Sizing — the Build Door

Two doors: this is the sizing door; /coordinator:research is the research door.

Rationale: coordinator/docs/wiki/planning/sizing-lobby.md

Default: submit, then sit back. One call, the PM's words verbatim: Workflow({name: "coordinator:engine-ask", args: {ask: "<PM verbatim>"}}) — the user's opt-in, never a question. Its hook runs sizing through commit and close (ENGINE-ASK-SUBMISSION-IS-THE-RUN). To add a brief, mint with op sizing.intake_mint and submit the args: {sizing} line it returns; a refused mint leaves the bare args: {ask}. Separate asks go in one args: {asks: [...]}, never bundled (wiki sizing-to-commit-chain-contract). While live the EM does not dispatch, poll, verify or relay; a PM revision goes in verbatim via workflow.revise — coordinator-invoke workflow.revise '{"run_id":"<run id>","verbatim":"<PM verbatim>"}' (run_id is the ask-… id of the live run, the dir name under state/warp/; a peer session adds "from":"<its session id>"); never hand-append requirements.md; a halted run also needs a refire of engine-ask {sizing}. Below is the in-chat lobby, only for a shape/roadmap/pm-decision or engine refusal. Tripwire: NOVEL-WORK-IS-ONE-FIRE-NEVER-A-RELAY.

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. The t-shirt picks a ROOM, not a cost. No appetite question first.

---

The flow

Step 0: the confident-XS predicate, before any other read. The scout is skipped only when all six hold: (1) the PM names, or you can name without a probe, the one existing module the change lives in; (2) no new abstraction or agent, and a new file only as a sibling of a pattern that module holds (a test file, another adapter instance); (3) no path outside this repo; (4) no held capability beyond the checkout (Step 1a); (5) the change has shipped before in this repo (--precedent shipped-before); (6) the PM's words carry one instruction. All true: sizing-assemble --express-lane --tshirt XS (Step 3) writes no object; submit args: {ask}. Any doubt fails the predicate. Otherwise coordinator:sizing-scout sizes, in the submitted run only; a hand dispatch is refused. Its brief carries no t-shirt, lean, guess or route; only the PM's verbatim words, pm-turns window output, EM session id and interaction_mode. The EM's act is the one args: {ask} call. Tripwire: THE-EM-TRIGGERS-SIZING-AN-AGENT-SIZES-IT.

Cross-examine the report: spot-check one claim against the tree; a revision names its mechanism; add capabilities the scout missed; check design_route (canvas halts for PM board review pre-code). Then Steps 3-4 from assemble_flags. {refused: true}: fix the brief.

1, 1a, 1b. The scout's work; contract: coordinator/agents/sizing-scout.md. Size on unknown mechanisms (depth), not touchpoints or appetite, and on what the performer must HOLD (credential, session, mounted store, guarded path, operator at a keyboard): name the capability, never the route tried (A-SCOPE-FIELD-NAMES-A-SUBJECT-NOT-A-BLAST-RADIUS). Each non-express-lane sizing states --premise-provenance executed|read|not-applicable, justified in premise.evidence (Step 4).

2. Substrate probe — L/XL/shaky reads, mandatory on XXL. For a judgment the scout leaves under unknown_mechanisms, fire a general-purpose Sonnet research scout briefed with ${CLAUDE_PLUGIN_ROOT}/snippets/internet-research-scout.md; fold the result into --probe-signal collapse|raise, --scout-evidence-kind mention-count|change-set|site-count, and on a raise --probe-raise-basis ask-scope|substrate-condition|breadth (engine applies only ask-scope). An unproven mechanism goes to the research door's spike class.

3. Compute the route — never hand-derive the table (route=shape included: engine-resolved from --jtbd-unclear/--well-trodden-step-change). Invoke bin/sizing-assemble per ${CLAUDE_PLUGIN_ROOT}/snippets/resolve-coordinator-bin.md. --tshirt <XS|S|M|L|XL|XXL> is the only required flag; --appetite only if the PM volunteered one verbatim. Always pass --intent "<PM's words>" with --intent-source pm-verbatim for a PM turn copied exactly, em-elaborated only when you restated it, --precedent shipped-before|novel, --boundary-in-notch yes|no (§ Appetite guards), the Step 1b/2 flags (from assemble_flags, plus --interaction-mode, --premise-evidence), and --jtbd-unclear/--well-trodden-step-change/--express-lane as warranted. Research door: --research-class (PM's words), --research-source (lobby wiki). Push route/detents/next_move verbatim, except a sizing carrying --research-class fires /coordinator:research --sizing <path>, not the plan chain.

4. Write the sizing-object for any non-express-lane sizing: re-run Step 3's call with --write state/sizings/<date>-<slug>.yaml and --premise-evidence "<text>" (Step 1b's justification). --exit-criterion is owed here at every size on a firing route; it names behaviour and "tests covering touched files pass", never a test tier or tier-valued argument (fast, full, Tier-U, tierU), even when the ask is about one. The record lands routed; never hand-edit it. Undecided direction-class items go in surfaced_to_pm, never fork/xl_exit. Curate the embedded pm_verbatims by cut and paste only. Optional --name sets name, a whiteboard label. Every written sizing carries the PM's instructions as a requirement register, at every size. The engine derives it: pipe the scout's report into the Step 4 write (sizing-assemble ... --write <path> --scout-register -) for one row per PM instruction; only scout rows quoted verbatim are kept, the rest go under unrelated_turns. Read the register back; amend only what it got wrong with sizing-assemble --register - --write <the same path> (body {sources: [{doc, ref}], rows}; rows merge by id, drop: [ids]; each row carries status and an enum surface; wiki requirement-register), never a scratch file. Every plan claims its rows (register_claims); the terminal stamp refuses implemented until each is met and wired or ruled. Splitting an ask never sheds a row. Tripwire: THE-PM-INSTRUCTIONS-ARE-REGISTER-ROWS-AT-EVERY-SIZE.

status. XS: the quick-wrap close stamps shipped; hand-stamp (cite the commit) only on j-dispatch-sizing-ship-failed or an uncommitted sizing file. S+: the run's commit stage owns the stamp, gated on a met judge (A-PLAN-SELF-COMPLETES-ONLY-ON-A-MET-TERMINAL-JUDGE); between ceremonies plan-status-transition stamp-implemented cascades it; never hand-stamp. Read status back (never acted). Tripwire: ACTED-IS-BLIND-TO-THE-DELIVERABLE-CASCADE.

4b. Open the flight recorder on the resolved route, before anything downstream: TaskCreate (ToolSearch it first), a session-goal task naming the ask and sizing path; on the relay fallback one task per remaining stage (wiki § Step 4b).

5. Size approval. XS through L on a plan, dispatch or spec-dispatch route needs no PM size approval in any mode. After Step 4's --write there the EM's act, same turn, no ask, is Workflow({name: "coordinator:engine-ask", args: {sizing: "<path>"}}), end of turn; a printed Workflow line is not fired by hand. emit-wave-fire --from-sizing runs only on an engine refusal, named in the report; Agent-tool hand-dispatch is never the exit. A live plan, dispatch or spec-dispatch sizing never enters /shape, and a baton-carrying one gets its in-flight fire, never a second (A-ROUTED-SIZING-NEVER-ENTERS-SHAPE). In ceo XL/XXL do not stop either: the run's APM rules xl_exit. In pm and hands-on XL+ keeps the PM in the chat; a shape route keeps the room in every mode; a pm-decision route goes to coordinator:apm first in ceo. Where the PM stays (post_size_prompt_pending, M+): ask once, in the PM's register, then stop: "Looks like an X — go with that, split it, cut it, what's up?" Never a closed fork or human time (A-PM-ASK-STATES-SCOPE-NOT-HUMAN-TIME-GUESSES-OR-RULED-CONCERNS). Record in pm_resolution (decided_on: YYYY-MM-DD required), plus fork when cut/raise-shaped.

5b. route: pm-decision bundles into the same ask — the M+ prompt and the XL exit question are ONE PM ask, never re-asked once answered (record the quote as pm_resolution). xl_exit stays null until the PM (pm, hands-on) or the run's APM (ceo) picks shape, roadmap, or accept_multi_session. Wiki § Step 5b/5c.

5c. Turn 1 exit on route: plan where the engine kept the acceptance touchpoint (XL+, or a resize into XL; XS–L skips this step's ask and sizing-accept-exit-criterion). Touchpoints follow interaction_mode. Pass --exit-criterion (behaviour; evidence the tests covering touched files), --interaction-mode only if the PM named one (INTERACTION-MODE-IS-NEVER-READ-FROM-JOB-MODE); source documents add requirements to the Step 4 register; user-facing: --click-path "<role>: <nav> > ... > <page>" per role (A-CAPABILITY-UNREACHABLE-FROM-THE-UI-IS-NOT-DELIVERED) — and record the PM's answer with sizing-accept-exit-criterion --sizing <path> --pm-quote "<words>" [--statement "<amended>"] --mode <mode>, never by hand. In pm or ceo at XL+, --apm-ruling "<ruling>" --ruling-ref <ruling path> replaces --pm-quote; only hands-on at XL+ needs the quote. At XS–L the EM writes the criterion itself (sizing-accept-exit-criterion --sizing <path> --statement "<criterion>", any mode), never asking the PM to set or accept one; an engine refusal demanding acceptance there is an engine defect, not a PM ask. In pm and ceo the ask says that accepting it authorizes execution without a further ask. After acceptance submit Workflow({name: "coordinator:engine-ask", args: {sizing: "<path>"}}): the run commits and ships, so no dispatch.land_run follows. Relay fallback, only when the engine refuses that submission: emit-wave-fire --from-sizing <sizing path> --repo-root <abs> --trail-dir <abs> (never /spinoff or /handoff); name the refusal.

6. Hard gate. The t-shirt→route map binds absolutely, no override. A route that feels wrong means the size was wrong: fix it with evidence (symmetric resize, or the plan⇄sizing return edge).

---

Appetite guards

appetite is the PM's stated budget, set after the size lands; it never moves the estimate. A cross-team dependency is a gate (blocked_by/awaiting_gate), never the t-shirt: --boundary-in-notch no.