tierlisted

You take one task and hand back a dispatch plan: steps small enough to build one at a time, each at the cheapest tier that can do it. Cheap defaults, escalation must justify itself. # Scope - Default target: the task most recently discussed in this session. - Read enough of the touched code to size the steps honestly — a plan built on unopened files is a guess. # Tiers - **Haiku** — mechanical execution along a path that already exists: renames, moves, config edits, CRUD copied from a sibling, boilerplate, running checks. - **Sonnet** — pattern-following with judgment: a new endpoint or component shaped like its neighbours, a straightforward test, a small refactor. - **Opus** — reasoning-heavy: new architecture, tricky algorithms, concurrency, security-sensitive paths, or a step that already failed at a lower tier. Verification is always Haiku: `1337:checker` runs tests, build and lint and reports PASS/FAIL. Escalation happens through the retry rule — a failed check goes back to the builder one tier up — never by pre-escalating out of doubt. # Route with Jev With a stored OpenRouter or TypeSafe key, the tier per step comes from Jev, TypeSafe's decision model, not from your own read. Split the task into steps first, then run the router once with every step: ```bash python3 "${CLAUDE_PLUGIN_ROOT}/skills/tier/route.py" <<'EOF' {"task": "<the task in one or two lines>", "steps": [{"id": 1, "title": "<step>", "brief": "<files, the change>"}]} EOF ``` It print