sequence-plans
Use when a project has more than one plan and you need to know which is runnable next — reads a committed manifest, runs a Kahn topological ready-set, and dispatches one plan at a time. Do NOT use for intra-plan step order (that is `write-plan` and `sdd`) or for greenfield spec work (that is `iterate`).
- License 许可
- Apache-2.0
- 所属仓库
- zak-keown/moe
- 最近更新
- 2026-09-05
- Stars
- ★ 0
- Forks
- 1
01 前言 Front Matter
- name
- sequence-plans
- description
- Use when a project has more than one plan and you need to know which is runnable next — reads a committed manifest, runs a Kahn topological ready-set, and dispatches one plan at a time. Do NOT use for intra-plan step order (that is `write-plan` and `sdd`) or for greenfield spec work (that is `iterate`).
02 资源文件 1
- 脚本scripts/plan-set.mjs741 B
03 SKILL.md 正文
Sequencing Plans
Overview
write-plan produces one plan. sdd (SDD) executes
one plan and ledgers its own tasks. Neither models the space above one plan.
This skill does — with one committed file per project, one CLI, and one loop.
Use it when brainstorming split a spec into more than one plan and each plan depends on the previous one (or on a diamond of previous ones) landing. The first version of this problem is deterministic and cheap. Nothing here holds state across sessions except the manifest, and the manifest is a git-tracked markdown file — a fresh session on another machine reads the same answer.
Announce at start: "I'm using the sequence-plans skill to run this plan set."
When NOT to use this
- One plan.
write-plancovers it. This skill exists for the cross-plan case. - Intra-plan step order. SDD's per-task ledger already handles that inside one plan.
- Greenfield spec work over a large surface.
iterateis the right cluster for that; its state model is per-iteration, and this skill will not help. - Parallel plan dispatch. Not covered here (
parallel-execution-optionowns the write-ban lift). Run one plan at a time.
The manifest
One file per plan set, committed alongside the plans it names:
docs/moe/plans/<project>-MANIFEST.md. It is a markdown wrapper around one
fenced YAML block. Fields:
plan_set_id— optional stable id unique across the manifests in this repository. It must match[A-Za-z0-9][A-Za-z0-9._-]*so qualified<plan-set-id>/<plan-id>references round-trip without ambiguity. When omitted for a legacy manifest, the CLI derives it from the filename by removing-MANIFEST.mdand validates the derived id with the same grammar.depends_on_plan_sets— optional list of plan-set ids that must be whollydonebefore any plan in this set is runnable. Defaults to[]. A blocked prerequisite blocks every dependent set transitively.id— short label unique in this manifest. Used byplan-set doneand by everydepends_onreference. Recommend the plan slug.plan— path to the plan markdown, from the repository root.depends_on— list of ids that must bedonebefore this plan is ready.[]when nothing precedes it.status— one ofpending,running,done,blocked. New entries start aspending.blockedis terminal and propagates: every dependent of ablockednode is treated as blocked too, so a plan whose foundation stopped by design is never handed back as "ready".commits— set byplan-set done, not by hand.<base7>..<head7>for the range of commits the finished plan produced.
A manifest must contain at least one plan. Empty plan sets are rejected rather than treated as vacuously complete prerequisites.
Example:
# <Project> Plan Set
Prose describing the project. `plan-set` reads only the yaml block below.
```yaml
plan_set_id: example-project
depends_on_plan_sets: []
plans:
- id: schema
plan: docs/moe/plans/2026-01-01-schema.md
depends_on: []
status: pending
- id: reader
plan: docs/moe/plans/2026-01-02-reader.md
depends_on: [schema]
status: pending
- id: writer
plan: docs/moe/plans/2026-01-03-writer.md
depends_on: [schema]
status: pending
- id: server
plan: docs/moe/plans/2026-01-04-server.md
depends_on: [reader, writer]
status: pending
```
The CLI
Resolve skills/sequence-plans/scripts/plan-set.mjs relative to this
loaded document. It is a plugin-owned Node launcher, not a global executable:
invoke the resolved resource with node, never as a bare plan-set command.
It locates the plugin's scheduler independently of the project working
directory and ships in the moe plugin alongside this skill.
Verbs:
plan-set next [--manifest <path>]— Print every id whose plan and plan-set dependencies are alldoneand whose own status ispending. With multiple discovered manifests, output is qualified as<plan-set-id>/<plan-id>. Scoped output remains the legacy bare id. Blocked nodes, sets, and their transitive dependents never appear.plan-set done <id> <base>..<head> [--manifest <path>]— Mark<id>asdoneand record its commit range. In aggregate mode with multiple manifests,<id>must be qualified as<plan-set-id>/<plan-id>. Fails if the id is unknown, if any plan or plan-set dependency is notdone, or if the range is not two 7+-char hex SHAs joined by...plan-set check [--manifest <path>]— Validate: unique ids, all plan files exist, all plan and plan-set dependencies resolve, neither DAG contains a cycle (Kahn), plan-set ids are unique, and no plan path is listed twice even across manifests. Exits non-zero on any failure. A cycle names its nodes on stderr.
If --manifest is omitted, plan-set discovers every matching
docs/moe/plans/*-MANIFEST.md in sorted filename order and operates on the
aggregate. If --manifest is present, it loads only that manifest and its
transitive prerequisite closure. Unrelated malformed siblings therefore do not
poison a scoped command, while prerequisite completion is still enforced.
The SessionStart hook
plan-set-notice is a bash SessionStart hook that fires on startup|clear|compact.
It looks in the session's cwd for docs/moe/plans/*-MANIFEST.md, and if one or
more exist and have runnable plans, it prints the bare or aggregate-qualified
ids and a plan-set next command as SessionStart additionalContext. Fully
deterministic — no model in the loop. Every failure path exits 0; a broken
notice must never break a session.
The hook exists so the cold-start case ("new session, four plans exist, which
are done?") is caught even when the transcript has no history and no skill
description is triggering. plan-set-notice announces; this skill carries the
loop.
The loop
-
Confirm the manifest set. Resolve skills/sequence-plans/scripts/plan-set.mjs relative to this loaded document, then invoke it as
node "<resolved-plan-set.mjs>" checkbefore anything else. A cycle, a missing plan file, or a duplicate id is a dead end before the first plan runs, andchecksays which one at once. Use--manifest docs/moe/plans/<project>-MANIFEST.mdonly when intentionally scoping to that set and its prerequisites. -
Pick the next plan. Invoke the resolved launcher as
node "<resolved-plan-set.mjs>" next.nextreturns a set; the loop takes the first line of it.NEXT=$(node "<resolved-plan-set.mjs>" next | head -n 1)If the output is empty, either everything is
done(the project is finished) or everything ready isblocked(a foundation stopped by design and its dependents cannot run). Do not dispatch in either case. -
Dispatch. Read the plan path for
$NEXTout of the manifest and hand the plan off tosdd(the recommended executor) orexecute-plan(the no-subagents fallback). Both skills already own their own per-task ledger, so this skill does not track task-level state. -
Record completion. When the plan finishes cleanly, capture the branch's base and head SHAs and mark it done:
node "<resolved-plan-set.mjs>" done "$NEXT" \ "$(git merge-base main HEAD | cut -c1-7)..$(git rev-parse --short HEAD)"Add
--manifestonly for a scoped, bare-id workflow.donerefuses to run if$NEXT's dependencies are not alldone— a safeguard against marking things done out of order. -
Loop. Return to step 2.
When to mark a plan blocked
blocked is terminal and propagates. Use it when a plan has stopped by
design — the spec changed, the approach is being reconsidered, an external
dependency is not going to land. Do NOT use it as a synonym for "in progress"
or "waiting for review": those are running.
The reason blocked matters is that its dependents are still pending in the
file. A naive "artifact exists = done" reader hands the agent a plan whose
foundation was never built. plan-set next never returns a dependent of a
blocked node.
Edit the manifest by hand to mark a plan blocked; there is no CLI verb for
it, on purpose — it is a decision, not a state transition.
Design notes
nextreturns a ready set, not a single id. The loop takes the first of it. The set is the seam a later cross-plan scheduler would read, once the parallel-write ban is lifted. Nothing here requires that ban lifted.- The CLI lives under
hooks/, not this skill's directory. SessionStart invokes the adjacent notice before any skill body is loaded. Keeping both runtime pieces together also gives every generated harness the same stable plugin-relative path. plan-setis Node, not bash. Kahn's algorithm plus per-line YAML parsing over a mutable file is short in JS and painful in bash. Node ships with every harness this fork targets.- The manifest is a markdown wrapper around one YAML block, not a bare YAML
file. The wrapper is where humans write the prose that says what the
project is for.
plan-setreads the yaml block and ignores everything outside it.
04 同仓库更多 Skills
write-skill
Use when creating new skills, editing existing skills, or verifying skills work before deployment
write-plan
Use when you have a spec or requirements for a multi-step task, before touching code
write-clearly
Apply Strunk's timeless writing rules to ANY prose humans will read—documentation, commit messages, error messages, explanations, reports, or UI text. Makes your writing clearer, stronger, and more professional.
windows-vm
Create, manage, or connect to a headless Windows 11 VM running in Docker with SSH access. Use when the user wants to spin up, stop, restart, or SSH into a Windows VM.
verify-completion
Use when about to claim work is complete, fixed, or passing, before committing or creating MRs - requires running verification commands and confirming output before making any success claims; evidence before assertions always
using-moe
Use when starting any conversation - establishes how to find and use skills, requiring skill invocation before ANY response including clarifying questions