next

Read CLAUDE.md and stories/_queue.json.

Autonomous loop

Repeat one iteration until a stop condition fires:

1. Reload state

Re-read stories/_queue.json (and CLAUDE.md if needed) so each iteration uses fresh statuses.

2. Hard stop: BLOCKED

If any story is BLOCKED -> tell the human, summarize queue, exit the loop (do not dispatch).

3. Choose the next batch (priority order)

Apply dispatch rules 3–10 from CLAUDE.md, in order: FAILING_VISUAL_QA -> FAILING_QA -> READY_FOR_VISUAL_QA -> READY_FOR_QA -> READY_FOR_BUILD -> READY_FOR_UX_DESIGN -> READY_FOR_LLD -> READY_FOR_ARCH.

Only consider stories whose deps are all DONE. Never dispatch a story with unmet deps. Use the first rule that matches at least one eligible story; collect all stories matching that rule (with met deps) as one batch. Route by track per CLAUDE.md's track routing table (e.g. READY_FOR_BUILD dispatches developer for track=backend and frontend-developer for track=ui/fullstack).

If no story matches rules 3–10:

• If every story is DONE -> short success summary, exit.
• If prd/current.md exists and the queue has no stories (empty array) -> dispatch project-manager once (same prompt as /ship-prd), parse <status>, update _queue.json, then continue the loop.
• Otherwise -> report queue state (e.g. DRAFT, waiting on deps, or stuck), exit — do not spin.

4. Parallel dispatch

If the batch has multiple slugs mapping to the same subagent role (after track routing), dispatch them in parallel with concurrent Task tool calls (cap 7 per iteration). If the batch is a single story, one Task is enough.

Maker ≠ checker: qa-engineer and visual-qa must always be fresh Task dispatches — never grade in the implementer's context.

Include in every subagent prompt:

• slug, track, absolute paths to inputs, expected outputs, target next_status on success
• prior QA/visual-QA report path when re-dispatching on FAILING_QA/FAILING_VISUAL_QA
• Identity reinjection: one line with the role's role_id, identity_marker, and non_negotiables from that agent's Identity block
• Effort / model: set Task model from CLAUDE.md effort routing (xhigh for project-manager / solution-architect / tech-lead / qa-engineer / visual-qa; high for ux-designer / developer / frontend-developer)

5. Apply results

When the batch finishes, parse every returned <status> block; for each story append a history entry and set status to next_status.

6. Continue

Immediately start the next iteration at step 1. Do not stop after one batch to wait for another /next unless a stop condition fired.

Stop conditions (summary)

• Any BLOCKED
• All stories DONE
• No eligible batch and not the "empty queue + PRD present" special case
• No-progress guard: if an iteration performed no successful dispatch and the queue is unchanged, exit and explain — avoids infinite loops

Keep a short running log of batches dispatched so the final user summary lists what ran and ends with the queue snapshot.