handling
• Magic numbers without constants
• Dead code or unused imports

Test Coverage Analysis (WARNING)

• New code paths without tests
• Modified code with outdated tests
• Missing edge case coverage

Documentation Analysis (WARNING for public APIs)

• Public APIs without documentation
• Complex logic without comments
• Missing README updates for new features

6.3 Classify Issues

Categorize each issue by severity:

• CRITICAL: Must fix - blocks merge (security vulnerabilities, data integrity risks)
• WARNING: Should fix - blocks merge (quality issues, missing tests)
• SUGGESTION: Nice to have - does NOT block merge

6.4 Output

Large-diff gate: If the injected "Diff size (lines changed)" value from pre-flight is > 500, display this recommendation before the standard review output and skip the automated fix loop — route to /code-review ultra instead:

Phase 6: Large Diff Detected
==============================
This PR changes [N] lines (>500). Standard auto-review may miss subtle issues.

Recommendation: Run `/code-review ultra` for multi-agent deep review on this PR.
  - /code-review ultra provides more thorough analysis for large changes
  - Auto-fix loop skipped for large diffs (too many moving parts for reliable auto-fix)

PR URL: [url] (open — awaiting manual review)
Branch: [branch-name] (preserved)

For diffs ≤ 500 lines, display the standard review output:

Phase 6: Auto-Review
====================
PR #[number] analyzed.

Issues Found:
  Critical: [N] (must fix)
  Warnings: [N] (should fix)
  Suggestions: [N] (non-blocking)

[If no blocking issues]
✓ All checks passed! Proceeding to merge.

[If blocking issues exist]
Found [N] blocking issues. Starting fix loop.

---

Phase 7: Fix Loop

If there are CRITICAL or WARNING issues, attempt to fix them automatically.

7.1 Loop Parameters

• Maximum attempts: 5
• Blocking issues: CRITICAL + WARNING only (suggestions don't block)
• Exit conditions:
  • All blocking issues resolved → proceed to merge
  • Unfixable issue detected → report and stop
  • Max attempts reached → report exhaustion and stop

7.2 Fix Loop Logic

• Attempt a fix for every blocking issue; mark unfixable ones per 7.4
• Any unfixable issue → exit immediately to the Phase 8 failure report (no commit that attempt)
• Otherwise commit the fixes, push, and re-analyze the PR
• All blocking issues resolved → exit to the Phase 8 success path; else continue to the next attempt
• 5 attempts exhausted with blockers still remaining → exit to the Phase 8 exhaustion report

Full pseudocode walkthrough: see the Fix Loop Pseudocode template in references/ship-output-templates.md.

7.3 Fix Strategies by Issue Type

Issue Type	Fix Approach
Hardcoded secrets	Replace with process.env.VAR_NAME reference
SQL injection	Convert to parameterized query
XSS vulnerability	Add sanitization/escaping
N+1 queries	Add eager loading or batching
Long functions	Split into smaller focused functions
Missing error handling	Add try/catch with appropriate handling
Magic numbers	Extract to named constants
Missing tests	Generate basic test stubs
Missing docs	Generate JSDoc/docstring comments

7.4 Unfixable Issue Detection

Mark an issue as "unfixable" if:

• Requires external changes: Database schema, external API, infrastructure
• Architectural decision needed: Multiple valid approaches, needs human choice
• Cyclical issue: Same fix keeps being undone or introduces new issues
• Beyond scope: Would require changes outside the PR's files
• False positive: Code is actually correct, analysis was wrong

When unfixable:

Unfixable: [ID] [title]
Reason: [why it cannot be auto-fixed]
Suggestion: [what the user should do manually]

---

Phase 8: Completion

Three possible outcomes: success (merge), failure (unfixable), or exhaustion (max attempts).

8.1 Success Path (All Blocking Issues Resolved)

Execute merge:

# --- GitHub ---
gh pr merge $PR_NUMBER --squash --delete-branch

# --- Gitea ---
tea pr merge $PR_NUMBER --style squash
git push origin --delete $BRANCH_NAME   # tea doesn't auto-delete the branch

# --- Both platforms ---
git checkout main
git pull

# Clean up local branch if it exists
git branch -d $BRANCH_NAME 2>/dev/null || true

Prune stale branches:

# Fetch and prune remote tracking branches that no longer exist on remote
git remote prune origin

# Find local branches where upstream is gone and delete them (merged only)
git branch -vv | grep ': gone]' | awk '{print $1}' | xargs -r git branch -d 2>/dev/null

Note: Only fully merged branches are deleted (-d not -D). Unmerged branches with deleted remotes are preserved and reported as warnings.

Emit the completion summary using the Phase 8.1 Success template in references/ship-output-templates.md.

8.2 Failure Path (Unfixable Issues Exist)

Do NOT merge. Report to the user using the Phase 8.2 Failure template in references/ship-output-templates.md.

8.3 Exhaustion Path (Max Attempts Reached)

Do NOT merge. Report diagnostics to the user using the Phase 8.3 Exhaustion template in references/ship-output-templates.md.

---

Performance

Scenario	Expected Duration
Clean PR, no issues	30-60 seconds (branch, commit, push, PR create, review, merge)
PR with fixable issues	1-3 minutes (includes fix loop iterations)
PR with CI checks	2-10 minutes (depends on CI pipeline duration)
Max fix loop (5 attempts)	3-5 minutes (excludes CI wait time)

Duration scales with the number of changed files (affects review time) and CI pipeline configuration. Gitea operations may be slightly slower than GitHub due to CLI differences.

Summary of Workflow

Phase	Action	Outcome
Pre-flight	Verify git, gh/tea, changes	Ready to ship
Phase 0	Detect platform (GitHub/Gitea)	CLI selected
Phase 1	Determine branch name	Branch name confirmed
Phase 2	Create branch	On new branch
Phase 3.1	Documentation gate	LAB_NOTEBOOK.md updated (if present), doc warnings reported
Phase 3.2	Stage and commit	Changes committed (including doc updates)
Phase 4	Push to remote	Branch pushed
Phase 5	Create PR	PR opened
Phase 6	Auto-review PR	Issues identified
Phase 7	Fix loop (up to 5x)	Issues fixed or marked unfixable
Phase 8	Complete	Merged, branches pruned, or failure report