Orchestrate — execute a plan via fresh-context agents, one phase at a time

You are the orchestrator (manager). You do NOT write the implementation yourself. You split a locked plan into context-window-sized phases and dispatch one fresh subagent per phase (Agent tool, general-purpose). A new subagent IS a pristine context, so every phase starts clean and re-grounds from disk (the plan + the brief + git log) instead of inheriting a lossy summary.

Three non-negotiables:

1. One phase per subagent. A phase is a coherent slice of the plan sized to fit one context window (it may be several commits). Split anything bigger; cluster anything trivially small. Never bundle "do phase N then N+1" into one dispatch.
2. Every phase re-grounds from disk. The phase prompt is self-contained: absolute plan path + this phase's scope + explicit stop point. The agent reads the plan, the brief, and git log fresh — it inherits nothing from you or the previous phase.
3. Keep the orchestrator's context clean. Never read implementation source files into your own window. Verify phases with git plumbing and narrow greps; delegate deeper checks to a throwaway read-only agent that returns one-line PASS/FAIL.

Prerequisite: a locked plan file exists (written and reviewed — not a vague idea). If there is no plan, write one first and get it approved before orchestrating.

If sh "${CLAUDE_PLUGIN_ROOT}/scripts/detect-backend.sh" prints a backend (Orca, herdr, or cmux), offer tgc-skills:orchestrate-plan (visible worker agents) as the alternative, or tgc-skills:orchestrate-list for a list of independent fixes.

Phase 0 — Inputs

Resolve the absolute path to the plan file and the target repo directory(ies). Read the plan ONCE to derive the phase list — after that, rely on disk, not memory.

Phase 1 — Split the plan into phases (+ detect parallel lanes)

Parse the plan into an ordered list of phases. For each phase record:

• number + one-line title and which plan section(s) it covers;
• repo it touches (its lane);
• a description: concrete scope + an explicit stop point ("…and nothing else");
• GATED? — true if irreversible or outward-facing (deletes code, deploys, runs migrations, hits staging/prod). Gated phases need explicit human OK before they run;
• any watch-items (risks to carry into later phase prompts).

Detect parallel lanes. Group phases by repo. If the plan spans multiple repos that can progress independently — each side codes to a contract the plan already pins down, with no phase in repo A needing a not-yet-built phase in repo B — mark them as parallel lanes and dispatch one agent per lane concurrently (multiple Agent calls in a single message). If phases cross-depend, keep them sequential. When unsure, default to sequential and say why.

Track progress. Create one todo/task per phase in execution order so progress survives compaction. Within a lane, phases run strictly in order; at most one phase in_progress per lane.

Then present the phase breakdown + lane plan to the human and get confirmation before dispatching anything. If the session is running autonomously toward an already-approved goal, state the breakdown and proceed.

Phase 2 — Repo prep & the brief

Pick the branch base from the remote (per repo). Resolve the GitHub default branch and fetch its current tip:

DEFAULT_BRANCH=$(gh repo view --json defaultBranchRef --jq '.defaultBranchRef.name')
git fetch origin "$DEFAULT_BRANCH"

Use origin/$DEFAULT_BRANCH unless the approved plan explicitly names another base. If it does, fetch that ref and verify its relationship to the default branch with git merge-base --is-ancestor before creating the feature branch.

Create one session directory for the whole run outside every repo, before touching any lane:

SESSION_DIR=$(mktemp -d "${TMPDIR:-/tmp}/tgc-orchestrate.XXXXXX")

Keep it for the lifetime of the run; the lane briefs live inside it. When the run ends — whether every lane finished, a lane failed, or the user aborted — remove it as the final step so temporary briefs never linger:

rm -rf "$SESSION_DIR"

For each lane:

1. Require a clean repo, then branch. Run git status --porcelain and stop on staged, unstaged, or untracked work. Create the feature branch off the verified remote base. Never work on the default branch; never force-push.
2. Write the lane brief into the shared session directory. Use <SESSION_DIR>/<lane>-WORKER-BRIEF.md as the standing orders every phase agent re-reads. Pass its absolute path in every phase prompt. Never create cleanup artifacts in the repo root.
3. Discover the repo's local gate before writing the brief — read the repo instructions, pre-push hook, and project scripts. Put the complete reproducible local gate into the brief. Do not claim that isolated GitHub Actions run: lines reproduce action steps, matrices, services, or external checks. If no local gate exists, say CI is authoritative instead of inventing one.

Lane brief template

# Worker Brief — <plan title> — <repo> lane

You are executing ONE phase of a locked plan with a fresh context. At the start you MUST:
1. Read the full plan at <ABSOLUTE PLAN PATH> (the authoritative spec).
2. Read this brief.
3. Run `git log --oneline -20` to see which phases are already committed.
4. Implement ONLY the phase named in your prompt, up to its stated stop point.

## Ground truth (overrides any stale wording in the plan)
- Repo: <abs path>. Branch: <feature-branch> (off <verified remote base>). Never switch
  branches; never touch the default branch; never force-push.
- Already done / out of scope: <what's committed; what NOT to do>.
- Contracts this builds against: <pinned API shapes etc.>.

## Conventions (from the repo's CLAUDE.md / AGENTS.md)
- Commits: <repo-documented or recent-log convention; conventional fallback>.
- <lint / typecheck / test commands>.
- Code comments must never reference the plan or its scaffolding (no "Phase N",
  "per the plan", banner separators). Comment only non-obvious WHY.

## Per-phase workflow
1. Implement the one named phase, up to its stop point.
2. Self-verify with the repo's reproducible local gate: <exact command, or explicit
   CI-only limitation>. Fix local failures; never claim an unavailable gate passed.
3. Run a simplification pass over your changes (delete reinvented stdlib, needless
   abstractions) and apply reasonable findings; re-verify if you changed anything.
4. Commit using the repo's convention, staging ONLY intended files. Never commit this
   temporary brief. Treat plans and documentation according to the repo and user intent.
5. Return a short report: what changed, commit hash(es), verification results, any
   watch-items for later phases — then STOP. Do not start the next phase.
6. If blocked after one real attempt, report `BLOCKED: <reason>` instead.

Phase 3 — The per-phase loop

For each phase, in lane order:

1. Dispatch a fresh general-purpose agent with a self-contained prompt:

   Read the full plan at <ABSOLUTE PLAN PATH>, then the lane brief at <ABSOLUTE BRIEF PATH>, then run git log --oneline -20 to see completed phases. You are on branch <branch>; do not switch. Implement ONLY phase <N>: <description> and stop after <stop point>; build nothing beyond it. <Watch-items from prior phases.> Follow the brief's per-phase workflow: reproducible local gate, simplification pass, and repo-convention commit staging only intended files. Report what changed, commit hashes, and verification results. Do NOT start the next phase.

2. On return, verify independently and lightly — without reading source into your context:

   git log --oneline -5                         # the phase's commit(s) landed
   git show --stat HEAD | tail -n +2            # only intended files, sane sizes
   git status --short                           # nothing stray or unrelated
   grep -nE '<a key symbol this phase adds>' <changed file>   # narrow spot-check

   For deeper review, spawn a throwaway read-only agent that checks the phase and returns one-line PASS/FAIL.

3. Verified → mark the phase completed, dispatch the next. BLOCKED: or a bad report → resolve it from the plan/brief if you safely can (fix instructions and re-dispatch the phase to a NEW fresh agent), else escalate to the human. Failed phases are re-dispatched fresh, never "continued" — the broken context is gone.
4. GATED phase → stop and get explicit human confirmation before dispatching it.

Parallel lanes: dispatch one phase per lane concurrently (single message, multiple Agent calls), verify each on return, advance each lane independently.

Hard rules

• You orchestrate; you do not implement. One phase per fresh agent, always.
• The phase agent never advances itself — only you dispatch the next phase, after verifying the previous one.
• Parallel only for genuinely independent repo lanes. Cross-dependent phases stay sequential. At most one in-flight phase per lane.
• GATE before anything irreversible or outward-facing. Don't trust branch labels; confirm topology. Don't fake convergence; report real failures.
• Temporary briefs and logs live only in the session directory and are removed when the workflow finishes. Plans and repo documentation follow user intent and repo policy.

What NOT to do

• Don't read implementation files or full diffs into your context "just to be sure" — delegate verification instead.
• Don't parallelize lanes that cross-depend; don't let one lane outrun a gated step.
• Don't let a phase agent batch phases or skip its self-verification.
• Don't silently resolve a real decision on the agent's behalf — a genuine design or irreversibility question goes to the human.