fix(ai): attach user media for Cohere (vision) requests

[ling-senpeng13]

Problem

Cohere is vision-capable — command-a-vision-07-2025 (and Aya Vision) accept image input, and Cohere's v2 chat API accepts image_url content parts (base64 data URI or URL). But conductor-ai dropped image input entirely:

1. CohereChatModel.toCohereMessage() built the message from message.getText() only — getMedia() was ignored.
2. Deeper: the request DTO CohereApi.ChatMessage.content was a bare String, so it couldn't even represent content parts.

Same class of bug as #1238 (Anthropic), #1241 (Gemini), #1243 (Grok/Perplexity), #1245 (HuggingFace) — Cohere was the last provider still dropping media.

Fix

• CohereApi.ChatMessage: widen content from String to Object (String or List<ContentPart>); add a ContentPart type (text / image_url + ImageUrl) and a ChatMessage.user(List<ContentPart>) factory.
• CohereChatModel.toCohereMessage(): for a UserMessage with media, emit text + one image_url content part per media item (base64 data: URI for bytes). Text-only messages unchanged.

Test

Three layers:

• CohereChatModelMediaTest (mocked HTTP, always runs — CI has no provider keys): captures the outgoing ChatCompletionRequest and asserts the user message serializes to a content-part list with an image_url part carrying the base64 data URI (plus the text part). FAILS before the fix, PASSES after.
• CohereAITest.IntegrationTests#testChatCompletionWithImageMedia (live API, self-skips without COHERE_API_KEY): drives the adapter against Cohere's real v2 chat API with a repo-committed image embedding a machine-unguessable token (MELON7391) and asserts the vision model transcribes it — a token, unlike a color, cannot be answered correctly by luck, so a pass proves the image arrived (pattern from test(e2e): Python media-input suite (image → vision model) agentspan-ai/agentspan#301).
• LLMChatCompleteTests#cohere_vision_imageMediaInMessages_modelSeesTheImage (live-API e2e, self-skips without COHERE_API_KEY): full server path — workflow messages[].media URL → HttpDocumentLoader download → adapter → live API — asserting the same token transcription. The image is committed in this repo (e2e/src/test/resources/assets/melon7391.png) and referenced at a raw URL pinned to its immutable commit SHA, so there is no third-party image host. COHERE_API_KEY is forwarded through the e2e docker-compose files like the other provider keys.

Also includes LLMChatCompleteTests#anthropic_haiku_vision_imageMediaInMessages_modelSeesTheImage — the same e2e shape through Anthropic, as a regression lock for the already-merged #1238 media fix, which shipped without an e2e test — plus its counterfactual (anthropic_haiku_vision_withoutMedia_tokenIsAbsent: same prompt, no media, token must be absent — proving the token can only come from the image). Both gated on ANTHROPIC_API_KEY and verified against the live API: the positive case transcribes MELON7391, the counterfactual completes without it.

The shared server pipeline for the e2e shape was verified against a live provider (Anthropic: same messages+URL input returns "Red" with image tokens counted in the prompt). The Cohere-gated tests are pending a COHERE_API_KEY — no key is available in CI or locally today, which is why the mocked test stays as the always-running regression guard.

Note (out of scope, provider-agnostic): workflow media passed as bare base64 never reaches any provider adapter — FileSystemDocumentLoader.supports() claims any string without :// and the document access policy rejects it. URL media is the supported end-to-end shape.

./gradlew :conductor-ai:test --tests "org.conductoross.conductor.ai.providers.cohere.*"
BUILD SUCCESSFUL

Notes

Image support is model-dependent (use a Cohere vision model such as command-a-vision-07-2025). This makes conductor forward the media in Cohere's documented image_url format.

🤖 Generated with Claude Code