Summary

openai-codex image generation against gpt-image-2-medium currently fails in two stages against the Codex backend:

1. Bundled/original plugin request shape fails with:

   • 400 invalid_request_error
   • Tool choice 'image_generation' not found in 'tools' parameter.

2. After removing the tool_choice block from plugins/image_gen/openai-codex/__init__.py, the request no longer 400s, but the backend still never emits any image-generation events/items. Hermes then fails with:

   • Codex response contained no image_generation_call result
   • error_type: empty_response

This appears to be a compatibility gap between the current openai-codex image-gen plugin request shape and the live Codex backend behavior, rather than a local config/cache issue.

Environment

• Repo: NousResearch/hermes-agent
• Checkout tested: 167b5648ea609aafa85f56c5714f7abda5091ed6
• git describe: v2026.4.30-227-g167b5648e-dirty
• Runtime: editable checkout under ~/.hermes/hermes-agent
• Profile: bot2
• Provider: openai-codex
• Image model/tier returned by Hermes: gpt-image-2-medium
• Base URL: https://chatgpt.com/backend-api/codex
• Auth type: Codex OAuth credential in profile auth.json
• image_gen.provider: openai-codex
• image_gen.use_gateway: false
• credential_pool_strategies.openai-codex: fill_first
• Credential pool simplified to a single usable openai-codex OAuth entry

What was ruled out

We explicitly ruled out the most likely local-environment causes:

• Gateway restart/cache issue
  • Reproduced before and after gateway restart
• Dirty working tree affecting behavior
  • Reproduced from a clean detached worktree at the same commit
• Credential-pool rotation confusion
  • Pool reduced to a single openai-codex OAuth credential
• Missing auth
  • Runtime selects a valid Codex OAuth token and reaches the backend
• Config mismatch around provider selection
  • Reproduced with direct plugin invocation, outside Telegram/gateway flow

Reproduction (original bundled plugin)

The bundled plugin currently sends a request shaped like this in plugins/image_gen/openai-codex/__init__.py:

with client.responses.stream(
    model=_CODEX_CHAT_MODEL,
    store=False,
    instructions=_CODEX_INSTRUCTIONS,
    input=[{
        "type": "message",
        "role": "user",
        "content": [{"type": "input_text", "text": prompt}],
    }],
    tools=[{
        "type": "image_generation",
        "model": API_MODEL,
        "size": size,
        "quality": quality,
        "output_format": "png",
        "background": "opaque",
        "partial_images": 1,
    }],
    tool_choice={
        "type": "allowed_tools",
        "mode": "required",
        "tools": [{"type": "image_generation"}],
    },
)

Actual result

{
  "success": false,
  "error_type": "api_error",
  "error": "OpenAI image generation via Codex auth failed: Error code: 400 - {'error': {'message': \"Tool choice 'image_generation' not found in 'tools' parameter.\", 'type': 'invalid_request_error', 'param': 'tool_choice', 'code': None}}",
  "provider": "openai-codex",
  "model": "gpt-image-2-medium"
}

Reproduction (minimal patch: remove tool_choice)

After removing only this block:

tool_choice={
    "type": "allowed_tools",
    "mode": "required",
    "tools": [{"type": "image_generation"}],
},

The same request no longer 400s, but Hermes still fails because no image tool call/result ever arrives.

Actual result

{
  "success": false,
  "error_type": "empty_response",
  "error": "Codex response contained no image_generation_call result",
  "provider": "openai-codex",
  "model": "gpt-image-2-medium"
}

Stream/event inspection

With tool_choice removed, direct stream inspection shows the backend completing the request without any image_generation events at all.

Observed event pattern:

• response.created
• response.in_progress
• response.output_item.added (message)
• response.content_part.added
• response.output_text.delta / response.output_text.done (sometimes empty)
• response.output_item.done (message)
• a second message output item
• response.completed

Not observed:

• response.image_generation_call.partial_image
• response.output_item.done with item.type == "image_generation_call"
• any final response.output items

Final response summary observed:

{
  "status": "completed",
  "output_text": "",
  "output_count": 0,
  "output": []
}

Additional variants tested

• tool_choice='auto'
  • No 400, but still no image-generation events/items
• partial_images=0
  • No 400, but still no image-generation events/items
• Omitting instructions
  • Backend returns 400 with Instructions are required

Expected behavior

One of the following should work reliably for the openai-codex image-gen plugin:

1. The bundled/original request shape should be accepted by the live Codex backend and return an image_generation_call item/result, or
2. The plugin should be updated to the currently supported request shape for Codex image generation so that Hermes can actually receive image events/results.

Actual behavior

• Original bundled request shape: fails at request validation (tool_choice 400)
• Minimal patched request shape: request completes, but the backend never invokes/returns image generation output

Relevant files

• plugins/image_gen/openai-codex/__init__.py
• tests/plugins/image_gen/test_openai_codex_provider.py
• potentially any Codex Responses compatibility layer assumptions around tool invocation

Suggested maintainer follow-up

• Verify the currently supported Codex Responses request shape for image generation against https://chatgpt.com/backend-api/codex
• Re-check whether tool_choice is still supported in this path
• If tool_choice is unsupported, update provider tests so they match the live backend contract
• If the backend now requires a different image tool invocation shape, update the plugin accordingly

Notes

A likely related but not identical issue is #5736 (openai-codex empty response.output on gpt-5.x agent-loop requests). This report is specifically about the bundled openai-codex image generation plugin path.