rail

name	rail
description	Ask which rail skill or flow fits. Router over the rail pack.
disable-model-invocation	true

Rail

A flow is a path through skills. Most work follows the main chain.

Main chain: idea → ship

1. /rail-setup — once per repo (all rail docs under docs/monorail/)
2. /rail-align — light grilling → writes align.md; map mode if foggy/multi-session
3. /rail-spec — write docs/monorail/<feature>/spec.md from align.md and/or a cleared map (refuses empty context)
4. /rail-slice — write tasks/NN-*.md with blockers
5. /rail-build — user-triggered; runs frontier tasks serially in one session (fresh implementer sub-agent per non-trivial task, trivial tasks inline); drives /rail-tdd, then marks done. Concurrent builds require separate git worktrees (see rail-build)

Planning chain (auto-continue)

align → spec → slice is one planning chain. When a stage finishes successfully in this session, continue the next stage in the same session by reading and following that skill's SKILL.md — do not stop and ask the user to type the next /rail-*.

• Light align writes align.md → continue /rail-spec → continue /rail-slice
• Cleared map (no open tickets / fog) → continue /rail-spec → continue /rail-slice
• Standalone /rail-spec auto-continues to /rail-slice; /rail-slice ends the planning chain (suggest /rail-build only)

Do not auto-continue into /rail-build — build stays user-triggered: the planning chain ends at slice so you can review the task list before implementation starts. Once a build starts, it runs tasks serially in the same session — finishing one task continues to the next frontier task with no fresh session and no context-clearing between tasks (non-trivial tasks are isolated in fresh implementer sub-agents). Builds ask nothing: defaults are run to queue-clear with one commit per task; the user states any deviation (review pause, no-commit, narrower run) in their command. If the run's context degrades, /rail-pass or a fresh sub-agent resumes it. To run several builds at once, set up one git worktree per task first — more sessions on the same cwd are not parallel-safe.

Stop auto-continue when: the user asked to stop after this stage; context is near limits (use /rail-pass); or an adequacy / map / fail-closed gate says stop.

If the window breaks mid-chain, durable align.md / map / spec.md still feed the next skill when the user resumes.

In-pack handoffs only

Handoffs (auto-continue or suggest) use the next rail skill by exact name (/rail-align, /rail-spec, /rail-slice, /rail-build, …). Do not substitute lookalike skills from other packs (e.g. /to-spec for /rail-spec, /to-tickets / /to-issues for /rail-slice) even if those are installed in the same agent.

Side paths

• No codebase → /rail-grill
• Context full / branch session → /rail-pass
• Hard bug → /rail-debug
• Review a branch / PR / WIP → /rail-review (opt-in; not part of build)

Vocabulary underneath

• /rail-tdd — red-green-refactor (also used inside build)

Map ticket types (not separate skills)

In /rail-align map mode, decision tickets may be Type: decision | research | prototype (see rail-align/map-mode.md). There is no /rail-research or /rail-prototype skill.

Not in this pack (yet)

triage, standalone wayfinder, architecture-improve tours — use other packs or wait for a later rail version.