| name | delegate-claude |
| description | Launch and follow one durable Claude Code session that independently leads and implements a substantial mission. Use when the user asks Codex to have Claude execute or hand off the work, or when an ambitious, exploratory, multi-file task would benefit from a long-horizon collaborator. Do not use for an obvious bounded fix or to produce a copy-paste prompt; use handoff instead. |
Delegate Claude
This is the inverse-direction workflow: Codex starts a Claude session for a
specific task. It does not define the standing Claude Code relationship, where
Claude leads the mission and may call Codex through /codex:rescue.
Invite Claude to lead a substantial piece of work as a collaborator. Tell it
what you want to become true, point it at the context worth seeing, and name
only the boundaries that are real. Do not send a governance document, a
pre-chosen plan, or a menu of decisions for Codex to make.
The invitation should feel like a useful message to a trusted maintainer. Add
the current branch, worktree state, or proof target only when it helps Claude
act well. Claude makes the in-scope calls, including product, architecture, and
local implementation choices. It may reframe the work, discard an inherited
approach, and explain the choice in its handoff. Do not interrupt it to decide
ordinary alternatives.
Use consult-claude for a one-shot read-only investigation. Use this skill
when Claude should investigate, decide, and carry the work through. Use
handoff only for a manual copy-paste prompt.
Boundaries that remain real
Local work and local commits are in scope. Destructive actions, external
writes, unrelated private-data access, and material expansion beyond the user's
request are not. Pushing, opening or merging a pull request, deploying, and
other external writes require separate user authorization after verification.
The launcher denies common direct forms of git push, gh pr create, and
gh pr merge on every launch and resume. It is not a shell or network sandbox:
deploys, APIs, and other external-write routes still rely on Claude following
the invitation and on later verification.
Run and watch
Resolve this skill's directory from its loaded SKILL.md path, then start one
durable session in the task's working directory:
bun <skill-dir>/scripts/delegate-claude.ts start --name <short-name>
Write the invitation to stdin and close it (Ctrl-D on a PTY). The launcher
uses high effort, normal project configuration, auto permission mode, and a
separate worktree. Capture DELEGATE_CLAUDE_JOB_ID=<id>; if it is absent after
a successful launch, inspect claude agents for the chosen name before trying
again.
Watch the session with:
bun <skill-dir>/scripts/delegate-claude.ts watch <id>
Read logs on state changes, blocks, long silence, or completion. A block about
an in-scope choice is a cue to tell Claude to make its best call, not a cue to
make the choice for it. Continue only for new user direction or newly granted
authority:
bun <skill-dir>/scripts/delegate-claude.ts continue <id>
continue refuses to interrupt a working turn unless --interrupt is
explicit. Cancel with claude stop <id> when the work is no longer relevant;
use claude respawn <id> only after a confirmed process failure.
When Claude finishes, this Codex-initiated workflow may need to inspect its
worktree and complete diff before the calling task uses or reconciles the
result. Choose the verification appropriate to that task and check for
unauthorized external actions. This does not override Claude's normal ownership
of a Claude Code mission. Treat the handoff as evidence, not proof. Publishing
remains a separate user-authorized action.
