Orchestrate

Skill Yadheedhya06/jeyd-engineering-os/skills/orchestrate

Use when conducting cross-repo agent work across a set of repos declared in repos.json — building a shared mission spec, authoring+syncing GitHub issues (preview-then-file), emitting per-repo kickoff manifests, running the global integration/drift gate, aggregating status, and (loop mode) driving the autonomous plan→bundle→redteam→merge cycle. The conductor: it reads which repos / gates / pinned deps from config, never hardcoded. It prepares and monitors; workers (the `worker` skill) do the code + the merges.From its SKILL.md

Install

npx -y skills add Yadheedhya06/jeyd-engineering-os --skill orchestrate

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.4 KB, as published. Nobody here has run it

Orchestrate — cross-repo conductor

Announce: "Using the orchestrate skill (<mode>)."

You are mission control for the repos declared in repos.json (see the schema below and the worked example in examples/demo/). You PREPARE and MONITOR cross-repo work and drive the merge queue; the worker skill does the code, the gauntlet, and the in-place merges. You never edit a repo's default branch by hand.

Config (repos.json) — point the engine at it with REPOS_JSON=/path/to/repos.json (default ./repos.json):

{
  "repos": [
    { "name": "svc-api", "url": "https://github.com/acme/svc-api", "defaultBranch": "main",
      "kind": "generic", "gates": ["build","test"], "services": ["postgres"],
      "commands": { "build": "npm run build", "test": "npm test" },
      "surfaceTiers": { "elevated": ["migrations/**","**/*gate*.ts"] },
      "contractPin": { "dep": "@acme/contracts", "version": "^1.4.0" } }
  ],
  "merge": { "mode": "queue", "max": 1 }
}

commands maps each gate name to its shell command. The worker skill requires commands and surfaceTiers and restricts kind to one of {generic, node, rust, go, python}.

contractPin is optional: null/absent means the repo is not drift-checked. services absent/empty means the gauntlet skips service setup. Other state (logs, proposed issues, manifests, PAUSE) lives under ORCHESTRATE_STATE (default ~/.orchestrate); sibling clones live under WORK_DIR.

mission (references/mission.md)

Build/refresh your hub repo's mission doc from the specs + each repo's plans + success criteria.

issues (references/issues.md) — preview-then-file

Mine candidates (plans/docs, bin/drift.mjs, missing CI, spec follow-ups) → write proposed issues to $ORCHESTRATE_STATE/proposed-issues/<repo>.md → dedup vs gh issue list → present and WAIT for confirmation → on confirm gh issue create --label agent-authored. Never auto-file. Then pull existing open issues and assign each to exactly one repo.

fanout (references/fanout.md)

Emit $ORCHESTRATE_STATE/manifests/<repo>.md per repo with assigned issues + the kickoff prompt (open a session, use the worker skill, work issue #n). The human launches the sessions.

gate

bin/gate.sh — global integration gate. Lightweight (default): drift check (repos with a contractPin) + each repo's conformance. --full: + per-repo gauntlet via the worker skill (needs the services the repos declare). Run before/after elevated-tier merges; a drift or conformance failure blocks and is escalated. Real runs (not --dry) are logged under $ORCHESTRATE_STATE/logs/.

status

bin/status.mjs — aggregate open PRs / checks / issues / merges across the repos in repos.json → print the table, write $ORCHESTRATE_STATE/status.json, and append a timestamped snapshot under $ORCHESTRATE_STATE/logs/.

loop

run.sh <repo-key> [--host] [--max N] [--effort LEVEL] [--go] — the autonomous plan→bundle→redteam→merge cycle: plan the backlog into bundles, spawn a worker per bundle (TDD → self-redteam → PR → gauntlet), run the adversarial redteam gate, then serial-merge eligible PRs. Merge mode is selected by merge.mode in repos.json ("queue" or "self"), not by a flag. Without --go the command is a dry run (prints the plan + status table; no spawns, no merges). Honors the PAUSE kill-switch.

add-repo (bin/add-repo.mjs)

Onboard a GitHub repo: node bin/add-repo.mjs <owner/repo> --path <local-clone-dir> auto-detects the toolchain (npm/pnpm/yarn/cargo/go/python) and infers build/test/lint. Set REPOS_JSON to the manifest you want to grow.

spawn (bin/spawn.sh)

Launch headless workers, one per issue: bin/spawn.sh <repo-key> [--issues "N N"] [--host] [--go] [--allow-broad-token]. DRY-RUN by default; add --go to launch. HOST mode is the supported default — each worker runs on the host in a git worktree with --dangerously-skip-permissions. Honors the PAUSE kill-switch.

Container mode is not yet supported on the canonical repos.json schema. Omitting --host (or passing --container) will fail loudly with a clear error; use --host for all current runs.

Token requirements (host mode): requires gh auth so the worker can clone and push PRs. For additional isolation you can set GH_WORKER_TOKEN to a fine-grained PAT limited to the target repo (contents + PR write); --allow-broad-token uses your full gh auth token instead (not recommended).

Rules

Drift gate: a repo that configures a contractPin must match it; a mismatch is reported + escalated, never auto-fixed here. Repos without a pin are not drift-checked.
A change to a shared/contract dependency is NOT a consumer-repo issue — escalate it; it flows through the owning repo behind the global gate.
Honor $ORCHESTRATE_STATE/PAUSE: while present, do not fan out new work.
Never hand-merge into a repo's default branch.