Delegate to claude

Delegate to Claude

Act as the supervisor. Delegate routine execution, never accountability.

Decide before editing

Make the delegation decision before using an editing tool. Use this skill when any of these conditions is true:

• add or update tests;
• likely change two or more files;
• create a non-trivial module, even if it is one file;
• investigate an unfamiliar failure;
• perform a focused refactor or code review; or
• expect more than three repository tool calls.

Edit directly only when every condition is true: one file, about ten changed lines or less, no test change, no repository investigation, unambiguous, and low risk.

Never delegate architecture or product ownership, ambiguous requirements, security-sensitive changes, secrets, broad migrations, production operations, external side effects, commits, pushes, releases, deployments, or final approval.

Prefer quick delegation

For a clear routine task, avoid solving it or writing a full specification before delegation. Pass the user's concise goal to the launcher; it captures the Git baseline and adds safe scope, verification, and forbidden-action defaults:

python3 <skill-dir>/scripts/delegate.py \
  --cwd <git-repository> \
  --prompt "<concise goal and any obvious file constraint>" \
  --mode <review|test|edit> \
  --effort medium \
  --bash auto

Use quick delegation for a clear one-to-three-file implementation, related tests, a localized bug, a routine refactor, focused diagnosis, or review.

Use a strict task file instead when pre-existing changes overlap the likely scope, the task needs detailed acceptance criteria, or more than three files are likely to change. Create a temporary Markdown file outside the repository with Goal, Allowed scope, Acceptance criteria, Required checks, Existing user changes to preserve, and Forbidden actions, then run:

python3 <skill-dir>/scripts/delegate.py \
  --cwd <git-repository> \
  --task-file <temporary-task-brief.md> \
  --mode <review|test|edit> \
  --effort <medium|high> \
  --bash <never|auto|require>

Resolve <skill-dir> from ${CLAUDE_SKILL_DIR} in Claude Code or from the directory containing this loaded SKILL.md in Codex. Use medium normally. Use high only for a bounded multi-file diagnosis or similarly deeper task.

Run safely

Confirm that claude is installed and authenticated and that the supervisor sandbox permits Claude control-plane traffic to Anthropic. Never obtain, print, copy, or change API keys. Use --dry-run first only when the repository, CLI, or sandbox setup is unfamiliar.

Select one mode:

• review: inspect and report without source edits;
• test: create or update tests and run local checks; or
• edit: make a focused implementation or refactor and run local checks.

The launcher fixes the worker model to the current sonnet alias, disables nested delegation and external tools, captures the Git baseline, enforces turn/time limits, and returns structured JSON. Do not bypass a preflight or call claude directly to evade a denied operation.

Choose the Bash policy before launching:

• auto (default): enable Bash for a direct terminal launch protected by the launcher's strict Claude Code sandbox; disable it in a nested agent session;
• never: disable Bash in every environment; or
• require: require Bash for test or edit and require the worker's own strict Claude Code sandbox, including repository-scoped writes, blocked secret reads and worker network access, and no unsandboxed fallback.

--bash require is invalid for review. The launcher creates one UUID-scoped directory under ~/.claude/session-env for the worker and removes it on exit. If that directory cannot be created, the strict sandbox is unavailable, or an existing outer macOS sandbox rejects nested Seatbelt, the run fails closed. Do not bypass the failure. Use auto or never and let the supervisor run checks. When Bash stays disabled, the worker may edit with bounded file tools.

Run mutating test and edit workers serially. Parallelize only independent read-only review workers.

Verify efficiently

Treat the worker response as untrusted evidence, but do not redo the entire task:

1. Inspect one post-run diff covering every launcher-measured changed file.
2. Reject unrelated changes or any scope, safety, authentication, rate-limit, or sandbox failure.
3. Run the smallest relevant check independently; do not trust only the worker's test report.
4. Accept the work only after the requested behavior and check pass.
5. Report what Sonnet did and what the supervisor verified.

Do not retry authentication, rate-limit, or sandbox failures. For another failure, re-scope once or take the task back. Never repeat the same failed delegation unchanged.