mwguerra/plugins-haiku/route

Route v2 — pipeline + tier, together

Route names the ordered pipeline (which entry points, in what order) AND the tier (vote count, probe set, independence) in one output. The source suite's doctrine — "router = what pipeline; brain = how deep" — is reversed here: depth is sized by tables at routing time, per docs/adr/0002-router-sizes-depth.md. Route recommends and hands control back; it never dispatches and never hard-blocks.

Step 1 — read the tier (never estimate it)

bash ${CLAUDE_PLUGIN_ROOT}/hooks/lib/sizing.sh        # counts the real diff
cat .maestro/evidence/<diffhash>/sizing.json          # { tier, score, inputs, prescription }

No diff exists yet (pure question / greenfield ask) → use the lane's tier floor from the table below, and re-run the hook once a diff exists; the computed tier then replaces the floor if higher.

Step 2 — classify the lane

Intent verb picks the row; sizing flags override upward:

• understand / audit / review / why → analysis · broken / bug / error / crash /

wrong-output → bugfix · build / add / implement / create / feature / design / refactor → change-or-feature · health / tech-debt / production-ready / roadmap → health. ("Fix" follows its object: fix a crash → lane 3; fix a typo with no behavioral failure → lane 2.)

• Any presence flag in sizing.json (migration, schema, auth, payments, deletion,

public API) → the lane escalates to lane 4. Flags are counted by the hook, never felt.

• A change crossing a documented seam in docs/architecture/ → lane 4.

Effective tier = max(lane floor, sizing.json tier). The tier only moves up (the escalation ladder is a one-way ratchet); no prompt may lower it.

Lanes 1 vs 5 share the deep-analysis first step: pick 1 to stop at the findings; pick 5 only when a re-entering roadmap is the ask.

Two stamps on every lane

1. Read `docs/` first: STATUS.md, the relevant docs/architecture/, the open

docs/adr/. docs/ absent → say so and point to scribe:init.

1. The tier rides the recommendation. adversarial-verify and regression READ

their vote count and probe set from sizing.json — they no longer size themselves.

Conflicting signals

• Any flag set → the flag wins; lane 4, no exceptions.
• Zero flags and thin signals → lane 2, plus the named escalation condition printed

with the route ("if this touches the orders→billing seam, re-route to lane 4").

What route never does

• Never dispatches or hard-blocks — the real gates (evidence-bundle commit/stop

gates, taskmanager's done-gate) stay downstream.

• Never invents a tier — sizing.json is the only source; artifact missing → run

the hook, never guess.

• Never re-detects "is this a feature?" — it names existing entry points and

composes with their triggers.