LexAI/docs/MODEL_ROUTING.md

Model routing

This file is the project's answer to "which model runs what." It is filled once, at first initialization, by /model-routing (or /project-init), and re-run whenever the model lineup or your plan changes. Everything else in the kit refers to lanes, never to a model ID — so the kit survives Cursor's model list changing under it.

Status

Field	Value
Routing filled	yes
Plan / access	Operator-confirmed suggested split (picker verification pending for lead)
Filled on	2026-08-13
Verified against the model picker	partial — operator accepted profile-table IDs; lead must match picker

The four lanes

Lane	Filled value	Roles that run on it	Best use	Avoid
lead	grok-4.5	the Cursor session itself — the model in your picker, not a file	framing, routing, judging evidence, fast-path edits	deep implementation it should have delegated
strong	claude-opus-5	critic · security-auditor · system-steward · planner · gauntlet-critic	adversarial review, security analysis, architecture, gauntlet refereeing, final synthesis	retrieval, boilerplate, deterministic work
mid	composer-2.5	builder · integrator · ux-ui-designer · ux-psychologist	implementation, debugging, ordinary planning, design work	novel high-consequence decisions without review
fast	composer-2.5-fast	scout · verifier · learning-steward	narrow search, running checks, extraction, lesson capture	architecture, ambiguous change, security sign-off

Routing test: can a cheap model succeed given a precise contract and a deterministic verifier? Yes → fast. Known-pattern implementation → mid. Otherwise → strong, then verify independently.

The referee is never cheaper than the builder. gauntlet-critic sits on the strong lane by construction: a referee weaker than the thing it judges rubber-stamps. This is also why Auto is disallowed on strong — a parity verdict from a router that may have silently downgraded is not a verdict. If the strong lane is collapsed, every parity call needs owner sign-off.

Escalate a role one lane only after a concrete failure at its current lane, and record a permanent escalation in docs/DECISIONS.md. When lanes span two vendors, that is a feature: put the second vendor on cross-model critique rather than on a second builder.

The lead lane is a human setting, not a file

This is the one thing Cursor does differently from every other harness in this repo. Subagent models live in frontmatter and are writable. The lead's model is whatever is selected in the Cursor model picker — no project file can set it, and no agent can change it.

So the lead row above is a recorded intent, not an enforced binding. Three consequences:

1. /model-routing asks you to select the lead model in the picker yourself, then records what you chose.
2. The sessionStart hook (.cursor/hooks/session-context.mjs) reads the model Cursor reports for the session and compares it against this row, so a drifted picker shows up as a line in the session context rather than as a mysteriously expensive week.
3. If you work in Auto mode, write Auto (Cost), Auto (Balance), or Auto (Intelligence) in the lead row. Auto is a legitimate lead choice — it is not a legitimate strong lane, because a router that may downgrade under load cannot be the independent judge the quality gates assume.

Where the lane values actually land

Filling this table is not the end of the job. /model-routing propagates the values, and all three must agree:

1. This table — the human-readable contract.
2. .cursor/agents/*.md frontmatter — each subagent carries lane: fast|mid|strong and gets its model: line written from that lane. model: inherit means "run on whatever the lead is running" — the safe default the kit ships with, not a bug, but also the reason an unbound kit has no cost split at all.
3. The model picker — set by you, for the lead lane, and re-checked by the sessionStart hook.

Model profiles

Cursor manages the model list itself, and it changes with releases. The families below were current when this kit was written (2026-08-11). Confirm every ID in the model picker before writing it — an ID that no longer exists fails the Task call rather than degrading gracefully.

Family	Typical IDs	Fits	Notes
Cursor Composer	composer-2.5, composer-2.5-fast	mid (Composer), fast (Fast)	Cursor's own agentic coding model — trained for exactly the builder/integrator loop, and usually the cheapest capable mid.
Grok	grok-4.5, grok-4.5-fast	mid or lead, fast	Cursor-tuned for long-running work; a reasonable lead when sessions are long.
Claude	claude-opus-5, claude-fable-5, claude-sonnet-5	strong (Opus/Fable), mid (Sonnet)	Strongest adversarial-review behavior in this list; the default strong pick.
GPT	gpt-5.6-terra, gpt-5.6-sol, gpt-5.6-luna	strong → mid → fast	A whole ladder inside one family; useful when you want the cross-model critic to come from elsewhere.
Gemini	gemini-3.1-pro, gemini-3.6-flash	strong/mid, fast	Flash is a strong fast lane for search-and-check work.
Auto	Auto (Cost), Auto (Balance), Auto (Intelligence)	lead only	Routes for you; never assign it to strong (see above).

A sensible starting split, if you have no preference: fast = composer-2.5-fast, mid = composer-2.5, strong = claude-opus-5 (referee and critic), lead = whatever you already like driving. Confirm all four in the picker.

Collapsed and constrained lanes

A lane is a role assignment, not a promise of four distinct models. Legitimate collapses:

• One model, four lanes. Supported. Separation of creation from judgment survives because every subagent gets its own clean context window — but the capability asymmetry is gone, so say so below.
• Two models. A cheap fast/mid plus a genuinely strong lane for critic and security-auditor is the highest-value split when budget is tight.
• Degraded lanes must be recorded. If strong is not genuinely stronger than mid, write it in the Notes and treat every high-risk gate as needing a human reviewer — the kit's gates assume an independent, more capable judge exists.

Notes (this fill)

• Four distinct models; no collapse.
• Cross-vendor critique: Composer builds, Claude Opus judges — intentional.
• Lead recorded as grok-4.5 from the suggested split (“whatever you already like driving”) matching this init session’s model family. Select it in the picker.

Change log

Date	Change	Reason
2026-08-13	initial routing filled — lead grok-4.5, strong claude-opus-5, mid composer-2.5, fast composer-2.5-fast	/project-init; operator accepted suggested split