Feature/new models

[lchoquel]

Summary by cubic

Adds gpt-5.5 and gpt-image-2, moves image generation to a rule‑driven pipeline with parameter checks, and fixes model routing across openai/huggingface/fal. Also declares PDF input for GPT‑5.4/5.5 on azure_openai and the gateway. Error messages now include the model_name for clarity.

• New Features

  • Added gpt-5.5 on openai, azure_openai, and the gateway (text/images/pdf).
  • Added gpt-image-2 with per‑model rules (no transparent background); updated alias best-gpt → gpt-image-2.
  • Introduced ImgGenParamSupport and PipeImgGen static validation; added e2e/integration/unit tests.
  • Declared pdf input for GPT‑5.4 series and gpt-5.5 on azure_openai and the gateway.

• Refactors

  • Unified image‑gen arg building in ImgGenArgsFactory; OpenAI worker now rule‑driven with defaults quality=medium and output_compression=100; positive size validation for OpenAI img‑gen.
  • Renamed taxonomy MODEL_NAME → MODEL_CHOICE (supports model_id/model_name); standardized taxonomies (gpt_image vs gpt_image_legacy); added InputFidelity and OutputCompression; threaded model_name through all workers and errors, and now forward it to the arg factory.
  • Enforced model_choice = "model_id" on openai/huggingface/fal rules; workers require/pass the correct model (prevents OpenAI fallback to DALL‑E 2).
  • Improved Azure img‑gen reliability and observability (token usage capture, 600s timeout); refreshed gateway/docs and bumped remote config URL.
  • Removed TODOS.md.

^{Written for commit f93bed7. Summary will update on new commits. Review in cubic}

[lchoquel]

@cubic-dev-ai

[cubic-dev-ai]

@cubic-dev-ai

@lchoquel I have started the AI code review. It will take a few minutes to complete.

[greptile-apps]

Greptile Summary

• Adds gpt-image-2 (OpenAI & Azure), gpt-5.5 LLM, and refactors the OpenAI image worker to a taxonomy-driven ImgGenArgsFactory pattern that also supports image-to-image via images.edit.
• Introduces ImgGenParamSupport for static blueprint validation and InputFidelity / OutputCompression taxonomy types.
• P1: All five OpenAI image model rule blocks in both openai.toml and azure_openai.toml are missing model_choice = \"model_id\". Because openai_img_gen_worker.py now delegates entirely to ImgGenArgsFactory, the model key is never inserted into args_dict, so every images.generate() / images.edit() call is sent without a model parameter.

Confidence Score: 3/5

Not safe to merge — all OpenAI image API calls will use the wrong default model or fail outright until model_choice is added to TOML rules.

A single P1 bug (missing model_choice in every OpenAI image model's TOML rules) means the model parameter is never passed to the OpenAI SDK, breaking all image generation and editing calls for this PR's new models. The rest of the refactor is well-structured.

pipelex/kit/configs/inference/backends/openai.toml and pipelex/kit/configs/inference/backends/azure_openai.toml — every [*.rules] block for image generation models needs model_choice = "model_id".

Important Files Changed

Filename	Overview
pipelex/kit/configs/inference/backends/openai.toml	Adds rules blocks for all 5 OpenAI image models but all rules are missing model_choice = "model_id", so model is never passed to the OpenAI SDK (P1 bug).
pipelex/kit/configs/inference/backends/azure_openai.toml	Adds gpt-image-2 and updates existing image model rules; same missing model_choice issue as openai.toml. Also adds gpt-5.5 LLM and pdf input for gpt-5.4 series.
pipelex/cogt/img_gen/img_gen_args_factory.py	New taxonomy-driven factory; passes model_id instead of model_name to error-message params in three places (P2). Logic is otherwise sound.
pipelex/plugins/openai/openai_img_gen_worker.py	Refactored to use ImgGenArgsFactory and adds images.edit path for image-to-image; size fallback parsing and format detection look correct. Relies on rules having model_choice (which is missing from TOML configs).
pipelex/plugins/openai/openai_img_gen_factory.py	Good refactor adding gpt-image-2 size validation; potential ZeroDivisionError if min_edge is 0 (P2). LEGACY_SIZE_TO_DIMENSIONS class var dict comprehension is correct.
pipelex/cogt/img_gen/img_gen_param_support.py	New module providing yes/no parameter support checks; clean implementation with proper abstain-on-unknown-taxonomy behavior.
pipelex/cogt/img_gen/img_gen_model_rules.py	Adds InputFidelityTaxonomy, OutputCompressionTaxonomy, ModelChoiceTaxonomy; renames/expands existing taxonomies for GPT Image 2. Clean additions.
pipelex/pipe_operators/img_gen/pipe_img_gen.py	Adds static blueprint validation against model rules; correctly skips unresolvable presets and models without rules.
pipelex/plugins/azure_rest/azure_img_gen_worker.py	Adds usage token extraction and increases timeout to 600s for gpt-image-2 generation time. Minor: model_name added to factory call.

Sequence Diagram

sequenceDiagram
    participant PipeImgGen
    participant ImgGenParamSupport
    participant OpenAIImgGenWorker
    participant ImgGenArgsFactory
    participant OpenAIAPI

    PipeImgGen->>ImgGenParamSupport: check_blueprint_params(rules, aspect_ratio, background, output_format)
    ImgGenParamSupport-->>PipeImgGen: unsupported reasons (or empty)
    PipeImgGen->>OpenAIImgGenWorker: _gen_image_list(img_gen_job, nb_images)
    OpenAIImgGenWorker->>ImgGenArgsFactory: make_args_for_model(rules, job, nb_images, model_id, model_name)
    ImgGenArgsFactory-->>OpenAIImgGenWorker: args_dict (prompt, size, quality, n, ...)
    alt has input images
        OpenAIImgGenWorker->>OpenAIAPI: images.edit(**args_dict)
    else text-to-image
        OpenAIImgGenWorker->>OpenAIAPI: images.generate(**args_dict)
    end
    OpenAIAPI-->>OpenAIImgGenWorker: ImagesResponse (b64_json, size, output_format)
    OpenAIImgGenWorker-->>PipeImgGen: list[GeneratedImageRawDetails]

Loading

Comments Outside Diff (1)

1. pipelex/cogt/img_gen/img_gen_args_factory.py, line 95-113 (link)

   model_id passed where model_name expected

   Lines 102, 111, and 178 all call make_args_from_* with model_name=model_id. The model_name argument is used exclusively in user-facing error messages (ImgGenParameterError). Since model_name (config name, e.g. "gpt-image-2") and model_id (provider ID, e.g. "gpt-image-2-2026-04-21") can differ, errors will cite the wrong identifier, making debugging harder. The model_name parameter was explicitly added to this function for this reason, so it should be forwarded correctly.

   Prompt To Fix With AI

   This is a comment left during a code review.
   Path: pipelex/cogt/img_gen/img_gen_args_factory.py
   Line: 95-113
   
   Comment:
   **`model_id` passed where `model_name` expected**
   
   Lines 102, 111, and 178 all call `make_args_from_*` with `model_name=model_id`. The `model_name` argument is used exclusively in user-facing error messages (`ImgGenParameterError`). Since `model_name` (config name, e.g. `"gpt-image-2"`) and `model_id` (provider ID, e.g. `"gpt-image-2-2026-04-21"`) can differ, errors will cite the wrong identifier, making debugging harder. The `model_name` parameter was explicitly added to this function for this reason, so it should be forwarded correctly.
   
   How can I resolve this? If you propose a fix, please make it concise.

Prompt To Fix All With AI

This is a comment left during a code review.
Path: pipelex/kit/configs/inference/backends/openai.toml
Line: 263-275

Comment:
**`model_choice` missing from all OpenAI image model rules**

None of the five new OpenAI image model rule blocks (`gpt-image-1`, `gpt-image-1-mini`, `gpt-image-1.5`, `gpt-image-2-2026-04-21`, `gpt-image-2`) include `model_choice = "model_id"`. Because `openai_img_gen_worker.py` now calls `openai_client.images.generate(**args_dict)` (and `.edit(**args_dict)`) entirely via the factory-built dict, and `ImgGenArgTopic.MODEL_CHOICE` is never hit, the `model` key is never added. The SDK call therefore omits the required `model` parameter, causing it to either fail or silently use an unintended default model. The same omission is present in all five matching blocks in `azure_openai.toml`. Adding `model_choice = "model_id"` to every `[*.rules]` block, the same way the unit-test helper `_make_gpt_image_2_rules()` already does, will fix this.

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: pipelex/cogt/img_gen/img_gen_args_factory.py
Line: 95-113

Comment:
**`model_id` passed where `model_name` expected**

Lines 102, 111, and 178 all call `make_args_from_*` with `model_name=model_id`. The `model_name` argument is used exclusively in user-facing error messages (`ImgGenParameterError`). Since `model_name` (config name, e.g. `"gpt-image-2"`) and `model_id` (provider ID, e.g. `"gpt-image-2-2026-04-21"`) can differ, errors will cite the wrong identifier, making debugging harder. The `model_name` parameter was explicitly added to this function for this reason, so it should be forwarded correctly.

How can I resolve this? If you propose a fix, please make it concise.

---

This is a comment left during a code review.
Path: pipelex/plugins/openai/openai_img_gen_factory.py
Line: 99-103

Comment:
**Potential division by zero in size validation**

`long_short_ratio = max_edge / min_edge` divides before the pixel-area floor check runs. If an `ImageSize` with `width=0` or `height=0` is supplied, `min_edge` is 0 and this raises `ZeroDivisionError` rather than the intended `ImgGenParameterError`. Adding a guard like `if min_edge == 0: raise ImgGenParameterError(...)` before this division would keep error handling consistent.

How can I resolve this? If you propose a fix, please make it concise.

_{Reviews (1): Last reviewed commit: "cleanup TODO" | Re-trigger Greptile}

[cubic-dev-ai]

2 issues found across 34 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="pipelex/plugins/openai/openai_img_gen_worker.py">

<violation number="1" location="pipelex/plugins/openai/openai_img_gen_worker.py:69">
P2: Don’t drop the `moderation` argument for `images.edit`. The OpenAI API supports `moderation` on edits, so this removes a requested safety setting.</violation>
</file>

<file name="pipelex/cogt/img_gen/img_gen_param_support.py">

<violation number="1" location="pipelex/cogt/img_gen/img_gen_param_support.py:196">
P2: `check_blueprint_params` omits `input_fidelity` validation, so unsupported blueprint `input_fidelity` values are not caught during static blueprint checks.</violation>
</file>

_{Reply with feedback, questions, or to request a fix. Tag @cubic-dev-ai to re-run a review.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 19bae2c459

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".