cross-harness-sub-agent

spawn a cross harness sub-agent

SKILL.md

GitHub

route

  • from codex, follow claude-subagent.md
  • from claude code, follow codex-subagent.md
  • from another harness, try claude code first, fallback to codex if it failed
  • the calling agent invokes the selected route directly
  • an explicit user target, model or effort wins
  • never use this for same harness sub-agents

checkout

<checkout> is an absolute path you reuse literally in every later command:

  • current checkout: the output of git rev-parse --show-toplevel
  • separate worktree: the last output line of bin/wt-run HEAD -- pwd

max 1 writer per checkout

task

give the sub-agent a self-contained prompt with the objective, scope, completion condition, edit permission and any context it can't discover

require the sub-agent to read ~/.agents/AGENTS.md when present and applicable repository/project AGENTS.md before starting

unless the user gives another review scope, compare the current branch with origin/develop and include committed branch changes plus staged, unstaged and untracked changes

for an exact-SHA read-only review, stop once the evidence supports a verdict; require the final report to include the reviewed SHA, compact read-only verification evidence, and either concrete file:line findings with the smallest remedy or a verdict line containing exactly NO_ISSUES_FOUND

for read-only reviews, require the report as the final response; the launcher persists it, so the reviewer needs no file writes

require no further cross-harness delegation, and a token-efficient verification command for you to run

for a direct cli route, pick a short unique <slug> and write the exact task with the file-writing tool to <checkout>/playground/cross-harness/<slug>-prompt.md (<prompt-file>). create its parent directory in a separate call. retain <checkout>, <slug> and <prompt-file> until completion; files under playground/ stay out of git

misc

read troubleshooting.md if something goes wrong