/captain-claude

⚠️ MANDATORY TIMEOUT — READ BEFORE ANY BASH CALL ⚠️

Every single Bash tool call in this entire skill MUST include timeout: 2700000 (45 minutes). No exceptions. Not even for config.sh read. The default 2-minute timeout will kill long-running commands like plan.sh and claude -p, destroying the user's work. Copy-paste this into every Bash call: "timeout": 2700000

/captain-claude

Full pipeline: Claude plans, Claude implements, Claude reviews until satisfied.

Usage

/captain-claude <task description and any ad-hoc instructions>
/captain-claude --skip-plan <path> <task description>

Flags

• --skip-plan <path> — Skip planning, use an existing plan file
• --max-rounds <n> — Cap review iterations (default: from config)
• --supervised — Pause after planning and after each review for human approval

Behavior

Parsing Input

Parse flags from the user's input. Everything else is natural language.

The user may include ad-hoc instructions for any phase inline. Look for cues like "for planning...", "when implementing...", "for review...", "the reviewer should...", "the planner should...", "make sure to...", or similar natural language directing a specific phase.

Extract:

• task description — the core task
• ad-hoc plan instructions — anything directing the planning phase
• ad-hoc implementation instructions — anything directing the implementation phase
• ad-hoc review instructions — anything directing the review phase

Read config using ${CLAUDE_PLUGIN_ROOT}/scripts/config.sh read. For each phase, merge ad-hoc instructions (if any) with the config values by appending the ad-hoc instructions to the config array.

Phase 1: Planning

1. If --skip-plan <path> was provided, verify the plan file exists and skip to Phase 2.

2. Run ${CLAUDE_PLUGIN_ROOT}/scripts/plan.sh "<task description>". Before calling, write the merged plan instructions into a temporary config override if ad-hoc plan instructions were provided. The script outputs a tab-separated line: <plan_path>\t<session_id>. Parse both values.

3. Verify the plan file exists. If missing, retry once, then report failure.

4. If --supervised was specified: present the plan and ask for confirmation before continuing.

5. Initialize state tracking:

   ${CLAUDE_PLUGIN_ROOT}/scripts/config.sh init-state "<task description>" "<plan file path>" <max_rounds> "<session_id>" <true|false for --supervised>
   

Phase 2: Implementation

Read the plan file and ${CLAUDE_PLUGIN_ROOT}/templates/implement-prompt.md.

Substitute the placeholders in the template:

• {{plan_contents}} → contents of the plan file
• {{implementation_instructions}} → merged config + ad-hoc implementation instructions, joined by newlines
• {{plan_file}} → path to the plan file

Execute the resulting prompt.

The review gate hook fires automatically when implementation tries to stop. It dispatches a separate Claude session to review, and blocks or allows the stop based on the verdict. The loop continues until the reviewer approves or max rounds (from config) is hit.

Note: ad-hoc review instructions need to be available to the hook. Write them to .captain-claude/state.json under an adhoc_review_instructions key during init-state so the hook can merge them with the config values.

Phase 3: Completion

1. Read the final state from .captain-claude/state.json
2. Present a summary: task, plan path, rounds, final verdict, review history.
3. If max rounds exceeded, offer to continue with --skip-plan.
4. Update state to "complete" or "failed".