v1.13.0

Added

• otito route <repo> "<request>" — the model router, promoted out of scripts/ into the CLI. Score a coding task before any tokens are spent on it and recommend a cheap, mid, or premium tier. otito answers the repository half deterministically (AX, containment, canonical risk flags); a System One model answers the request half with calibrated probabilities over three narrow typed questions, in one call. --tier-only prints the tier for any host to map, --host <id> prints that host's model name, and maps live in .otito/model-route.json, so nothing about the scoring is specific to one IDE. scripts/model-route.mjs remains as a thin wrapper for the 1.12.0 prototype invocation.
  • The router is not the gate and must never become one. It runs before work starts and decides how much model to spend; the gate runs after the diff exists and decides whether a change may merge. A failed, unreachable, or unkeyed model call falls back to a local estimate that labels itself uncalibrated, and costs a tier, never a verdict. The gate never asks.
  • It stays advisory. The share weights and band thresholds were chosen by judgement and have never been compared to an outcome — the same critique the calibration thesis makes of inferRisk. Promoting it past advisory needs a backtest measuring regret: tasks routed cheap that ended in a revert or a repair, weighed against the spend avoided.
• An advisory routing footer under commands that already hold an impact pass and an AX score. Offline by construction — an ordinary otito command still makes no network call.

Changed

• Nineteen report commands now get otito's shared document treatment. Fourteen commands honoured colour and theme while nineteen printed raw Markdown and looked nothing like review or impact. renderDocument gives them all the same header box, headings that read as headings, and quiet rules, rather than a hand-written renderer per command.

Fixed

• The router read an empty match as a contained change. Zero candidates is absence, not containment: when otito matches nothing, AX describes an empty set and containment reads high for the same reason there is nothing to spread across, so the arithmetic produced a confident-looking cheap tier for exactly the request the repository understood least. Found by dogfooding: "fix scanning multi date event ticket", asked against a repository with no such code, scored containment 100 on zero files and routed cheap. The floor jumps straight to the ceiling rather than stepping one tier — a one-tier step would leave a high-AX empty read at mid, the same mistake one notch quieter. scoring.evidence carries the candidate count and a sufficient flag so a caller can say "no recommendation" instead of printing a tier that rests on nothing.
• The offline router escalated on evidence it never measured, in two ways, both through a fail-safe bump rather than the score. Fixture corpora no longer carry a risk flag: otito has no checkout of its own, so add refund handling to checkout ranked an eval fixture first and escalated to premium on money-flow evidence from a corpus that ships nothing. Test data still counts toward reach — it is a real file the change touches — but risk has to be about shipped code. And the offline estimator now reports no confidence rather than the peak of its own distribution: distribute(0.6), the baseline for any request the keyword lists do not recognise, peaks at 0.40, under the 0.55 floor, so every unrecognised request was bumped a tier and the offline default was "spend more". A measured confidence below the floor still escalates.
• A test that only passed on a machine without TYPESAFE_API_KEY. askJev falls back to the environment, so the missing-key assertion was handing a real key to its own stub and asserting the wrong error. It now controls the variable instead of depending on it.