claude-delegate

Delegate coding, repository analysis, file edits, test runs, or code review to the local Claude Code CLI without embedding an Anthropic API key. Works with Claude Pro/Max/Enterprise accounts that authenticate via OAuth and have no API key. Invokes claude -p with the existing Claude CLI login and returns Claude's final output.

Use this skill when the user wants OpenClaw to hand a coding or repository task to Claude Code and receive a result back. It is meant for local, trusted operator setups where claude is already installed and signed in.

Preconditions

• claude (Claude Code CLI) is installed on PATH.
• Claude CLI has already been authenticated by the operator with claude auth login (browser OAuth — works for Pro, Max, and Enterprise accounts). No API key is required or used. Enterprise accounts sign in with their org credentials/SSO; some orgs route through a corporate gateway (claude gateway --config), which the CLI handles automatically once configured.
• Do not ask the user for an API key.
• Run inside the target repository or pass --cwd.
• OpenClaw must allow this agent to use exec for the wrapper script.

If Claude is not authenticated, tell the operator to run claude auth status then claude auth login once and complete the browser sign-in. Do not request, print, copy, or inspect ~/.claude/.credentials.json.

Wrapper

Prefer the bundled wrapper:

bash

{baseDir}/scripts/claude-delegate.sh --cwd /path/to/repo --prompt "summarize the repository and list the riskiest files"

For file creation or edits (auto-accept workspace edits):

bash

{baseDir}/scripts/claude-delegate.sh \
  --cwd /path/to/repo \
  --sandbox workspace-write \
  --prompt "Create docs/architecture.md explaining the current service layout."

For machine-readable event logs while still returning the final Claude answer:

bash

{baseDir}/scripts/claude-delegate.sh \
  --cwd /path/to/repo \
  --sandbox workspace-write \
  --json-log .openclaw/claude-runs/last.jsonl \
  --output .openclaw/claude-runs/last.md \
  --prompt "Review the auth changes and fix any failing tests."

To pass CLI-style context:

bash

npm test 2>&1 | {baseDir}/scripts/claude-delegate.sh \
  --cwd /path/to/repo \
  --sandbox workspace-write \
  --prompt "Summarize the failing tests and make the smallest safe fix."

Use --prompt-file path when a generated prompt is already on disk. Use --stdin-file path when another tool wrote context to a file. Use --allowed-tools "Read,Edit,Bash(git *)" to restrict which tools Claude may call, and --model sonnet (or opus) to pick a model.

Delegation Workflow

1. Restate the exact Claude task in one compact prompt.
2. Choose the narrowest permission level:
   • read-only (maps to Claude plan mode) for analysis, review, summaries, planning.
   • workspace-write (maps to acceptEdits) for creating or editing files in the repo.
   • danger-full-access (maps to bypassPermissions) only on an isolated machine/container with explicit operator approval.
3. Run the wrapper from the repository root or with --cwd.
4. Return Claude's final answer to the user.
5. If edits were allowed, inspect git diff --stat and relevant diffs before claiming files were changed.

Safety Rules

• Never pass ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN, or raw tokens to Claude. The wrapper unsets API-key env vars before invoking claude, forcing the saved OAuth login.
• Never read, reveal, summarize, or copy ~/.claude/.credentials.json.
• Do not run this skill from public or untrusted chat channels.
• Do not give Claude secrets, credentials, private keys, or production data unless the operator explicitly approved that exact data flow.
• Prefer an isolated checkout or git worktree for write tasks.
• Treat Claude output as another agent's report; verify important claims locally.

Failure Handling

• claude: command not found: ask the operator to install Claude Code CLI (claude install or npm).
• Authentication failure (exit 4/33/34): ask the operator to run claude auth status / claude auth login with their Pro/Max/Enterprise sign-in, then retry. No API key needed.
• Permission errors (exit 5/11): retry with --sandbox workspace-write only if file changes are intended, or narrow --allowed-tools.
• Non-zero exit codes other than above: report the code; Claude Code uses distinct codes for network errors (3), config errors (6), budget/turn limits (16/25), and API/server overload (17/27/28).