feat: add Prism provider

[rajitkhanna]

Summary

• add Prism as an OpenAI-compatible inference provider
• document Prism's Chat Completions, Responses, and Anthropic Messages API surfaces
• add DeepSeek V4.1 Flash and DeepSeek V4 Flash with Prism pricing and reasoning controls
• inherit V4.1 image capability while keeping V4 text-only
• add a monochrome Prism provider logo

Sources and verification

• Model catalog: https://docs.prisminference.com/models
• Live model IDs: https://api.prisminference.com/v1/models
• Pricing: https://prisminference.com/pricing
• Chat Completions: https://docs.prisminference.com/api-reference/chat-completions
• Responses: https://docs.prisminference.com/api-reference/responses
• Anthropic Messages setup: https://docs.prisminference.com/guides/coding-agents
• OpenAPI contract: https://docs.prisminference.com/openapi.yaml

Prism intentionally exposes reasoning_effort=none|low|medium|high; none disables reasoning and V4.1 maps Prism medium/high to the underlying high/xhigh tiers. Both submitted model IDs are returned verbatim by the live Models API. DeepSeek V4.1 Flash accepts base64 image inputs through Chat Completions, Responses, and Anthropic Messages; remote image URLs are not supported. DeepSeek V4 Flash remains text-only.

This PR remains draft until the V4.1 image-input API and documentation changes are deployed.

Validation

• bun validate

[github-actions]

Action items

• [high] [possible mistake] providers/prism/models/prism/deepseek-v4-flash.toml:12 - Check: Relay reasoning_options must match this host’s wire API and the DeepSeek V4 Flash lab/peer baseline, not a generic GPT effort enum. Why: Both Prism models set effort = ["none", "low", "medium", "high"]. First-party DeepSeek Flash uses toggle + ["low", "high", "max"] (and Pro uses high/max); established openai-compat peers (e.g. ofox, tokengo) follow that shape. medium and off-via-none look like an invented L/M/H gateway default unless Prism’s API actually accepts and applies those values. Action: Confirm against Prism docs or a live request. If Prism only maps DeepSeek-style controls, use toggle + ["low", "high", "max"] (or the exact Prism set) with a leading wire-path comment for the toggle. If Prism truly exposes reasoning_effort=none|low|medium|high, cite that host surface in the PR body and keep effort only (no toggle with none).
• [high] [possible mistake] providers/prism/models/prism/deepseek-v4.1-flash.toml:9 - Check: Same reasoning-options baseline for DeepSeek V4.1 Flash on a multi-model relay. Why: Same none|low|medium|high set as above, while lab/OpenRouter/ofox-style Flash peers use toggle + low|high|max (or host-documented variants). Incorrect options break client reasoning controls for this provider. Action: Align with verified Prism wire behavior or the lab/peer Flash set; document the exact request fields.
• [medium] [possible mistake] providers/prism/models/prism/deepseek-v4.1-flash.toml:6 - Check: Provider base_model files may override modalities/attachment only when this host truly differs. Why: Lab deepseek/deepseek-v4.1-flash has attachment = true and modalities.input = ["text", "image"], but this file forces attachment = false and text-only input/output. That strips vision for every consumer of Prism’s V4.1 Flash entry. Action: Verify Prism’s catalog/API for this model ID. If vision is supported, drop the redundant overrides and inherit the lab modalities/attachment. If Prism is text-only, keep the overrides and note that host limitation in the leading comment or PR body.
• [medium] [possible mistake] providers/prism/models/prism/deepseek-v4-flash.toml:1 - Check: Model ID is the path under providers/<provider>/models/ and must match the provider API id. Why: Paths yield ids prism/deepseek-v4-flash and prism/deepseek-v4.1-flash. Most DeepSeek relays use deepseek/… or a bare deepseek-v4-flash id; a prism/ org prefix is unusual unless Prism’s /v1/models list literally returns those strings. Action: Match filenames to Prism’s live model ids (from the cited Models API). Move/rename if the API uses deepseek/… or unprefixed ids.

[rajitkhanna]

Verified the reviewer's four points against Prism's public contracts:

• reasoning_effort=none|low|medium|high is Prism's documented API surface: https://docs.prisminference.com/api-reference/chat-completions
• both prism/... IDs are returned verbatim by the unauthenticated live catalog: https://api.prisminference.com/v1/models
• Prism currently accepts text-only message content, so V4.1 intentionally overrides the base model's image capability
• the model TOMLs now cite the exact reasoning documentation directly

bun validate still passes after the documentation update.

[github-actions]

No actionable findings.