fix(ai): attach user media to Anthropic requests

[ling-senpeng13]

Problem

Images passed as input via the media path never reached Anthropic (Claude) models. The model received a text-only message and hallucinated a response instead of reading the image.

The pipeline was correct up to the provider conversion: media is stored on the ChatMessage, and LLMHelper.getMessage() downloads the URL to bytes and builds a Spring AI UserMessage with media. But AnthropicChatModel.convertMessage() built the user message from userMsg.getText() only and silently dropped userMsg.getMedia():

case USER -> {
    UserMessage userMsg = (UserMessage) msg;
    messages.add(Message.user(userMsg.getText()));   // getMedia() ignored
}

The OpenAI provider (OpenAIResponsesChatModel.convertMessage()) already forwards media as image content parts; the Anthropic converter never got that treatment. The Anthropic ContentBlock API already modeled type="image" with a base64 Source — the capability existed, it just wasn't wired up.

Fix

• AnthropicChatModel.convertMessage(): when a UserMessage carries media, build a text block plus one image block per media item (base64-encoding the downloaded bytes), mirroring the OpenAI path. Text-only messages are unchanged.
• AnthropicMessagesApi: add a ContentBlock.image(mediaType, base64Data) factory.

Test

Adds AnthropicChatModelMediaTest, which mocks the Messages API, captures the outgoing MessagesRequest, and asserts the user message serializes to a content-block list containing an image block with source.type=base64, media_type=image/png, and the verbatim base64 payload — plus the accompanying text block.

Validated both directions:

• Before the fix: FAILS — user content serializes to a bare String (media dropped).
• After the fix: PASSES.

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

Scope

Anthropic provider only; OpenAI/Gemini media paths are untouched. convertMessage is the single message-building path for Anthropic (used by the normal call path), so all Anthropic requests are covered.

End-to-end validation (full agentspan SDK e2e suite)

Built a server carrying this fix (the 0.3.0 agentspan server baseline + this exact conductor-ai change) and ran the full agentspan Python SDK e2e suite against it.

Result: 123 passed, 8 failed, 19 skipped/xfailed (150 tests, 23m).

The media-input suite (Suite 25) passes in full, including the Anthropic case — which is the direct target of this fix and was previously failing/skipped:

test_vision_reads_text_from_image[openai]       PASSED
test_vision_reads_text_from_image[anthropic]    PASSED   ← fixed by this PR (was: hallucinated, no image)
test_without_media_token_is_absent[openai]      PASSED
test_without_media_token_is_absent[anthropic]   PASSED

A direct probe confirms it: anthropic/claude-sonnet-4-5 + an image of the text MELON7391 now returns MELON7391 (before the fix the model received no image and hallucinated unrelated text).

The 8 failures are unrelated to this PR — all are a pre-existing version mismatch between the newer standalone agentspan Go CLI and the older 0.3.0 server's REST API (not the Anthropic provider):

• TestSuite16CliSkills (5) — CLI skill --version flag unknown / skill load HTTP 400 against the 0.3.0 server.
• TestSuite2ToolCalling, TestSuite4McpTools, TestSuite5HttpTools (1 each) — credentials delete → HTTP 404 No static resource api/credentials/... (endpoint absent on the 0.3.0 server).

None touch the LLM provider path; they fail identically with or without this change.

Exact-version confirmation: the same result was reproduced building conductor-ai from the v3.32.0-rc.3 tag with this fix cherry-picked (not just the 3.30.2 baseline) and running it inside the matching agentspan server — AnthropicChatModelMediaTest passes at that source, and Suite 25 passes end-to-end (4 passed, Anthropic included).

🤖 Generated with Claude Code