fix(codex): route Claude system prompts to Responses API instructions

[Shanlans]

Problem

When translating Claude requests to the Codex Responses API, the ChatGPT OAuth upstream rejects system/developer role messages inside the input array with:

System messages are not allowed

Two sources produced such messages:

1. The top-level Claude system prompt was emitted as a role:"developer" message inside input, while the top-level instructions field was hard-coded to "".
2. Any role:"system" message injected into the conversation (e.g. by Claude Code Agent Teams, CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, which adds an "Available agent types…" system message) was passed through verbatim into input.

Both caused upstream 400 System messages are not allowed.

Fix

internal/translator/codex/claude/codex_claude_request.go:

• Fold the top-level system field (array or string) into the top-level instructions string.
• During message conversion, fold any system/developer role message's text into instructions as well, instead of emitting it into input.
• Never emit system/developer roles into input.

Result envelope now looks like:

{ "model": "...", "instructions": "<system prompt>", "input": [ { "role": "user", ... } ] }

Tests

Added to codex_claude_request_test.go:

• TestConvertClaudeRequestToCodex_SystemBecomesInstructions — top-level system → instructions, no system/developer in input.
• TestConvertClaudeRequestToCodex_InConversationSystemMessageFolded — in-conversation role:"system" message folded into instructions, absent from input.

go test ./internal/translator/codex/claude passes.

Verified end-to-end against a live Codex (gpt-5.6-sol) route with Agent Teams enabled: previously 400 System messages are not allowed, now 200.

🤖 Generated with Claude Code