Daily cycle — factory

Experimental. A day on the factory layer has four beats: turn an idea into a reviewed design, pick how the work should run, ask the pipeline what to start next, and bring the finished branch home. This is the steady-state loop behind First Steps — factory, which walks the same beats once with a first task.

/arch — an idea becomes a design

Invoke /arch <topic>: the external design contour turns a raw idea into a reviewed design plus a routed handoff. On a harness without skills, read .claude/skills/arch/SKILL.md and follow it by hand. A design that comes out of /arch plan-complete also changes how the task routes — a plan-complete kickoff from /arch is exactly the case the tier rubric routes whole-pipeline to the executor tier; the rubric owns the conditions.

preset — pick how the work runs

Launch presets are named run configurations for the pipeline: /pipeline <task> --preset <name>. The list below is vendored from the shipped preset data (.claude/skills/pipeline/references/presets/*.json, read at framework staging 94a3a9efcd; the landing carries no mechanical regen check — re-vendor on upgrade):

• aif — Autonomous overnight aif-handoff dispatch (mode=autonomous, project-default profiles, no marker)
• economy — Cost-conscious whole-line on executor tier (mode=whole-line-executor, marker=Z.AI GLM-5.3 SDK)
• night — Night-mode unattended single-session (mode=mode-a-inline)
• sdd — Interactive single-feature SDD (mode=in-session)

Resolve details (tier routing, predicates) live in the preset JSON itself and .claude/skills/pipeline/references/mode-overrides.md.

status — ask the pipeline what to start next

Invoke /pipeline with no task: it reads your kickoffs plus .ai-factory/orchestrator-prompts/plan.md (created on first run), ranks them, and emits a launch table. An empty backlog just renders the overview with zero open umbrellas — that is normal, not an error. The table's ranking is where the tier rubric meets your backlog: design-heavy rows plan on the top tier; mechanical rows run whole-line on the executor tier. That consumer staging dir is the template-stage-created twin of the framework's own staging home: in the framework repo the in-flight umbrella kickoffs live under .claude/orchestrator-prompts/<umbrella>/kickoff.md (see setup.d/LAYERS.md:121), while consumers get an empty .ai-factory/orchestrator-prompts/ created by the installer (setup.d/30-templates.sh:17).

dispatch → harvest — start it, then bring it home

Dispatch the launch table's top row, then bring the finished branch back with /harvest. The worker works on its own branch inside your compiled gates. If a task stalls or the runtime misbehaves, /aif-doctor is the diagnostic entry point. Under the hood the dispatch beat is the runtime-bridge CLI tsx .claude/vendor/runtime-bridge/src/cli/dispatch.ts <kickoff-path> [--force] (packages/runtime-bridge/src/cli/dispatch.ts:4), invoked for you by the PostToolUse hook when a kickoff's first line is <!-- bridge: auto -->, or run manually on demand — it exits 0 on every dispatch outcome (including the ManualBackend fallback), 2 when the dispatch spec itself is invalid, 1 on a call defect (dispatch.ts:32-38).

Its scope is the aif-handoff runtime itself — a task stuck or crash-looping, new tasks staying backlog at capacity, a broken runtime — and it stays invokable even when the dispatcher is NOT running (.claude/skills/aif-doctor/SKILL.md:3). The install also delivers executable heal helpers and keeps them executable when the gated suite was installed (setup.d/10-skills.sh:188).

Next

Why work routes to different tiers at all: Overview — multi-model pipeline. What degrades when a capability is absent: Degradations.