adlc-workstream

name: adlc-workstream description: Plan grounded long-running workstreams with Codex goal-managed execution.

ADLC Workstream

Use this when an epic is too large for one plan execution pass and needs durable Codex goal context, milestone sequencing, step cards, Kanban state, and execution handoffs.

Process

1. Resolve effective ADLC paths from .adlc/config.yaml; use adlc resolve-config when the CLI is available.
2. Ground the epic in live repo evidence, relevant interview specs, existing plans, architecture, rules, and user constraints before creating steps.
3. For long-running or multi-sprint execution, create or attach the active Codex goal and record the goal objective in WORKSTREAM.md.
4. Inspect the dirty tree before delegation. Commit small coherent existing work first; stop and ask for large, mixed, or unrelated dirty trees.
5. Create or update paths.workstreams/<slug>/ with a workstream overview, Kanban board, evidence log, decision log, milestone cards, step cards, and Codex handoff.
6. Decompose the epic as workstream -> milestone -> step -> task. Each step should be one reviewable, commit-capable slice with a bounded write scope.
7. Move each step through ready -> build -> review -> fix -> test -> commit -> done, with review/fix/test loops handled autonomously unless a human-gated blocker is explicit.
8. Assign each step an execution lane: coordinator, worker, parallel, or human-gated.
9. For Codex handoff, include the exact ADLC workstream/step path, bounded write scope, verification commands, expected gate outputs, and goal-continuity instructions.
10. Keep blocked decisions explicit and avoid inventing dependencies that are not supported by evidence.

Rules

• Workstreams coordinate execution; they do not replace adlc-plan, adlc-implement, adlc-verify, adlc-review, or adlc-commit.
• Codex goals manage the active long-running objective. ADLC artifacts remain the detailed source of truth for scope, state, evidence, and decisions.
• Interview specs remain binding clarification artifacts. Carry their non-goals, decision boundaries, pressure-pass findings, and acceptance criteria into milestones and step cards.
• Every milestone must have a concrete outcome, dependency notes, exit criteria, and release or rollback considerations when relevant.
• Every step must have evidence, scope, acceptance criteria, verification, commit checkpoint, and next-stage rules.
• Split any step whose acceptance criteria span unrelated subsystems, cannot be verified in one pass, or would create a mixed commit.
• Review, fix, and test gates are autonomous by default. A failed gate with actionable code, docs, tests, rules, or security findings should route to the next ADLC gate or fixer, then return to review/verify without operator approval.
• Use human-blocked only for explicit human decisions, credentials or external accounts, destructive or production operations, legal/security sign-off, scope or product ambiguity, or a user-requested approval point.
• Do not encode an ordinary post-fix review-required handoff as a human blocker. Create or advance an autonomous review/verify gate, or mark the fix complete with the downstream gate pending.
• Do not create hidden automation. ADLC exports work; execution lanes perform it through their normal Codex lifecycle.
• Do not stop global worker services while stopping a task. Gateway, daemon, scheduler, or service shutdown requires an explicit operator request.
• Long-running state must live in ADLC artifacts and Codex goal status, not chat memory.

Output

Report the workstream root, Codex goal status, milestone count, step count, execution lanes, blocked decisions, current Kanban state, and the next runnable Codex step.