Problem
Original text(PsychQuant/codex-pro#1 Diagnosis 的 Residue 宣告,close-time 裁決 file 為 follow-up):
「"100% deterministic headings" is unreachable without upstream structured-output / JSON-schema mode in codex-call (a parallel-ai-agents cross-repo concern, out of scope here). The honest scope is "reduce failure rate enough that a hard e2e assertion is stable", not "eliminate the failure mode."」
codex-pro 的 producer skills(review / adversarial-review)靠 prompt-level literal-token instructions 讓 Codex 輸出必要的 H2/H3 heading 結構。v0.4 hardening + 12-combo e2e matrix 已把觀測失敗率壓到 0(heading 斷言已 promote 為 hard),但這本質上是 probabilistic — LLM 不是 schema-validated endpoint。
Type
feature
Expected
bin/codex-call 評估支援 structured-output / JSON-schema mode(backend-api 若有對應參數),讓 downstream(codex-pro 等 consumer)能以 schema 強制輸出結構、把 heading contract 從 prompt-level 承諾升級為 API-level 保證
Actual
- codex-call 僅
--instructions + --prompt-file 純文字介面;輸出結構全靠 prompt 紀律
Impact
- codex-pro 的 result-file heading contract(
## Summary/## Findings、adversarial 4-H2)永遠殘留 probabilistic 失敗尾巴;model 換代(如 gpt-5.5 → gpt-5.6-sol)時需重新實測
Source: residue from PsychQuant/codex-pro#1 at /idd-close time (Step 3.6)
Problem
codex-pro 的 producer skills(review / adversarial-review)靠 prompt-level literal-token instructions 讓 Codex 輸出必要的 H2/H3 heading 結構。v0.4 hardening + 12-combo e2e matrix 已把觀測失敗率壓到 0(heading 斷言已 promote 為 hard),但這本質上是 probabilistic — LLM 不是 schema-validated endpoint。
Type
feature
Expected
bin/codex-call評估支援 structured-output / JSON-schema mode(backend-api 若有對應參數),讓 downstream(codex-pro 等 consumer)能以 schema 強制輸出結構、把 heading contract 從 prompt-level 承諾升級為 API-level 保證Actual
--instructions+--prompt-file純文字介面;輸出結構全靠 prompt 紀律Impact
## Summary/## Findings、adversarial 4-H2)永遠殘留 probabilistic 失敗尾巴;model 換代(如 gpt-5.5 → gpt-5.6-sol)時需重新實測