Summary

• Problem: openai-codex models using openai-codex-responses could build a request with input: [] when the context only had a systemPrompt.
• Why it matters: ChatGPT's Codex backend rejects that shape with a 400 before the stream can complete.
• What changed: Codex Responses payloads now keep system prompt text in top-level instructions and add a minimal sentinel user input item when there are no converted messages.
• What did NOT change (scope boundary): Other OpenAI Responses providers, auth, model catalog, and tool conversion behavior are unchanged.

Change Type (select all)

• Bug fix
• Feature
• Refactor required for the fix
• Docs
• Security hardening
• Chore/infra

Scope (select all touched areas)

• Gateway / orchestration
• Skills / tool execution
• Auth / tokens
• Memory / storage
• Integrations
• API / contracts
• UI / DX
• CI/CD / infra

Linked Issue/PR

• Fixes openai-codex provider sends empty input[] when context has only systemPrompt → ChatGPT 400 #73820
• This PR fixes a bug or regression

Root Cause (if applicable)

• Root cause: Codex Responses moved the system prompt to top-level instructions, but still sent the converted message array as input; with no user/assistant/tool history, that array was empty.
• Missing detection / guardrail: No focused test covered the openai-codex-responses system-prompt-only payload shape.
• Contributing context (if known): The Codex backend requires one of input, previous_response_id, prompt, or conversation_id; an empty input array is treated as missing.

Regression Test Plan (if applicable)

• Coverage level that should have caught this:
  • Unit test
  • Seam / integration test
  • End-to-end test
  • Existing coverage already sufficient
• Target test or file: src/agents/openai-transport-stream.test.ts
• Scenario the test should lock in: openai-codex-responses with systemPrompt and messages: [] produces non-empty sentinel input while preserving top-level instructions.
• Why this is the smallest reliable guardrail: The bug is in buildOpenAIResponsesParams, so direct builder coverage catches the rejected payload shape without live Codex calls.
• Existing test that already covers this (if any): N/A
• If no new test is added, why not: N/A

User-visible / Behavior Changes

OpenAI Codex provider turns that only have system-prompt-backed instructions no longer send input: [] to the ChatGPT Codex backend.

Diagram (if applicable)

Before:
systemPrompt only -> instructions set + input: [] -> backend 400

After:
systemPrompt only -> instructions set + one sentinel input_text item -> request is non-empty

Security Impact (required)

• New permissions/capabilities? (Yes/No): No
• Secrets/tokens handling changed? (Yes/No): No
• New/changed network calls? (Yes/No): No
• Command/tool execution surface changed? (Yes/No): No
• Data access scope changed? (Yes/No): No
• If any Yes, explain risk + mitigation: N/A

Repro + Verification

Environment

• OS: macOS
• Runtime/container: Node v24.14.0, pnpm 10.33.0
• Model/provider: openai-codex / openai-codex-responses
• Integration/channel (if any): N/A
• Relevant config (redacted): N/A

Steps

1. Build OpenAI Responses params for an openai-codex model with API openai-codex-responses.
2. Pass a context with systemPrompt set and messages: [].
3. Inspect the request payload.

Expected

• instructions contains the sanitized system prompt.
• input is non-empty.

Actual

• Before this fix, input was [].

Evidence

• Failing test/log before + passing after
• Trace/log snippets
• Screenshot/recording
• Perf numbers (if relevant)

Human Verification (required)

What you personally verified (not just CI), and how:

• Verified scenarios: focused unit coverage for system-prompt-only Codex Responses payloads; focused transport test suite; changed-file format; changed-lane gate.
• Edge cases checked: system prompt cache boundary stripping is preserved in instructions; fallback input stays a small sentinel; non-Codex providers are not routed through the new fallback.
• What you did not verify: no live ChatGPT/Codex backend call was made.

Review Conversations

• I replied to or resolved every bot review conversation I addressed in this PR.
• I left unresolved only the conversations that still need reviewer or maintainer judgment.

If a bot review conversation is addressed by this PR, resolve that conversation yourself. Do not leave bot review conversation cleanup for maintainers.

Compatibility / Migration

• Backward compatible? (Yes/No): Yes
• Config/env changes? (Yes/No): No
• Migration needed? (Yes/No): No
• If yes, exact upgrade steps: N/A

Risks and Mitigations

• Risk: The sentinel input adds a tiny user item to otherwise empty Codex requests.
  • Mitigation: The behavior is limited to openai-codex + openai-codex-responses and only when the converted input would otherwise be empty; the system prompt remains in instructions.

AI-Assisted Disclosure

• AI-assisted: Yes
• Testing degree: Focused local validation plus pnpm check:changed.
• I confirm I reviewed the diff and understand the code change.