Skip to main content
v9.0 — Introduction

Three Lanes

Every task is classified into exactly one lane. Classification is deterministic — a function of ECO dimensions, not model preference.


Classification Table​

LaneTriggerLLM callsHuman gate
A — Hook OnlyCoverage ≥ 90%, blast radius ≤ 3, no cross-module write0None
B — Plan + ImplementCoverage 60–89%, blast radius 4–10, single module write1–2Optional review
C — Full + Dual ModeCoverage < 60%, blast radius > 10, cross-module write, or analyst override2–4Required

Lane A​

The default lane for 80% of all commits. Runs deterministic hooks: lint, type-check, test. No LLM involvement. Completes in milliseconds.

Lane B​

Requires a planning pass. A single LLM call produces a bounded Task Execution Envelope (TEE). Implementation is constrained to the TEE's declared scope.

Lane C​

Full evidence collection, dual-mode planning (breadth + depth), required human review gate before any write to the codebase. Used for refactors, cross-module changes, and anything the analyst flags as high-risk.


Classification is a floor, not a ceiling

lane = max(forced_min_from_eco, analyst_override). The ECO can force a task up to a higher lane; it can never force it down.