OpenCode runtime: openai/gpt-6-luna completes cooks without the review_form, so green candidates exhaust their budget and never open a PR #2951

What happened (2026-10-14, homeboy-lab, controller v0.422.1)

Two cooks on --backend opencode --model openai/gpt-6-luna produced correct patches and passed their deterministic gates, but every attempt ended with:

• OpenCode completed without required structured output(s): review_form.
• the only missing required output was review_form; deferring it to the cook loop's form-only follow-up
• final: cook finished execution_budget_exhausted after the run was recorded pass, status candidate_recoverable

Runs: agent-task-0c9edae1-2113-49c3-a80f-bffc0780a6e2 (Extra-Chill/extrachill-cli#157) and agent-task-734c656f-b3ff-434d-b008-c47a22b58704 (Extra-Chill/extrachill-analytics#305). Both used all 3 attempts, and the form-only follow-up never got the form either.

Workaround: homeboy agent-task finalize-pr --recover <attempt> --review-form @form.json --review-form-tool … --review-form-model … --review-form-author … with a form written by the operator agent (opened Extra-Chill/extrachill-analytics#307).

Expected

The OpenCode runtime gets the review_form out of this model, either through its structured-output path or a form-only follow-up that the model actually answers. A green candidate with a missing form shouldn't burn the whole provider budget and end as recoverable.

Notes

The same cooks on Pi never got this far (see Extra-Chill/homeboy#15891), so it's unclear whether other OpenCode models have the same problem. A good first check: does the form-only follow-up prompt reach the model, and what does the model reply?