subagent-orchestration

name	subagent-orchestration
description	Use when executing FSD goals with independent work packages. Dispatches a fresh subagent per goal with file-backed handoff and bounded review.

Subagent Orchestration

Dispatch one fresh agent per independent FSD goal. Keep briefs, reports, diffs, and review evidence on disk; messages carry paths and verdicts.

Announce: "I'm using subagent-orchestration for file-backed goal dispatch."

Preconditions

• Goal has approved FSD authority, exact acceptance/test refs, and no unresolved OPEN-* blocker.
• For UI scope, the pointer includes ui_delivery_role, required_gate, and qualified pinned contract refs. CONTRACT_ENABLER is the only bounded DRAFT/BLOCKED exception; FIRST_VERTICAL_SLICE requires READY_FOR_SLICE; SCALE_OUT_SLICE requires a VALIDATED baseline and the pinned first-slice issue at FIRST_VERTICAL_SLICE_VERIFIED.
• HARDENING requires every applicable UI delivery slice dependency to be verified, final verification refs, and a named Business Owner for UAT.
• Parallel goals do not share unmerged files or mutable validation resources.
• Search existing code/tests before assuming anything is absent.

File-Backed Process

1. The scheduler writes a JSON array of allowed repository-relative target paths, then creates one package from the lightweight issue pointer and that scheduler-owned scope:

   node .agent/tools/work-package.mjs create \
     --run <run-id> --goal <goal-id> --brief <issue-path> \
     --paths-file <scheduler-scope.json>
   

2. Send the implementer only the returned briefPath, reportPath, read-only pathsPath, exact target paths, and this constraint: implement one goal, use TDD when behavior changes, never edit the scheduler-owned scope, keep full evidence in reportPath, and return at most 15 lines containing outcome, paths, verification status, and blockers.

3. Run parallel goals only in isolated worktrees/workspaces. Review rejects a changed scope digest and any new working-tree edit outside the allowlist.

4. After implementation, freeze one working-tree review package:

   node .agent/tools/work-package.mjs review \
     --run <run-id> --goal <goal-id> --base <review-base>
   

   The scheduler-owned allowlist is mandatory. A parallel goal without an isolated workspace must fall back to sequential execution.

5. One fresh reviewer reads the brief, report, and patch once, then writes two separate verdicts: SPEC and QUALITY. Use references/review-contract.md for the detailed checklist.

6. Batch critical/important fixes into one correction wave. Rebuild the patch and re-review once. After two failed revision cycles, escalate.

7. Record the result:

   A FIRST_VERTICAL_SLICE must use a real provider or backend to prove auth/permission, success, and representative failure, with integration-checking evidence. Only then may its status be recorded verified; mock-only evidence is insufficient. Recheck ui_delivery_role, required_gate, and pinned contract revision immediately before recording any UI result.

   node .agent/tools/work-package.mjs record \
     --run <run-id> --goal <goal-id> --status verified \
     --verification "<short command result>"
   

Invariants

• Never paste whole BRD/PRD/FSD/ADR or diff bodies into dispatch messages.
• Never invent schema, API, authorization, workflow, role, state, or UI behavior.
• Shared builds/tests run serially; safe search and isolated edits may fan out.
• Commits remain routed through /sc-go and only when requested.
• A spec failure and a quality failure remain distinct even though one reviewer reads the package once.

Related Skills

Use context-engineering, executing-plans, test-driven-development, code-review, and verification-before-completion as their active branches require.