The main loop may now move down to a cheaper model by default (mainModelSwitch true, fail-closed picker). Both caps measure the COLD READ a switch costs: the context minus what the target model's prompt cache already holds, from a per-model served record written on every main step (1 h TTL, cleared by /clear and /route reload; a smaller context than the record counts as cold). downgradeMaxTokens defaults to 120000, upgradeMaxTokens keeps 200000; no return bypass is needed since a warm return is cheap by construction. The prompt classifier writes the full phase route: coding and bookkeeping guesses move main down under the switch and the cap; a guess never ranks above the session model; a mid-turn ultrathink floor keeps its model; a user /effort typed between turns drops the guess for that turn only (engine baseline is now per turn). A model-loaded helper skill inside a best-tier run moves effort only. Main routed onto a model whose window the context outgrew returns to the session model. The skip line is said once per cap crossing; the toast names a cap hold; the status line shows the switch only when off.
3.4 KiB
Route doctrine — phase-level model and effort on the main loop (BDR-107)
Shared include, companion of lib/model-gate.md: the gate fixes WHICH model
reflects, the model-router mod (mods/model-router) fixes HOW HARD each
phase thinks. Rungs: low (fix a line, run a script) · medium (day-to-day) ·
high (refactor, resisting bug) · xhigh (architecture, audit before
validation) · max (stuck error, judged need).
The tool
mcp__model-router__route (params phase | effort | clear). It is a
deferred tool: when not loaded, run
ToolSearch("select:mcp__model-router__route") once per session. A route
applies from the next request on, paired with
another tool call or not (pairing only saves a request). The answer always
names the id and effort main runs on. A skill with a row routes itself on
load; a skill without one changes nothing, the last ROWED skill wins.
Wiring points
- Dispatch span starts →
route(phase="orchestrate"), sent with the dispatch. - Reflection resumes (challenge synthesis, verdict, plan revision) →
route(phase="reflect")or"plan"per the skill's own level; the line before everylib/challenge-plan.mdcall. - Bookkeeping tail (memory commit, doc commit) →
route(phase="apply"). - Escalation →
route(phase="escalate"): verify-secure loop caps and ship-feature STEP 4b. Not automatic: the challenge fail-safe and "gone WRONG → STOP"; their STOP text names the levers below. - Built-in judgment dispatch (
general-purposemodel="opus",model: "fable"skill-runners) → expliciteffort=on the Agent call (xhighfor opus reviewers,highfor fable runners). A main route never reaches a child. Typed agents run on their row, never on a shift. - After a prose gate that ends the turn, the resumed reflection phase starts with its own route call.
Run slot and levers
A best-tier skill row survives the end of the turn (a run spans prose
gates); /route clear, /route off and a user /model drop it. Levers for a
relaunch: ultrathink in the prompt (turn floor) or /route effort=max
(sticky, /route clear after).
Builtin /effort is NOT a lever inside a run: rows and routes outrank it.
Limits
- A skill typed while a background agent is live routes only through the typed marker (unverified live 2026-10-10).
- Headless (
-p, SDK) runs the hooks, so routing works there too. - A plain typed prompt with no route in force is classified by a small
model: the guessed phase sets the turn's effort and, under
mainModelSwitchanddowngradeMaxTokens(max cold read, default 120000), may move main DOWN, never up. A user/modelonly moves the baseline the guess starts from (/route classify offto stop guessing; never-p). - Mod off: typed agents fall back to their
model:/effort:frontmatter. - First use of a row asks once, with context (Later, Keep or Change: model,
then effort, then Everywhere or this project only); the answer is kept in
mods/model-router/routing.json./route pendinglists what is still unconfirmed,/route ask off|ontoggles the dialog,/route forget <name|all|projects>takes a decision back.
Measure the split any time: python3 ~/.claude/lib/effort-audit.py
(thinking/output/cache tokens per scope, model and effort).
Never
- A route inside a dispatched agent: its row rules there.
- Max is for diagnosis, not for retrying the same fix harder.