Fix/generate-image

[makiroll1125]

What

• Add OpenAI Image 2.0 image generation functionality
• Change Gemini image gen model to Nano banana 2

Why

Updated CraftBot to use state-of-the-art image gen models
Closes #219

[ahmad-ajmal]

Requesting changes - architectural concern + a few correctness issues

Main blocker: provider branching duplicates infrastructure we already have

This change hardcodes if provider == "openai" / elif "gemini" inside the action, with two SDK imports, two client inits, and two error-mapping blocks. We already have provider abstraction elsewhere in the repo:

• MODEL_REGISTRY + InterfaceType enum (used by VLM/LLM)
• LLMInterface / VLMInterface wrappers in app/llm_interface.py and app/vlm_interface.py that hide the provider-specific SDK calls
• describe_image.py (lines 62–70) is the reference pattern: read the configured provider from MODEL_REGISTRY[provider][InterfaceType.VLM], then delegate

As-is this PR builds a third parallel provider system that future image providers (Stability, Replicate, xAI, OpenRouter image, etc.) will all have to extend by adding another elif branch here. It also introduces a new image_generation.preferred_provider setting that parallels the existing vlm_provider / llm_provider pattern instead of joining it.

Could we route this through MODEL_REGISTRY with a new InterfaceType.IMAGE_GEN and an ImageGenInterface in agent_core, mirroring how VLMInterface is set up, so generate_image.py ends up looking like describe_image.py? Reusing InterfaceType.VLM directly is tempting since some providers serve both through one endpoint, but the capability sets differ (Claude / ByteDance support VLM but not gen) and users will want to pick providers independently for each.

Other issues worth fixing while you're in here

• OpenAI aspect-ratio map is wrong. "16:9": "1536x1024" is 3:2, "9:16": "1024x1536" is 2:3. The canvas constraint is real (gpt-image only has 3 sizes), but silently mismapping → at least append to warnings. (I skimmed real quick so please verify)
• Silent 4K downgrade for OpenAI. "4K": "high" returns at most 1536×1024. Either reject 4K for OpenAI or warn.
• quality dropped on the edit path. images.generate(..., quality=...) is passed, but images.edit(...) isn't - reference-image runs silently render at lower quality.
• images.edit ≠ "style reference." The existing reference_images field is documented as style guidance (how Gemini uses them). OpenAI's images.edit treats inputs as compositional/mask inputs. Same input, very different output between providers.
• Provider-selection UX doesn't match the PR description. The description says "asks the user" when both keys are present, but the code silently defaults to Gemini - there's no signal in the response telling the calling LLM that a choice is available. Once provider_preference is saved, there's also no way to clear it.

Happy to pair on the ImageGenInterface refactor if it'd help.

@zfoong What do you think about this change? Worth the effort?

[ahmad-ajmal]

Follow-up: extract these error strings to a message catalog

Even setting aside this PR, both error-mapping blocks (Gemini + OpenAI) are the same five patterns repeated with provider-specific text spliced in:

Template key	Triggers (substrings)	Placeholders
provider_rate_limit	quota, rate, insufficient_quota, billing	{provider_label}, {error}
provider_invalid_key	invalid + key, invalid_api_key	{provider_label}, {error}
provider_access_denied	permission, access	{provider_label}, {model}, {error}
provider_safety_block	safety, blocked, content_policy	{provider_label}, {error}
provider_model_not_found	not found, 404	{provider_label}, {model}, {help_url}, {error}

Suggested shape - anchors on the existing get_os_language() setting in app/config.py:341, which currently isn't wired to anything:

• app/i18n/errors.en.json - flat catalog of templates with {placeholder} substitution
• app/i18n/errors.<lang>.json - locale overrides; missing keys fall back to en
• app/i18n/__init__.py exposing two helpers:
  • t(key, **kwargs) -> str - formatted lookup with locale + en fallback
  • classify_provider_error(exc, *, provider, model) -> str - does the substring matching against a triggers table in the catalog and returns the formatted message in one call

The error block in this PR collapses to roughly:

except Exception as e:
    return {
        "status": "error",
        ...,
        "message": classify_provider_error(e, provider=provider, model=model_id),
    }

Benefits:

• one mapping table instead of two near-identical blocks per provider
• localization-ready without touching call sites
• adding a new provider = adding a row to the catalog, not another if/elif ladder
• avoids writing a bespoke message for every case - templates cover ~90% of them
• i think with the help of claude we can implement this pretty quick and easy

@zfoong What do you think about this too?
Happy to file this as a separate issue once we figure out which agent_core module owns it (since the same templates will be useful for the LLMInterface / VLMInterface paths too).