feat(models): support provider-owned image models

[alwaycoding]

Summary

• make image generation a provider-owned capability, with image-only custom providers and separate chat/image inventories
• support standard OpenAI Images, Ark Images, and explicit extra_body compatibility modes
• preserve provider/model/profile selection, Ark compatibility, and legacy generate_image behavior
• validate and recover persisted image model configuration without affecting chat inventories

Review Fixes

• Use OpenAI pixel dimensions and omit response_format by default for GPT Image; canonical DALL-E models explicitly request base64.
• Preserve an explicit ark-images protocol on custom providers through save/reload.
• Add an optional model-level responseFormat: "b64_json" setting for gateway aliases, wired through the editor, persistence, and request payload. No image-URL download path is introduced.
• Normalize legacy 1K to 1024x1024 only for standard OpenAI Images, across the settings editor, inventory validation, and thread tools. Editing a model no longer replaces its size set with all options or resets its default to auto. Ambiguous legacy sizes require explicit correction; Ark and extra-body presets remain unchanged.
• Apply the old GPT Image size restrictions only to known GPT Image 1 variants/snapshots. GPT Image 2 and its dated ID accept the configured 1792x1024 option while DALL-E/GPT Image 1 restrictions remain enforced.
• Merge current main (f2bb156) and preserve its desktop internationalization. New image settings and provider selection labels are available in English and Chinese.

Latest integration: afa506bcbb9613cd0935f1cf958297d626d92a09.

Verification

• full Bun suite: 904/904 tests passed across 164 files
• focused image/runtime/UI/i18n suite: 85/85 tests passed, including 8 mounted React regressions
• full repository ESLint: passed with zero warnings/errors
• full cross-package TypeScript check: passed
• Prettier and git diff --check: passed
• desktop renderer production build and web production build: passed (existing chunk-size warnings; web also reports its tolerated target/top-level-await warning)
• final PR diff scan: 27 files, no .trae/evaluation assets or credential matches

The prior live Agnes smoke at 9a93982 generated a 1K PNG. Live OpenAI/Agnes requests and native CEF interaction were not rerun for this integration commit; provider behavior is covered with injected responses and mounted React tests.

No Agnes-specific identifiers or endpoints are hard-coded in the implementation. Evaluation assets and credentials are excluded from this PR.

[MagicCube]

Found two issues with image protocol handling. The 33 focused tests I ran passed, but do not cover these cases.

[MagicCube]

[P1] Build a compatible payload for standard OpenAI Images

The new openai-images mode still sends metadata.size verbatim, while configuration and runtime validation only permit 1K/2K/3K/4K. Standard OpenAI Images requires pixel dimensions such as 1024x1024. This branch also always sends response_format, which GPT Image models do not support. Consequently, configuring a custom provider against the standard OpenAI endpoint cannot generate an image with the available settings. Please support protocol/model-appropriate sizes and output parameters, and add coverage for the standard mode alongside the extra-body fixture. Official parameter definitions.

[alwaycoding]

Fixed in 9a93982. Standard OpenAI Images now uses the official pixel-size options, maps the legacy 1K value to 1024x1024, rejects ambiguous larger presets before fetch, and only sends response_format for DALL-E models (GPT Image omits it). The settings editor and generate_image contract now expose protocol-compatible sizes. Added focused GPT Image, DALL-E, and invalid-preset coverage.

[MagicCube]

[P2] Preserve an explicit Ark protocol on custom providers

Normalization unconditionally drops api: "ark-images", but a custom provider defaults to openai-images when api is absent. I reproduced this by adding a custom provider, saving an image configuration with api: "ark-images" through updateProvider, and constructing a new ModelManager with the same settings directory: the API changes from ark-images to undefined. After restart, generation therefore uses a different payload and omits Ark's watermark/stream fields. Please only omit the API when it equals defaultApi, or preserve the normalized API explicitly.

[alwaycoding]

Fixed in 9a93982. Image protocol normalization now omits api only when it equals the owning provider default, so a custom provider explicitly configured with ark-images survives persistence and reload. Added a restart regression test and exposed Ark Images in the custom-provider image API selector.

[alwaycoding]

Pushed follow-up fixes in 814298d:

1. Gateway model aliases can explicitly request b64_json through a persisted model setting. GPT Image defaults still omit response_format.
2. Legacy 1K standard-OpenAI models retain their size/default when edited. Both new and existing thread tools remain usable after save/reload, with matching UI display. Native Ark/extra-body presets are unchanged.
3. GPT Image 2 no longer inherits the GPT Image 1 size whitelist; 1792x1024 works for its regular and dated IDs. Known GPT Image 1 and DALL-E limits remain checked.

Verification: 72/72 top-level focused tests, including an isolated suite of 8/8 mounted React regressions; changed-file lint/typecheck, Prettier, and desktop/web production builds passed. The PR description records build warnings and the live-provider/native-CEF checks that were not rerun.

[alwaycoding]

Resolved the current main conflicts in merge commit afa506bcbb9613cd0935f1cf958297d626d92a09 (base f2bb15600d1934bad62cf2595374ef1b27056ac2).

Resolution details:

• preserved upstream desktop/provider internationalization and localized the new image API, response-format, provider-selection, and validation labels in English and Chinese
• preserved provider-owned image inventories, custom image-provider selection, connection persistence, and legacy OpenAI 1K normalization
• adapted the mounted settings regression fixture to the new real I18nProvider

Local verification:

• full test suite: 904/904 passed (the Python generator test ran with jinja2 in an isolated /tmp virtualenv)
• full lint: passed
• full cross-package typecheck: passed
• desktop renderer and web production builds: passed
• final PR diff scan: 27 files, no .trae/evaluation assets or credential matches

GitHub now reports the PR as mergeable; checks may still need to settle.