ADR 0017: Per-call Responses routing

Date: 2026-09-20 Status: Accepted Supersedes: the superseded worker-era decisions (capsules, per-unit delegation, worker budgets, publication transactions) are retired; their machinery is not carried into V1

Context

The final technical review of the proposed V1 solution returned "approve with changes" on routing per model call inside one Codex session, with four contract tightenings: prove Codex per-call switching feasibility first; move completion verification outside the dynamic routing loop; select a valid (model, reasoning_effort) pair in a single Jev Choice; and separate GPT-6 "allowed to consider" from "must use".

A prior TaskUnit/worker design existed: capsules, capability catalogs of agents/Skills/MCPs, sixteen-check Policy Guard, isolated baselines, publication transactions, observation windows and evidence transitions. The review's per-call verdict retired that runtime: the unit of routing becomes the model call after a tool result, executed by the same session through a local Responses proxy — no workers, no capsules, no per-call isolation state machine. The granularity is implementable on the Responses request path, but task quality and real savings remain unproven in this product.

Decision

1. Routing unit: the model call. A local Responses proxy gates each call (router OFF, infrastructure bypass, privacy refusal), builds compact allowlisted routing state and validated candidate pairs, asks Jev one Choice over (model, effort) pairs, forwards natively with unchanged streaming events, and records actual pair and usage.
2. Fallback: low confidence, timeout, malformed or transport failure falls back to the configured baseline (default Terra/medium) with a recorded reason — fixed reliability handling, never a semantic second selector. A transport failure skips Jev for the remainder of the task. The kill switch restores the host's originally specified model and never downgrades user-chosen Sol/GPT-6.
3. Jev version: production pins a validated version; jev-latest runs in shadow only; requested/resolved versions are always recorded.
4. GPT-6: absent by default. One-shot eligibility from verified reasoning-blocker evidence; explicit user mandate is a separate hard constraint. Root takes over if Jev persistently avoids an evidenced-necessary GPT-6 call; Jev's answer is never rewritten.
5. Verification: independent, at the task boundary, outside the economic routing loop; acceptance read from the original request; fixed verification tier for semantic judgment; self-report is never evidence. Two correction cycles maximum, then Root takeover.
6. Observation: Router Compass records per-call facts and one task conclusion; UNKNOWN is never zero; savings claims follow the observed / replay / controlled-benchmark evidence ladder; replay is an offline script only.
7. P0 gate: before any routing logic is trusted, a real Codex A→B→A tool-loop proof must pass (auth, model/effort fidelity, tool-call IDs, SSE, cancellation, continuation/compaction). A critical failure stops per-call promotion and reopens the host interface question.

Consequences

• Five runtime modules (proxy, routing, verification, telemetry, entry) plus offline bench replace the worker-era modules; the SDD set is rewritten and task-capsule.md deleted.
• The old benchmark harness's UNKNOWN→zero conversion is a defect; its passing tests are not economic evidence.
• Cache economics are measured, not optimized; a switch threshold, if ever justified, arrives as a new policy version validated by the same ladder.
• Route-selection intelligence remains in exactly one place — Jev, through one adapter seam — with no heuristic classifier, task-kind table, tier ladder or second selector anywhere in the repository.