feat: add Agnes AI provider support (text, image, video)

[susirial]

Summary

Adds support for three Agnes AI models as configurable providers in OpenChatCut:

Model	Type	Provider ID	Default Model ID
Agnes 3.0 Flash	Text / LLM (OpenAI-compatible Chat Completions)	agnes	agnes-3.0-flash
Agnes Image 2.5 Flash	Image generation (OpenAI Images-compatible)	agnes-image	agnes-image-2.5-flash
Agnes Video 2.5 Flash	Video generation (OpenAI Videos-compatible, async)	agnes-video	agnes-video-2.5-flash

All three share the same base URL (https://apihub.agnes-ai.com/v1) and API key, configurable in Settings → Agent Brain / AI Generation → Image / AI Generation → Video.

Changes

Text (LLM)

• Registered Agnes as an openai-compatible LLM provider preset in shared/llm-providers.ts (auto-generates LLM_AGNES_API_KEY / LLM_AGNES_BASE_URL / LLM_AGNES_MODEL config keys).
• Added the three LLM_AGNES_* keys to KEY_NAMES so they can be persisted via the settings API.

Image

• New callAgnesImageProvider in server/plugins/image-provider-clients.ts — calls /v1/images/generations with size tier (1K/2K/3K/4K), ratio aspect ratio, and optional image[] references (Data URI for img2img). Returns data[0].url or data[0].b64_json.
• Wired agnes-image into ValidImageRequest, validateImageRequest (max 4 outputs, size tiers, aspect-ratio whitelist), the dispatch branch, and ImagePluginOptions.
• Added AGNES_IMAGE_* keystore keys, settings vendor page, route option, and caps detection.

Video

• New generateAgnesVideo in server/plugins/video.ts — POST /v1/videos to create a task, then polls GET /v1/videos/{id}?model_name= until completed. Supports text, keyframe (first/last frame), and reference (up to 5 images + 3 audios) modes. Fixed 720p, 4–12s duration.
• Wired agnes-video into VideoRequest/ValidVideoRequest, validateAgnes, runVideoOperation dispatch, the resumer provider list, and VideoOptions.
• Added AGNES_VIDEO_* keystore keys, settings vendor page, route option, and caps detection.

Shared & agent-facing

• computeCaps() now detects AGNES_IMAGE_API_KEY / AGNES_VIDEO_API_KEY so the settings UI correctly shows configured state at runtime.
• __CONFIGURED_CAPS__ build snapshot updated in config/vite.config.ts.
• Added Agnes entries to agent tool schemas (generate-schemas.ts), tool-input mapping (generate-tool-input.ts), capabilities routing (capabilities.ts), skill docs (image-gen/video-gen SKILL.md + new agnes-image.md/agnes-video.md references), and connection probes (key-probes.ts).
• Fixed Fal model routing: the Fal model catalog now only surfaces in the agent prompt when Fal is the actively routed image/video provider, so it no longer shadows another default such as Agnes.

Verification

• npx tsc -b --force — passes
• npm run build — succeeds
• Settings UI confirmed: Agnes appears in Agent Brain, Image, and Video vendor lists; saved API keys persist and show "configured"
• curl POST /api/keys with Agnes keys returns source: runtime and caps.image/caps.video: true

Notes

• package-lock.json was intentionally excluded from this PR (the local npm install only reformatted the lockfile with the installed npm version; no dependency changes).