Claude Delegate

Claude Delegate

You are the orchestrator. Delegate one bounded coding task to a separate implementer — a Claude Code CLI session — then review what it produced and land it yourself. You write the brief and own the judgment; the separate Claude session edits the working tree; you verify and commit.

This skill is not a signal for the current Claude to implement directly. Use it only after the human explicitly asks for delegation to another Claude Code process or session.

When not to use this

• The human asked the current agent to implement the task directly.
• The task is small enough to do inline and the human did not request delegation.
• The claude CLI is missing or unauthenticated (claude auth status).
• The task needs a stronger host boundary than Claude Code's tool permissions and shell-only sandbox provide. Use an isolated container or VM for that requirement.

Prerequisites

1. claude --version succeeds.
2. claude auth status reports an authenticated session. On macOS the live credentials sit in the login Keychain; when the orchestrator's own sandbox blocks Keychain access (Codex's sandbox does), claude falls back to a possibly stale credentials file and reports loggedIn: false even though the login is valid. Re-run the check — and the dispatch itself — with that sandbox escalated or outside it before concluding the CLI is unauthenticated.
3. The target repository is the directory passed with --cd.
4. On Linux/WSL2, Claude's sandbox dependencies are installed. The normal relay profile is configured to fail when the sandbox is unavailable instead of silently running shell commands unsandboxed. Existing merged settings can still affect the effective boundary.

The loop

1. Write the brief

The separate session has no orchestrator chat history. It receives the brief on stdin and can inspect the target working tree.

Claude Code automatically discovers the target project's CLAUDE.md and normal local Claude configuration because the relay does not use --bare. It does not generically auto-load AGENTS.md. Read AGENTS.md yourself and copy every load-bearing constraint and the real gate commands into the brief. Tell the implementer not to commit. Keep one task per brief.

Template and details: references/writing-the-brief.md.

2. Dispatch

bashCopy code
node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo
# review/diagnosis only:                 add --read-only
# continue the latest session:           add --resume-last
# continue the recorded session:         add --session <id>
# choose limits:                         add --max-turns 40 --max-budget-usd 10
# set the auto-compact window:           add --autocompact 400k
# hard relay deadline:                   add --timeout 2h
# inspect every option:                  node .../relay.mjs --help

<skill-dir> is this installed skill directory, the folder containing this SKILL.md.

The relay runs claude -p --output-format stream-json --verbose, sends the brief through stdin, and writes artifacts under the system temp directory by default. It never uses --bg or --bare, and it never commits. See references/dispatch-and-poll.md.

--autocompact <auto|tokens> passes Claude Code's auto-compact window setting on every new or resumed invocation, and needs Claude Code 2.1.221 or newer — older builds fail the dispatch with unknown option '--autocompact'. The installed CLI owns the accepted range (auto, or 100k–1M tokens); the relay records the requested value but does not claim that Claude applied or enforced it.

3. Wait

The relay blocks until Claude exits. Use the orchestrator's background-command facility, or run it in the foreground and wait. Completion means the process exited and result.json exists.

• A pre-run usage error exits 2 and writes no result.json.
• A missing claude exits 127 and writes status: "claude_unavailable".
• Timeout and caught relay signals terminate the whole implementer process tree and preserve an outcome artifact.

Read finalMessage, touchedFiles, resultSubtype, and the raw artifact paths from result.json.

4. Review

Treat the implementer's report and gate outcomes as claims:

• Review edits to existing tests before a green gate means anything.
• Re-run the project's actual gates yourself.
• Read the complete diff against the brief, starting with touchedFiles.
• Inspect untracked and staged content as well as the ordinary diff.
• Run relevant guard skills if installed.

Full checklist: references/review-and-land.md.

5. Land

The orchestrator commits only after the gates pass and the diff holds. For rework, resume the same Claude session with a delta brief:

bashCopy code
echo "Keep the implementation, replace the mocked DB test with the migrated fixture, and remove the
unused import." | node "<skill-dir>/scripts/relay.mjs" --session <id> --cd /path/to/repo

Review a resumed run exactly like the first run.

Permission profiles

The normal profile is deliberately explicit:

• acceptEdits permission mode.
• Built-in tools restricted to Read, Glob, Grep, Edit, Write, and the platform shell.
• On macOS, Linux, and WSL2, Claude's shell sandbox is enabled with startup failure on missing dependencies and no unsandboxed retry. Commands that stay sandboxed are auto-approved so ordinary gates can run headlessly. The sandbox governs shell processes and their children only; merged local or managed sandbox settings can add effective paths or exclusions.
• Configured MCP discovery and Claude.ai connectors are disabled, all MCP tools are denied, and skills, commands, and Claude's Agent tool are unavailable to the child. Project CLAUDE.md, hooks, normal authentication, session persistence, and other local settings still load.
• String rules deny common direct shell forms of git commit, git push, and nested claude, plus any command containing claude-delegate. Aliases, scripts, and wrappers can bypass them, so they are only a speed bump; the brief's no-commit instruction and orchestrator review remain the boundary.

Native Windows does not support Claude's shell sandbox. The relay restricts the tool surface and pre-approves PowerShell so the run remains non-interactive, but that shell is not OS-isolated. Native claude.exe and npm claude.cmd launch paths are implemented; Windows verification is pending.

--read-only uses plan mode with only Read, Glob, and Grep. It removes edit, write, and shell paths, then compares parsed git porcelain and fingerprints the working-tree identity and index entries of Git-visible paths that were already dirty. readOnlyViolation is true when either signal proves a change, false when coverage is complete and detects none, and null when coverage is incomplete. This is a reporting tripwire, not an OS boundary: ignored paths and perfect restores are outside it, local hooks can write, and concurrent changes cannot be attributed to Claude.

--dangerously-skip-permissions is an explicit opt-in to Claude's bypassPermissions mode. The restricted tool surface, direct commit/push deny rules, and supported-platform shell sandbox remain, but direct file tools can cross normal permission boundaries. Use it only with the human's explicit acceptance.

Complementary to native Claude features

Claude subagents, agent teams, and background sessions are useful when the current Claude environment is already the orchestrator and native coordination is the goal. This skill is complementary: it provides a cross-orchestrator contract — self-contained brief → dispatch → artifacts → review → land — and keeps the commit with the orchestrator.

References

• references/writing-the-brief.md — context, CLAUDE.md versus AGENTS.md, real gates, report contract, and delta briefs.
• references/dispatch-and-poll.md — flags, profiles, artifacts, result.json, polling, and failure recovery.
• references/review-and-land.md — generated-code review, the commit boundary, and session rework.
• references/multi-task-queues.md — sequential queues, progress tracking, constraint carry-forward, and final coherence.