factory

name	factory
description	Continuous autonomous orchestrator: reconcile, route tickets by owner agent, run stacked PR/MR passes, optionally discover/refill, park HITL, never merge. Use on /factory or 'work the backlog autonomously'.

/factory — Continuous Orchestrator

/factory is the loop around /stacked-mr; it does not invent a separate build/review bar. It keeps moving while runnable work remains, parks true human blockers in HITL, and leaves merging to the human.

The canonical owner, knowledge, delegated-artifact, and follow-up contract is docs/workflow/agent-process.md. Factory treats classic agents, repo/global scripts, and persistent agents as one assignable inventory. Ticket ownership is always the agent_id + agent_scope + agent_kind tuple; model policy, quality gates, and output judges live on the selected agent definition and should not be reimplemented in this skill.

MCP Fast Path

Use MCP for deterministic orchestration state when available:

• list_tickets({repo, status}) / get_ticket({slug})
• list_agents({repo})
• update_ticket_agent({slug, agentId, agentScope, agentKind})
• update_ticket({slug, status: 'in-progress'})
• update_ticket_run({slug, runId: $TERMINAL_RUN_ID, runSource: 'agent', runStartedAt, runStatus: 'running'}) when a run id exists
• request_agent_artifact({repo, title, prompt, agentId, agentScope, agentKind})
• file_ticket({repo, title, body, type, priority, source, agentId, agentScope, agentKind})
• emit_activity({kind, repo, title, detail})

Fallback command helpers: .claude/bin/list-agents and .claude/bin/request-agent-artifact.

Invocation

/factory
/factory "payments + auth"
/factory --discover
/factory --max-stack 12

Loop

/merge-sync
  -> select runnable owner lane
  -> mark in-progress and link pickup run/session when available
  -> /stacked-mr pass
  -> handle stuck/HITL lanes
  -> if queue empty and --discover, run discovery/check agents
  -> repeat until no runnable scoped work remains

Responsibilities

Step	Factory responsibility
Reconcile	Run /merge-sync; clear stale stuck tickets when blockers are gone.
Route	Group work by agent_id / agent_scope / agent_kind across classic and persistent agents; split multi-agent work into linked tickets.
Observe	At pickup, set in-progress and write agent_run_* fields with update_ticket_run when a run id is available.
Build/review	Call /stacked-mr; do not duplicate its implementation or review logic. The review checkpoint is the code-review agent, with /code-review only as a compatibility launcher. /stacked-mr also yields one joint /digest over the batch (its Phase 2.5) — the human's read surface; do not gate on it.
Discovery	When --discover is enabled, run discovery/check agents and file owner-scoped tickets.
HITL	File true human blockers, then continue independent lanes.
Context	Compact at pass boundaries; state of record lives on disk.

Delegation

Factory stays thin. Use these patterns:

Pattern	Use for
request_agent_artifact	Focused cross-domain knowledge questions with durable reports.
list_agents	Resolve assignable classic and persistent owner agents.
code-review agent	Review/test artifacts. /code-review remains a compatibility launcher.
/stacked-mr	Full build stack.
Fresh worktree/session	Long-running owner lane that needs isolation.

Keep only summaries and artifact paths in factory context.

Discovery

Discovery is off unless --discover is passed. When enabled:

1. Run the relevant discovery/check agents.
2. Use list_agents({repo}) or .claude/bin/list-agents.
3. File every finding as a ticket with exactly one owner and a concrete quality/eval expectation.
4. Split multi-phase or multi-owner findings into linked tickets with depends_on.
5. Loop back to /merge-sync.

Final State

Report state, not a handoff request:

• passes run
• PR/MR stack in merge order
• tickets at the bar
• stuck/HITL lanes
• follow-up/discovery tickets filed

Hard Rules

1. Never merge.
2. Never alter /stacked-mr's review bar.
3. Never create ownerless tickets.
4. True human-only decisions go to HITL; review failures iterate.
5. No "tell me when ready" while runnable scoped work remains.
6. Do not hardcode persona names or /code-review; route through agent definitions.