curl 'https://chatgpt.com/backend-api/codex/responses' \
-H 'Accept: text/event-stream' \
-H 'Version: 0.144.1' \
-H 'Originator: axonhub' \
-H 'Session-Id: e7dc302b-4dd3-4699-b6e4-68c09bcd1a51' \
-H 'Content-Type: application/json' \
-H 'Authorization: ******' \
-H 'X-Forwarded-By: lzc-ingress' \
-H 'Conversation_id: e7dc302b-4dd3-4699-b6e4-68c09bcd1a51' \
-H 'Chatgpt-Account-Id: 22cb1255-0fd9-42a1-af47-ad509654a2b6' \
-d '{"text":{"format":{"type":"json_object"}},"input":[{"role":"user","type":"message","content":[{"text":"Title: Unicode 令牌预算估算应采用 UTF-8 字节上界\n\nMemory:\n令牌预算估算不能对 Unicode 文本采用可能低估风险的简化 fallback，而应使用保守的 UTF-8 字节上界。非 ASCII 文本可能使简化估算显著低于真实 tokenizer 结果，进而导致上下文预算严重低估。修改相关逻辑时应覆盖所有令牌估算调用路径，并保留 Unicode 回归测试，以防止仅修复单一路径后其他运行时路径仍发生低估。","type":"input_text"}]}],"model":"gpt-5.4","store":false,"tools":[{"name":"classify_memory_unit_type","type":"function","parameters":{"type":"object","required":["unit_type","confidence","rationale"],"properties":{"rationale":{"type":"string","description":"Short reason grounded in the memory text."},"unit_type":{"enum":["fact","preference","decision","plan","procedure","learning","context","event"],"type":"string","description":"The single best unit_type."},"confidence":{"type":"number","description":"0..1 confidence in the classification."}}},"description":"Classify one durable memory into Nowledge Mem'\''s closed unit_type vocabulary."}],"stream":true,"include":["reasoning.encrypted_content"],"reasoning":{"summary":"auto"},"instructions":"Classify one durable memory into exactly one Nowledge Mem unit_type.\nAlways call classify_memory_unit_type.\nUse the closed vocabulary only:\n- fact: A stable objective statement, observation, or data point with no stronger behavioral, temporal, or procedural role.\n- preference: A user'\''s or owner'\''s preferred behavior, style, or constraint.\n- decision: A committed choice, tradeoff, or rejected alternative.\n- plan: An intended future action, roadmap item, or open next step.\n- procedure: A reusable how-to, runbook, command sequence, workflow, or operational rule for doing something again.\n- learning: A reusable lesson, gotcha, correction, or failure pattern that does not itself prescribe a repeatable workflow.\n- context: Background context, project state, identity, or scoping information.\n- event: A time-bound occurrence, meeting, milestone, incident, or deadline.\nImportant distinctions:\n- fact is for objective statements only; do not use fact as an '\''unknown'\'' bucket when another type fits.\n- preference is an always-on behavioral or style constraint.\n- procedure is a repeatable workflow, runbook, operational rule, command sequence, or how-to; it is evidence for skills but is not itself a skill.\n- learning is a lesson/gotcha/correction; if the memory includes the fix, prevention rule, or repeatable way to handle the gotcha, choose procedure.\n- decision is a committed choice; if the durable value is how to repeat the chosen process, choose procedure.\n- schema, API, storage, architecture, naming, or product-boundary choices are decision when the durable value is '\''use this design / do not use that design'\'', even if phrased imperatively.\n- do not classify a one-off design choice as procedure merely because it says '\''use X'\'' or '\''do not add Y'\''; choose procedure only when it gives a reusable execution workflow for future work.\n- if the memory describes how to handle future instances of a recurring situation (for example: for large imports, do not auto-run expensive work; show a plan; let the user schedule a bounded job), choose procedure.\n- decision records the chosen design; procedure tells an agent or system what to do next time that situation happens.\n- plan is future intent; if it is already a reusable way of working, choose procedure.\n- context is background state or identity when no stronger type applies.\n- crystal is not a unit_type; crystals still use one of the eight types.\nWhen uncertain, choose the closest valid type with low confidence; choose fact only for objective statements.\n\nReturn the result as one valid json object.","parallel_tool_calls":true}'
报错如下：
Response input messages must contain the word 'json' in some form to use 'text.format' of type 'json_object'.
请求如下：