Plan

coordinator:plan

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 (single-file fix, typo, link repoint, no abstraction) → just do it. Implementation-only ambiguity mid-coding → harness Plan tool inline. Architectural-tier (cross-system irreversible, multi-stakeholder) → surface to PM first. PM in exploration mode or problem-shape unconverged → coordinator:shape (Branch A). Spec vague or multi-subsystem → coordinator:brainstorming. Writing a SKILL.md → plugin-dev:skill-development. Plan written and needs review → coordinator:review. Stuck pattern → Branch E.

A new skill/agent scaffold is Branch B, not architectural-tier. "Novel", "new surface", "be careful", "this is unfamiliar" are conservative-bias tells. Architectural-tier requires one of four positive criteria: cross-system-irreversible, multi-stakeholder, security/privacy boundary, naming-collision-with-product-policy. If you cannot name which fires, it is not architectural.

Route is resolved by the caller/engine before this skill loads (Branch A). The route-specific procedure detail below this point — Branch B substrate verification, Branch C composition lenses, Exit terminal specifics, Branch D drift handling, Branch E friction — is retrieved by the plan-assemble brief op, which reads the segment set scoped to the resolved route. Segment content: 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 over whatever base route the object carries — wiki § Shape-is-a-conditional-room, not this file. → Arrival split, for route: plan and route: spec-dispatch alike. Fresh inbound (the lobby resolved this ask for the first time, or the baton cites an object not yet conformed against this pass) → cite the object's intent (verbatim), estimate, and appetite when present (optional, usually absent — never block or backfill on it) in B.0's restatement, then proceed to Branch B. An unanswered post-size prompt does not gate plan authoring and is not re-asked here or at the execution gate. Via the return edge (Branch B's verified-scope-collapse row re-invoked sizing this pass) → conform identically but resume at Branch C, since Branch B already ran all-green — re-running wastes it and lets the collapse row fire twice. → 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 proportional doubt-check, the concurrent-session pre-flight, the cross-plan conflict scan, and both scaffold-plan invocation points. SKIPS Branch C's four-lens composition (an S-lane body is four parts: problem sentence, file scope, acceptance criteria, test surface) and the Opus plan review — the rest of Branch C runs at S-lane weight. 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 PM's pick lands in xl_exit, and a null xl_exit never means accept. split is retired: an ask decomposing into independently shippable pieces is roadmap. Preconditions: shape — JTBD unstated, or you cannot falsifiably restate the problem 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 — this is a wall, not a courtesy. The do-not-build-a-wall anti-scope ruling is lifted for plan alone; it still governs shape, roadmap-planning, and goal-setting unchanged. Do not soften it back. → What the machine actually enforces. scaffold-plan --sizing-object refuses an unresolvable path; assert-plan-sizing-citation sweeps frontmatter only, never body prose. Neither catches ABSENCE — an omitted flag writes clean, and a body-only citation is invisible to both. Only the trampoline covers absence, as EM behaviour. Cite the wall as a rule that binds you, never a mechanism that would have caught you.
• 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, never offer sizing as a question. On return, re-enter Branch A; the conform detent fires because an object now exists, so 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 → plan is not the room; abandon the pass and dispatch directly (clean by construction — scaffold-plan runs at Exit, so nothing is scaffolded yet); shape/roadmap → the named room; pm-decision → surface the offered exits. Termination: sizing always writes an object before returning, so re-entry lands on the conform detent, never back here.
  • Anti-gaming clause. Satisfied only by an artifact on disk in state/sizings/ citing this ask — not by already holding file:line substrate, not by silently concluding the work is plan-tier in your head, not by asking "want me to plan that?".
  • Provenance-blind, not just PM-direct-blind. A picked-up cross-repo memo's ask is novel work in THIS repo, sized against the sender's substrate — it needs the lobby too. The trampoline fires for it.
  • 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 via a deliverable_id some plan on disk carries, the ordinary mid-execution baton's only link back. Every citation must resolve; an unresolvable one is a broken pointer, not a size — never re-litigate a resolved citation, and never audit a baton for one you were not handed. A baton citing none is unsized whatever its provenance — a spinoff mints its own fresh deliverable_id, naming only itself — and the trampoline fires. Tripwire: A-BATON-IS-NOT-A-SIZING-ARTIFACT. (2) The express lane is not a plan-side carve-out — PM ruling: "just do it doesn't make it to planning anyway. that's an exit for the sizing lobby." Such an ask never reaches plan.
• Trivial? (single-file change, no new abstraction, scope obvious) → Just do it. No plan doc. See ${CLAUDE_PLUGIN_ROOT}/snippets/em-operating-doctrine.md § How to Plan and Hand Off.
• 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, not brainstorm: the PM has chosen scope, and your job is to plan the named work, not re-litigate it. Continue to Branch B. The architectural-tier check still fires — an axiom does not override an architectural surface-to-PM. This row disposes; it does not admit. An axiom carries no exemption from the sizing wall: with no object on disk the trampoline row above fires first — it is provenance-blind by design — and re-entry lands on the conform detent.
• Trigger arrived via a PM-handed pickup whose handoff prescribes a plan? → Already authorized — do NOT ask "want me to plan?". A handoff is PM-authored, so an embedded plan trigger satisfies the /plan keyword-gate transitively. A T3 spinoff fork stays separately PM-gated (surface it as a one-line candidate) but must not block or be conflated with the plan. 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. This is the seam where premature convergence happens. Precedence: a PM session axiom and the architectural-tier check both win — /shape is the exploration branch, not an override of a directive to plan named work. Discriminator: the PM HAS a problem and wants confirmation you understood it → shape; the PM does not know what to build at all → 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, not a single well-understood ask the sibling implements on their own surface) → Invoke coordinator:plan anyway — do NOT execute off the handoff's t-shirt. Handoff sizing runs systematically low here: investigation reveals the structural problem, planning reveals the implementation blast radius (N repos, M reviewers, contract fixtures still converging). Continue to Branch B.
  • The existence of a sibling is not the trigger — negotiated co-design is. A memo (one ask, the sibling implements it on their own surface) does not escalate here — it is a blocked_by/awaiting_gate fact, not coordination cost. If you cannot name the coordinating party and the shared contract still being negotiated, this row has not fired.
  • A multi-wave plan that sends a memo carries a wave-boundary re-read of it. The memo is an output of the plan, and a later wave of the same plan can contradict it — a wave-1 promise of "C5 ships only the report-dict half, zero contract movement" survived intact until wave 3 moved the contract, leaving the sibling holding a promise this tree had stopped keeping, with no surface on either side able to notice. Write the re-read into the spine, not into anyone's memory: at each wave close, enumerate the memos this plan has sent and ask whether the wave just landed contradicts any of them; where it does, the correction memo is part of that wave's deliverable. Prefer scope claims stated at their ref ("as of C5 as planned") over unqualified promises about what the plan will not do — the promise is written before the waves that could break it exist. 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 eventual task list to the COMPLETE problem set.
• 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.

Known boundary: this floor protects the plan path only — work mis-triaged as trivial at Branch A bypasses /shape and the doubt-check. Mitigation is EM alertness at that row, not a second doubt-check.

Exit — Route-Selected Terminal

Condition: body drafted and saved to docs/plans/YYYY-MM-DD-<slug>.md. Which terminal fires is selected by the inbound sizing route. 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 as one named unit spanning two invocation points either side of body authoring. This holds for every terminal; only what follows the commit differs by route.

Invocation point 1 — scaffold. The generator is the single emission point for frontmatter — never hand-author field lists against schemas/plan.schema.json. If frontmatter was hand-authored, conform it now through the same invocation.

Invoke the .exe launcher by absolute path via the PowerShell call operator (Shape W) — ladder and shapes: 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 — mandatory to the tool, not merely to you. coordinator-doc-new --type plan hard-refuses without an explicit --sizing-object/--no-sizing-object, exit 1, nothing written. Branch A refused you entry without a sizing object, so you hold its path. The flag also writes the reverse edge onto the cited sizing (plan FK plus status flip, same transaction) — omit it and the sizing never learns it was routed.

Read the scaffolded file before authoring the body — coordinator-doc-new writes it via Bash, so the first Write/Edit bounces until it is read once.

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. Scope stays to the single plan doc, never a sweep commit. Distinct from the review-integrator's commit-after-integrate discipline.

Test Surface

No runtime test for this skill body — prose-doctrine, not executable code. The applicable automated check is skill-body lint / frontmatter validation; the grep-asserts below stand in for a test harness, each a single grep -c against the named file.