2.9 KiB
Owner-authorized model routing
Claude commander and independent reviewer: claude-opus-5-5.
Codex default: gpt-6.1-sol. Alternative: gpt-6-astra, only when Claude Opus 5.5 selects it for the task and records a reason.
Use the persisted owner_model_cli.py coordinator entry point for commander and Codex dispatch/run, with OWNER_RUNNER_CONTROL pointing to the reviewed Runner checkout. It preserves ordinary Runner decisions, leases, contracts, sandbox, counters, acceptance and review gates. It is not a change to worker source or registered project digests.
The commander result adds codex_model and codex_model_reason. For a Codex run, the latest recorded commander proposal must be for the same task, have action run, and select an allowed model with a nonempty reason. Pass that exact recorded decision to ordinary dispatch. A model-selection-only wait decision never starts an implementation.
Example (replace PROJECT, TASK, CONTROL with actual registered values):
OWNER_RUNNER_CONTROL=CONTROL python3 -B owner_model_cli.py --project PROJECT commander TASK --request 'Choose next action and Codex model under owner policy'
OWNER_RUNNER_CONTROL=CONTROL python3 -B owner_model_cli.py --project PROJECT dispatch recorded-decision.json
New-project defaults are also set in the Runner policy. Existing frozen project configurations are not rewritten; the explicit adapter applies owner-authorized model selection at invocation. Bare legacy CLI/MCP dispatch does not provide this routing; use this entry point for future Codex invocations. The P24 GUI continuation was observed using Claude Opus 5.5 Medium; do not claim the old Gemini selection remains. CLI adapter authentication compatibility is separate from this actual GUI execution.
Configuration and selected argv are distinct from actual inference evidence. Confirm the worker execution artifact after its next authorized run; no implementation is launched solely to test this setting.
The same validated recorded run decision is also delivered in full to the Codex implementer prompt. Its concrete execution direction is subordinate to the immutable original spec, allowed paths and acceptance. Model choice alone is insufficient for task handoff. Check the actual run packet for the exact recorded direction before relying on delivery. This corrects the original model-only adapter; it does not change Runner permissions, registration, counters or acceptance.
The P18-only allocation-placeholder resolution and same-source infrastructure retry are exact task/base/spec/failed-receipt guarded, retain counters and call ordinary real validation. They do not accept evidence or override reviews. p18_review_context.py contains the bounded immutable Git source context; full Runtime unchanged-method review is not claimed. Independent review requested additional legacy precedence/lifecycle context, which remains explicit.