[payne0420]

Adds end-to-end Codex image generation support — both the chat-flow hosted tool (what Codex CLI uses) and the legacy /v1/images/generations endpoint (what OpenWebUI and similar clients use).

Problem

Codex CLI ships with the default-enabled feature flag:

[features]
image_generation = true

When enabled, Codex CLI injects an OpenAI Responses-API hosted tool into every POST /v1/responses request:

{ "type": "image_generation", "output_format": "png" }

OpenAI's Responses API then executes gpt-image-1 server-side and streams back response.output_item.done events where item.type === "image_generation_call" carrying the base64 PNG in item.result. Codex CLI saves that PNG to $CODEX_HOME/generated_images/<session>/<call_id>.png and renders a TUI artifact panel.

Through OmniRoute this feature silently no-oped. Two distinct breakages:

1. /v1/responses path — normalizeCodexTools() in open-sse/executors/codex.ts dropped every tool whose type !== "function". The hosted image_generation tool was stripped before the request reached OpenAI, so upstream never produced an image_generation_call output item. The same break affected other hosted tools (web_search, file_search, …).
2. /v1/images/generations path — OmniRoute's image handler had no Codex branch, so a DALL-E-style request with model: codex/gpt-5.4 fell through to a generic OpenAI-compatible code path that built a bogus URL (chatgpt.com/backend-api/codex/responses/images/generations), which Cloudflare 403'd.
3. store: true default — once the hosted tool was preserved, the Codex backend rejected requests with "Store must be set to false" because OmniRoute's executor defaults store to true for the OpenClaw chaining scenario.

Fix

Two commits on this branch:

fix(sse): preserve Responses API hosted tools in Codex executor

• Add a whitelist of Responses-API hosted tool types (image_generation, web_search, web_search_preview, file_search, computer, computer_use_preview, code_interpreter, mcp, local_shell).
• Teach normalizeCodexTools to preserve hosted tools (shape: type set, no .function, no .name, type in whitelist) alongside the existing function-tool handling. Unknown non-function types are still dropped locally with a console.debug.
• tool_choice handling is unchanged — it only inspects type === "function", so hosted tool_choice values pass through.

feat(sse): route DALL-E-style image generation through Codex hosted tool

• Register codex (alias cx) in the image provider registry with format: "codex-responses".
• Add handleCodexImageGeneration to the image handler: builds a /v1/responses body with tools: [{ type: "image_generation", output_format: "png" }], POSTs to the Codex endpoint with proper ChatGPT-OAuth headers (chatgpt-account-id, session_id, User-Agent), reads the SSE stream, extracts image_generation_call items, and returns OpenAI image response shape ({ data: [{ b64_json }] } or { data: [{ url: "data:image/png;base64,..." }] } depending on response_format).
• Export a shared extractImageGenerationCalls() helper for testability.
• Default store: false in the executor when a request contains an image_generation tool — avoids the "Store must be set to false" upstream 400 for the chat-flow path. Explicit client store values are respected.

Scope

• Only open-sse/executors/codex.ts, open-sse/config/imageRegistry.ts, and open-sse/handlers/imageGeneration.ts are modified on the production side. Other executors and providers are untouched.
• No new settings, feature flags, or DB columns.
• n > 1 for the DALL-E-style endpoint is handled sequentially (the hosted tool generates one image per call); clients that need parallel generation should use multiple requests.

Test plan

• node --import tsx/esm --test tests/unit/executor-codex.test.ts — 19/19 pass (7 new cases covering hosted-tool preservation, tool_choice regression, and the image-gen store default).
• node --import tsx/esm --test tests/unit/image-generation-handler.test.ts — 36/36 pass (5 new cases covering extractImageGenerationCalls, /responses round-trip, b64_json vs url response formats, upstream error propagation, and the no-image-emitted error path).
• node --import tsx/esm --test tests/unit/tools-filter-anthropic-format.test.ts tests/unit/codex-stream-false.test.ts tests/unit/t19-codex-responses-empty-content.test.ts — 17/17 pass, no adjacent regressions.
• npm run typecheck:core — clean.
• ESLint on changed files — clean.
• Manual end-to-end smoke test (chat flow): POST /v1/responses with tools: [{ "type": "image_generation", "output_format": "png" }] and "store": false against a Codex ChatGPT-OAuth connection — SSE stream contains response.output_item.done with item.type === "image_generation_call" and a non-empty base64 item.result. Verified in user testing.
• Manual end-to-end smoke test (DALL-E shim): POST /v1/images/generations with { "model": "codex/gpt-5.4", "prompt": "...", "response_format": "b64_json" } against the same connection — returns { data: [{ b64_json: "..." }] }.
• tests/manual/image-generation.http has both request blocks appended.

[gemini-code-assist]

Code Review

This pull request introduces support for OpenAI-hosted tool types, such as image_generation, web_search, and file_search, within the Codex executor. It adds a set of recognized hosted tool types and updates the normalizeCodexTools function to correctly identify and preserve these tools while filtering out malformed or unknown ones. Comprehensive unit tests and a manual test case have been included to verify the transformation logic and ensure proper handling of tool choices. I have no feedback to provide.

[payne0420]

This fixes stuff like Open WebUI image generation:

[diegosouzapw]

Awesome feature! 🚀 This has been successfully merged into release/v3.7.0. Supporting Codex image generation and DALL-E shimming is a huge plus. Thanks for the contribution!

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 3449aa1daf

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Clamp Codex image n before sequential upstream loop

This handler uses body.n directly to control a per-image for loop, issuing one upstream /responses request per iteration. Because the request schema does not enforce an upper bound for n, a client can send a very large value and trigger excessive sequential upstream calls in a single API request, causing runaway latency/cost and tying up server capacity. Add a strict max (or reject oversized n) before entering the loop.

Useful? React with 👍 / 👎.

[chatgpt-codex-connector]

Compute store default after Codex tool normalization

store is defaulted from hasImageGenerationTool(body) before normalizeCodexTools(body) removes invalid hosted tools. If an image_generation entry is later filtered out (for example because it also has a name or function field), store remains forced to false even though no image tool is sent upstream, which can unexpectedly disable response persistence and break previous_response_id chaining for the remaining valid tools. Derive this default from the normalized tool list instead.

Useful? React with 👍 / 👎.