coordinator:plan

<!-- Purpose: Decision-tree router for plan-writing. Predicate classification and evidence: state/audits/2026-08-08-plan-skill-predicate-classification.md. A plan-router may never auto-resolve a triage exit classified untrusted-gate. -->

Trigger: EM is about to plan implementation work carrying decision weight (multi-file, new abstraction, cross-system, scaffolds new agents/skills, reverses a prior decision) OR the PM typed "write a plan", "break this down", "plan the implementation".

When NOT to use: Trivial work → coordinator:sizing submits it, no plan doc. Implementation-only ambiguity mid-coding → harness Plan tool inline. Architectural-tier (four criteria below) → surface to PM first. Exploration mode or unconverged problem-shape → coordinator:shape. Vague or multi-subsystem spec → coordinator:brainstorming. SKILL.md authoring → plugin-dev:skill-development. Plan needs review → coordinator:review. Stuck → Branch E.

A new skill/agent scaffold is Branch B, not architectural-tier. Architectural-tier requires one of four positive criteria: cross-system-irreversible, multi-stakeholder, security/privacy boundary, naming-collision-with-product-policy. If none can be named, it is not architectural.

Route is resolved by the caller/engine before this skill loads (Branch A). Route-specific detail (Branches B-E, Exit terminals) is retrieved by the plan-assemble brief op, scoped to the resolved route. Segments: coordinator/skills/plan/residue/.

Dispatch authorization — invoking this skill IS the request. The dispatches named below are constitutive steps of this skill, not a separate thing to get cleared: invoking a skill requests the actions that skill performs. A harness line permitting dispatch "unless the user requested it" is therefore satisfied here, not overridden — no precedence claim is needed and none is made. Re-asking spends the very context the dispatch exists to protect. The rule attaches to skill entry and dissolves no PM-authored gate: keyword-gated skills gate entry, and every gate a skill names for itself still binds — per-session cross-repo-commit assent, ask-before-external-action, and any other this skill's own body names. Tripwire: UNATTRIBUTED-HARNESS-LINE-IS-NOT-PM.

---

Branch A — Triage: should I plan, and at what altitude?

Condition: a planning trigger has arrived; decide whether a plan doc is the right artifact.

• An incoming sizing-object is present? (coordinator:sizing resolved this ask and handed a state/sizings/<id>.yaml, or the EM is picking up a sizing-routed baton citing one) → Conform, don't gate — skip the rest of triage — but conform per the object's actual route. The D5 shape-entry gate still wins — wiki § Shape-is-a-conditional-room. → Arrival split. Fresh inbound, accepted sizing, any route the engine carries (dispatch, spec-dispatch, plan) → submit Workflow({name: "coordinator:engine-ask", args: {sizing: "<path>"}}), the single submission; the run plans, executes, reviews and commits — not Branch B/C, no /review --surface plan. Engine refusal → the relay fallback: cite the object's intent (verbatim), estimate, and appetite when present (never block or backfill on it) in B.0's restatement, then Branch B. Via the return edge (verified-scope collapse re-invoked sizing) → conform identically but resume at Branch C; the submission fires only fresh. → route: spec-dispatch is the S lane: set scope_mode: spec-dispatch; the Exit resolves its light terminal. OWES full Branch B substrate verification, B.0's doubt-check, the concurrent-session pre-flight, the cross-plan conflict scan, the pre-dispatch plan-reviewer pass (after the scan, before emit-and-dispatch), and both scaffold-plan invocation points. SKIPS Branch C's four-lens composition (an S-lane body: problem sentence, file scope, acceptance criteria, test surface) and the Opus plan review. NEVER skips scoped-commit discipline or ask-before-external-action. → route: shape / roadmap / pm-decision → plan is not the room. Do not enter Branch B; route to coordinator:shape or coordinator:roadmap-planning, or surface the choice. The engine sets pm_decision_pending and never auto-selects — the pick (PM, or APM in ceo) lands in xl_exit; a null xl_exit never means accept. split: independent initiatives; roadmap: one initiative, several workstreams. Preconditions: shape — JTBD unstated or not falsifiably restatable in the PM's vocabulary; roadmap — spans ≥2 workstreams, or carries/needs an initiative/goal FK; accept_multi_session — neither holds and the PM explicitly assented (xl_exit: accept_multi_session). → plan requires a sizing-object and trampolines back without one. Lifted for plan alone. → Enforcement is EM behaviour: the tools check a cited path or frontmatter; only the trampoline catches ABSENCE.
• No sizing-object, and this ask has not been through the lobby — whatever its provenance? → STOP. Do not enter Branch B. Invoke coordinator:sizing directly — never ask permission to size. On return, re-enter Branch A; this row fires at most once per invocation. All six routes: plan → conform, Branch B as fresh inbound; spec-dispatch → conform, Branch B at S-lane weight; dispatch → abandon the pass and submit Workflow({name: "coordinator:engine-ask", args: {sizing: "<path>"}}); shape/roadmap → the named room; pm-decision → surface the offered exits.
  • Anti-gaming clause. Satisfied only by an artifact on disk in state/sizings/ citing this ask — not by holding file:line substrate, not by concluding the work is plan-tier, not by asking "want me to plan that?".
  • Provenance-blind: a picked-up cross-repo memo's ask needs the lobby too.
  • Carve-outs — exactly two, no more. (1) A continuation is exempt only if its baton cites work already routed: a resolving sizing_object, or a plan (via origin_plan_id/plan_ids, or a deliverable_id some plan on disk carries). Every citation must resolve; never re-litigate a resolved citation. A baton citing none is unsized whatever its provenance. Tripwire: A-BATON-IS-NOT-A-SIZING-ARTIFACT. (2) The express lane never reaches plan.
• Trivial? (single-file change, no new abstraction, scope obvious) → Submit it (Workflow({name: "coordinator:engine-ask", args: {sizing: "<path>"}})). No plan doc.
• Implementation-only ambiguity? (choosing between two valid shapes mid-typing) → Harness Plan tool inline. No plan doc.
• PM has set a session axiom? ("we are going to do X" / "build Z this session" — a directive naming the work, not a question about it) → Disposition flips to plan: plan the named work, do not re-litigate it. Continue to Branch B. The architectural-tier check still fires. This row disposes; it does not admit: with no sizing object on disk the trampoline row above fires first.
• Trigger arrived via a PM-handed pickup whose handoff prescribes a plan? → Already authorized — do NOT ask "want me to plan?". A T3 spinoff fork stays separately PM-gated (surface it as a one-line candidate). Continue to Branch B.
• PM in exploration mode, OR problem-shape unconverged, OR the EM detects it is guessing at the problem? → Propose coordinator:shape before committing to plan; its exit gate chains back here. A PM session axiom and the architectural-tier check both win. The PM HAS a problem → shape; the PM does not know what to build → brainstorming.
• Cross-repo work whose shared contract is itself the unknown, even when a handoff says "ready to execute" / "XS"? (negotiated co-design — the contract/hookspec/shared fixture takes round-trips to converge) → Invoke coordinator:plan anyway — do NOT execute off the handoff's t-shirt. Continue to Branch B.
  • Negotiated co-design is the trigger, not a sibling's existence: a memo (one ask) does not escalate; no nameable coordinating party and contested shared contract → this row has not fired.
  • A multi-wave plan that sends a memo carries a wave-boundary re-read of it, written into the spine: at each wave close, enumerate the memos sent and ask whether the wave just landed contradicts any; the correction memo is part of that wave's deliverable. Tripwire: A-MULTI-WAVE-PLAN-NEVER-RE-READS-ITS-OWN-SENT-MEMOS.
• Non-trivial (default for everything else)? (multi-file, new abstraction, cross-system, scaffolds an agent/skill, reverses prior teardown, touches shared schema) → Continue to Branch B, and scope the task list to the COMPLETE problem set; ADR, lineage and docs chunks are not execution-wave rows: list them under a non-empty ## Distill pass heading, referenced from the closing handoff or the /workstream-complete step, which hands them to /distill.
• Architectural-tier? (the four positive criteria above) → Surface to PM: "this looks architectural — propose /staff-session, want me to draft the brief?" Wait for PM.

---

Exit — Route-Selected Terminal

Condition: body drafted and saved to docs/plans/YYYY-MM-DD-<slug>.md. The inbound sizing route selects which terminal fires. shape/roadmap/pm-decision never reach here — Branch A diverted them before Branch B.

The plan file MUST be produced and committed via scaffold-plan, not hand-authored. It owns frontmatter-skeleton emission and the write-time commit across two invocation points either side of body authoring. Every terminal holds this.

Invocation point 1 — scaffold. The generator is the single emission point for frontmatter — never hand-author field lists against schemas/plan.schema.json; hand-authored frontmatter is conformed through the same invocation.

Invoke the .exe launcher by absolute path via the PowerShell call operator (Shape W) — ladder: ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md.

`& "$env:COORDINATOR_SETTINGS_HOME\bin\coordinator-doc-new.exe" --type plan --title "<title>" --sizing-object state/sizings/<the-object-that-routed-you-here>.yaml --out docs/plans/YYYY-MM-DD-<slug>.md`

Pass --sizing-object. coordinator-doc-new --type plan hard-refuses without an explicit --sizing-object/--no-sizing-object, exit 1, nothing written. The flag also writes the reverse edge onto the cited sizing (plan FK plus status flip).

Read the scaffolded file before authoring the body.

Invocation point 2 — commit. Commit the moment the body is saved, then proceed to review, per snippets/scoped-commit-route.md — pathspec exactly docs/plans/<slug>.md, subject plan(<slug>): draft; never a sweep commit.

---

Test Surface

No runtime test for this skill body — prose-doctrine. The grep-asserts below stand in, one count per token against the named file.

#	token	file	expect	threshold reason
1	Eighth dimension	core + shared segment	≥2	table self-citation
1	trampoline: true	core + shared segment	≥2	same
2	plan⇄spike	core + shared segment	≥2	same
2	plan⇄spike	skills/spike/SKILL.md	≥1	both sides
3	coordinator:plan	skills/spike/SKILL.md	≥1	viable exit
3	coordinator:shape	skills/spike/SKILL.md	≥1	not-viable exit
4	"viable"	schemas/spike-result.schema.json	≥1	schema-backed enum
4	"not-viable"	schemas/spike-result.schema.json	≥1	same enum
5	plan⇄sizing	core + plan-route + shared segments	≥3	split across segments
6	plan requires a sizing-object and trampolines back without one	this file	≥2	1 = gate is gone