fix-native-codex-lightweight-evaluator

Fix: native Codex lightweight evaluator dispatch

Metadata

Statusdone
Assignedagent-906
Modelcodex:gpt-5.6-sol
Created2026-07-26T22:30:40.242455833+00:00
Started2026-07-26T22:45:19.016960554+00:00
Completed2026-07-27T00:05:00.664901844+00:00

Description

Description

The fresh native Codex 5.6 canary resolved evaluator to codex:gpt-5.6-luna/high, but wg evaluate run failed before spawning Codex with error[WG-PI-ROUTE-REQUIRED]. The configured direct Codex starter promises native Luna evaluation, yet the lightweight evaluator dispatch path accepts only Pi. Implement native Codex CLI one-shot dispatch for explicit evaluator/agency codex: routes, preserving exact model and structured reasoning, with no Pi/Claude fallback.

File scope: src/service/llm.rs and focused evaluator/dispatch tests only. Implement directly — do not decompose further.

Validation

  • Failing test first proves wg evaluate run with evaluator route codex:gpt-5.6-luna reaches native Codex rather than WG-PI-ROUTE-REQUIRED
  • Codex receives native model gpt-5.6-luna and model_reasoning_effort="high"
  • Pi and Claude sentinels remain uninvoked; native Codex failure does not cross runtimes
  • Existing Pi evaluator behavior remains unchanged
  • cargo fmt --check passes
  • cargo clippy passes
  • cargo build and cargo test pass
  • cargo install --path . --locked updates the installed WG binary for the dependent canary

Depends on

Required by

Log