Add image generation, and fix non-streaming /v1/responses returning empty output

[mazxxy]

This started as a fork I needed for myself: I wanted image generation out of a Codex account, and there was no way to get one out of ChatMock. While building it I ran into a bug that is worth fixing on its own, so I'm sending both here instead of keeping it private.

The bug underneath

response.completed from the codex backend always carries "output": []. The items only ever exist in the response.output_item.done events. Since aggregate_response_from_sse() returns the object from the final event, every POST /v1/responses with stream: false answers with an empty output — plain text included, not just images. It looks like a success: status: "completed", usage filled in, no content.

On current main:

curl http://127.0.0.1:8000/v1/responses -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.4-mini","input":"say ok","stream":false}'
# "output": []

Items are now collected by output_index and rebuilt into the final object when it comes back empty. Streaming was never affected.

Image generation

The endpoint already accepts the image_generation tool; only ChatMock had no way to receive the result. What the backend actually does, measured rather than assumed:

• the drawing model is gpt-image-2-codex, chosen server-side
• tool_choice: {"type": "image_generation"} works and forces the call
• consumption shows up in tool_usage.image_gen, so it comes off the ChatGPT plan quota
• n inside the tool is refused (Unknown parameter: 'tools[0].n'), so n > 1 repeats the request, capped at 4 and stated in the error message
• size and quality are accepted but echoed back as "auto"; the resolution follows the prompt, so the requested size is written into the instructions as well. Asking for 1536x1024 landed exactly; asking for 1024x1024 returned 1254x1254

What that turns into:

• /v1/images/generations in the OpenAI Images shape, so existing clients work unchanged. response_format: "url" is rejected, mirroring gpt-image-1
• image_generation allowed in responses_tools on /v1/chat/completions and /api/chat, with the image embedded as a data URL in the message content
• --image-model / CHATGPT_LOCAL_IMAGE_MODEL for the model that orchestrates the call
• 9 tests added, 27 passing

One detail worth calling out: in streaming chat the image is emitted only after the think tag is closed. Without that it lands inside <think> and disappears in any client that collapses reasoning, which is the whole point of that mode. The close-tag block was duplicated twice in sse_translate_chat(); it is now a single _close_think_tag() helper used in all three places.

How to try locally

python chatmock.py serve --port 8000

# the fix: this returned an empty output before
curl http://127.0.0.1:8000/v1/responses -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.4-mini","input":"say ok","stream":false}'

# image, OpenAI Images shape
curl http://127.0.0.1:8000/v1/images/generations -H "Content-Type: application/json" \
  -d '{"prompt":"a minimalist geometric fox logo","size":"1024x1024"}'

# image inside a chat request
curl http://127.0.0.1:8000/v1/chat/completions -H "Content-Type: application/json" \
  -d '{"model":"gpt-5.4-mini","messages":[{"role":"user","content":"draw a blue cube"}],"responses_tools":[{"type":"image_generation"}]}'

python -m pytest tests -q

Every path above was run against the live backend, streaming and non-streaming, on both the OpenAI and Ollama routes.

Checklist notes

• Rebased on latest main. No issue reference — opening this cold; happy to file one first if you prefer that order.
• README.md and DOCKER.md updated (new flag, new env var, usage section).
• Defaults unchanged: the image path only activates when a request asks for the tool. The aggregator fix does change existing behaviour, but only from "empty" to "the content that was already paid for".
• Ollama routes updated alongside the OpenAI ones, per CONTRIBUTING.
• No entry points moved and no existing parameter names or payload shapes changed.

Disclosure

AI was used to write this patch (Claude Code). Everything it claims was verified against the live backend and the test suite rather than taken on faith, but review it as you would any patch from a stranger. Happy to rework anything, split the aggregate_response_from_sse fix into its own PR since it stands alone, or close this if it's not a direction you want the project to go.