claude-code-dispatch（Claude 代码 dispatch）

Claude Code Dispatch

Delegate coding tasks to Claude Code CLI. Use this when a task requires capabilities beyond what the current agent can do: file editing, shell commands, multi-file debugging, code review with file access, or any work that needs direct filesystem interaction.

Prerequisites

Claude Code must be installed and authenticated on the host machine. The host's ~/.claude/settings.json must pre-authorize the tools this skill will use. Example minimal config:

{
  "permissions": {
    "allow": [
      "Read",
      "Edit",
      "Glob",
      "Grep",
      "Bash(git:*)",
      "Bash(npm:*)"
    ]
  }
}

Without pre-authorized permissions, Claude Code will fail in non-interactive mode because it cannot prompt for approval.

When to Use

Route tasks to Claude Code when they require:

• File reading, editing, or creation on the host filesystem
• Shell command execution (builds, tests, git operations)
• Multi-file debugging or refactoring
• Code review with actual file access
• Any task where the agent needs to interact with the local filesystem

Do NOT use when:

• The task is purely conversational (answering questions, planning)
• You can handle it with your own capabilities or other skills
• The task involves external APIs you already have access to

Usage

bash {baseDir}/scripts/invoke-claude.sh \
  --prompt "Fix the broken import in src/App.tsx" \
  --workdir "/path/to/project" \
  --model sonnet \
  --tools "Read,Edit,Glob,Grep,Bash"

Parameters

Model Selection Guide

• sonnet (default): Fast. Good for single-file fixes, simple refactors, code review, running tests.
• opus: Slower but more capable. Use for complex multi-file debugging, architectural changes, tasks requiring deep reasoning across many files.
• haiku: Fastest and cheapest. Use for trivial tasks like renaming, formatting, or simple lookups.

Tool Presets

Common tool combinations for different task types:

• Read-only (code review, analysis): Read,Glob,Grep
• Standard editing (default): Read,Edit,Glob,Grep
• Full access (builds, tests, git): Read,Edit,Glob,Grep,Bash

Output Format

The script returns a structured summary:

STATUS: success|failure
MODEL: sonnet|opus|haiku
COST: $0.04
DURATION: 12s
RESULT: <Claude Code's response>

Exit codes: 0 = success, 1 = Claude Code error, 2 = preflight failure, 3 = timeout.

Examples

Fix a build error:

bash {baseDir}/scripts/invoke-claude.sh \
  --prompt "The build is failing with 'Cannot find module ./utils'. Diagnose and fix." \
  --workdir "/Users/dave/WebProjects/summer-camps" \
  --tools "Read,Edit,Glob,Grep,Bash"

Code review:

bash {baseDir}/scripts/invoke-claude.sh \
  --prompt "Review the changes in the last commit. Focus on correctness and security." \
  --workdir "/Users/dave/WebProjects/summer-camps" \
  --model opus \
  --tools "Read,Glob,Grep,Bash"

Run tests and fix failures:

bash {baseDir}/scripts/invoke-claude.sh \
  --prompt "Run the test suite, then fix any failing tests." \
  --workdir "/Users/dave/WebProjects/summer-camps" \
  --tools "Read,Edit,Glob,Grep,Bash" \
  --timeout 600

Limitations

• No concurrent invocations to the same working directory. Queue tasks sequentially.
• No multi-turn conversations. Each invocation is independent. For follow-up work, dispatch a new invocation with more context in the prompt.
• Result truncation. Responses longer than 2000 characters are truncated at the nearest line boundary.

Notes

• Cost is reported for observability. Subscription users are not charged per-token but the field shows what the task would cost on the API.
• Claude Code inherits environment variables from the host, including any API keys loaded by load-openclaw-env or similar scripts.
• The prompt is passed via stdin pipe, not shell interpolation, so special characters in prompts are safe.