Fast Lane Routing Manual

For new work, prepare fast-lane-request-v2 with prepare_model_neutral_fast_lane_request; it pairs only with work-package-v3. Read its plan-v3 assignments as requirements, not a model registry:

• Use route_requirements to judge complexity, capability, effort, and cost.
• Choose from the model IDs and supported efforts listed by the current collaboration.spawn_agent tool metadata. Routine work should stay cost-conscious; design, research, and exceptional cross-module work should use the strongest suitable current model at its highest worker-safe effort.
• selection.state="unselected" is a normal planned state. Call record_model_selection with the exact chosen model ID, effort, and reason, then pass that exact route to the dispatch tool. The record does not claim availability or dispatch.
• An explicit_intent with state="required" is non-substitutable. If the current tool rejects it, report the route unavailable; never silently fall back to another model.
• Model IDs are bounded strings, not a product whitelist. Do not add a model name to DevKit routing tables merely because Codex introduces it.

Legacy request-v1/plan-v2 records remain exact replay contracts: their model, effort, route, worktree, lease, context, and predecessor bindings are not defaults to reinterpret.

• A missing, altered, or unattested route is `NO_SAFE_WORK`.
• A missing first index snapshot in a newly opened project is a cold-start precondition, not a route verdict. Follow workflow-design's one-time project_index_register -> project_index_sync sequence before compiling or declaring a degraded mode.
• Cross-session projection is host-owned; the manual never opens a session.
• `index_context` is bounded evidence, not an invitation to poll or rescan.
• Refill follows a validated terminal event only; retained work stays retained.

See `mcp-tools/devkit_fastlane/FASTLANE_CONTRACT.md` and its references for the complete schema and executable compiler boundary.