Delegate bounded coding tasks to isolated Claude sessions

Lets an orchestrating agent hand one bounded coding task to a separate Claude Code CLI session - an implementer with no orchestrator chat history - then review what it produced and land it itself. Dispatch runs through the bundled relay, which wraps Claude's non-interactive claude -p --output-format stream-json --verbose mode, sends the brief on stdin (never argv, and capped at 10 MB by Claude Code itself), and writes a structured result.json under the system temp directory by default. The relay never adds --bg (Claude documents background mode as incompatible with -p) and never adds --bare (bare mode would skip CLAUDE.md and OAuth/keychain authentication) - so the child session still discovers the target project's CLAUDE.md and local Claude settings normally, but does not generically auto-load AGENTS.md; the orchestrator must read that itself and copy load-bearing constraints and real gate commands into the brief.

The normal permission profile is deliberately narrow: acceptEdits mode, a built-in tool surface restricted to Read, Glob, Grep, Edit, Write, and the platform shell; on macOS, Linux, and WSL2, Claude's own shell sandbox is enabled with startup failure (not a silent unsandboxed retry) when the sandbox is unavailable; MCP discovery and Claude.ai connectors are disabled, all MCP tools are denied, and skills, commands, and Claude's own Agent tool are unavailable to the child; and string rules deny common direct shell forms of git commit, git push, and nested claude invocations, though these are only a speed bump since aliases and wrappers can bypass them - the brief's no-commit instruction plus orchestrator review remain the real boundary. --read-only switches to plan mode (Read/Glob/Grep only) and adds a Git-visible-change tripwire (readOnlyViolation: true/false/null) by fingerprinting the working tree before and after. --dangerously-skip-permissions opts into Claude's bypassPermissions mode and requires the human's explicit acceptance. Rework resumes the same session with a delta brief via --resume-last (maps to Claude's --continue) or --session <id> (maps to --resume <id>), re-passing the same permission profile.

Use it to delegate a bounded coding task to a separate claude implementer and then review its diff yourself, only after the user explicitly asked for delegation to another Claude Code process or session - this skill is not a signal for the current Claude to implement directly. Do NOT use it when: the human asked the current agent to implement the task directly; the task is small enough to do inline and delegation was not requested; the claude CLI is missing or unauthenticated (checked with claude auth status, though a sandboxed orchestrator can get a false loggedIn: false if it blocks Keychain access - re-check outside that sandbox before concluding otherwise); or the task needs a stronger host boundary than Claude Code's tool permissions and shell-only sandbox provide, in which case use an isolated container or VM instead.

Input is a bounded, block-structured brief (<task>, <repo_constraints> copied from AGENTS.md, <verification_loop> naming the real gate commands, <action_safety> forbidding commits, <structured_output_contract>), delivered on stdin. Dispatch:

node "<skill-dir>/scripts/relay.mjs" --brief brief.txt --cd /path/to/repo

Output is result.json (schema delegate-relay.result.v1) with status (completed/failed/timeout/aborted/claude_unavailable), finalMessage, touchedFiles (final git status --porcelain, not attribution), permissionMode, sessionId, resultSubtype, numTurns/usage/totalCostUsd, and readOnlyViolation on read-only runs - plus raw artifacts brief.txt, events.jsonl, final.txt, stderr.txt, and profile.json. The implementer never commits; the orchestrator does, after re-running the project's real gates and reading the complete diff.

Wraps the claude CLI (Claude Code), requiring it installed and authenticated, Node 18+, and git; Claude's shell sandbox requires macOS, Linux, or WSL2 (native Windows pre-approves PowerShell instead, without OS-level shell isolation, and that launch path is implemented but not yet verified). Supports --model, --effort (low/medium/high/xhigh/max/ultracode), --max-turns, and --max-budget-usd caps on the delegated session, and can apply a fleet lane's dials from a companion delegate-setup configuration via --lane.

Orchestrating agents - including another instance of Claude Code itself - that need to delegate a scoped coding task to a separate, sandboxed Claude Code session while keeping review-and-commit authority, and that need a cross-orchestrator contract complementary to native Claude subagents or background sessions. Adapted from the community amElnagdy/delegate-skills repository under the MIT license; this catalog entry is a docs-only import, since the executable scripts/relay.mjs is not bundled here.