| name | repo-orchestrate |
| description | Hand the entire repo-build-pr workflow to a harness of your choice (Claude Code headless or Codex): the chosen harness acts as orchestrator and validator — it creates its own worktree, implements (sub-delegating via repo-delegate where useful), runs the gate, runs the repo-review battery with cross-harness adversarial dispatch, and either stops ready-to-publish (default) or pushes and opens the draft PR (--publish). Use when the user wants to pick which harness orchestrates a build end to end, delegate a whole task AFK, or run the same build prompt under a different orchestrator for comparison. |
Repo orchestrate
Full-workflow delegation. repo-delegate hands off one scoped brief;
this hands off the whole loop — the chosen harness is the orchestrator
and validator, following repo-build-pr end to end. Same structure as the
sibling skills: one prompt template (PROMPT.md), one runner per
harness, a template for adding more.
Trust model — read before dispatching
The orchestrator commits on a branch it creates, runs your toolchain, and in
publish mode pushes with your credentials and opens a PR. Two levels:
- Default (no
--publish): build + validate + review, stop before any
push. Codex runs in a workspace-write sandbox (network enabled for
pnpm install --frozen-lockfile); Claude runs acceptEdits with a
git/pnpm/cargo allowlist.
The caller inspects the worktree and publishes.
--publish: the orchestrator may git push and gh pr create --draft
(never mark ready, never merge). On Codex this requires
danger-full-access (the workspace sandbox cannot reach your ssh agent);
on Claude the allowlist gains git push and gh pr. Only use on tasks
and prompts you wrote yourself.
Either way the orchestrator's report is a claim; the worktree diff, gate
output, and battery verdicts are the evidence.
How
-
Write the build prompt to a file — exactly what you would type after
/repo-build-pr: objective, constraints, acceptance criteria, scope
boundaries. Headless orchestrators cannot ask clarifying questions, so
answer the obvious ones in the prompt; anything left open is taken
conservatively and recorded as an assumption in the report/PR body.
-
Pick the orchestrator harness and dispatch (from the repo root of the
main checkout):
scripts/run-codex.sh -t <build-prompt.md> [-C <repo-root>] [-b <base>] [--publish] [-o <out>]
scripts/run-claude.sh -t <build-prompt.md> [same flags]
--dry-run prints the filled orchestrator prompt. Runs take long —
dispatch in the background and read the -o report file.
-
Verify: read the report, then the worktree diff and battery verdicts
yourself. Default mode: publish with your own hands. Publish mode: check
the draft PR before requesting reviews.
Gotchas
- Worktrees go under
<repo>/.worktrees/<branch> — inside the
workspace so the Codex sandbox can write them (sibling ../ paths are
outside the sandbox), per repo-build-pr's worktree strategy.
- The orchestrator must copy
.env + clovy-api/.env from the main checkout
into its worktree (repo-build-pr worktree strategy) — gitignored, so the
sandbox can read them; keep secrets you can't lose out of delegated repos.
- Cross-harness review still applies one level down: the orchestrator
dispatches the adversarial axis to a harness other than itself via
repo-review/scripts/run-*.sh.
- Fresh worktrees need
pnpm install --frozen-lockfile; the prompt says so,
but budget for it in runtime expectations.
Extending
Add scripts/run-<harness>.sh per
scripts/HARNESS-TEMPLATE.md: same CLI, prompt
from fill-prompt.sh, the two-level trust model preserved (sandboxed
default, explicit escalation for publish), report to a file.
