feat(llm): feed tool screenshots back into vision models

[alexk-dev]

Summary

• preserve tool-produced image artifacts in history as lightweight metadata instead of inline base64
• rehydrate tool screenshots back into multimodal LLM requests for vision-capable models
• keep non-vision models and broken/missing artifacts on the existing text-only path

Why

Tools like pinchtab_screenshot already return image attachments to the user, but the next LLM call could not see those images.
That meant the agent could capture a screenshot and still be blind to it on the following reasoning step.

What changed

• DefaultHistoryWriter now stores image artifact references on tool messages via toolAttachments metadata
• Langchain4jAdapter now:
  • detects whether the target model supports vision
  • loads tool image artifacts through ToolArtifactService
  • appends a synthetic multimodal context message next to the matching ToolExecutionResultMessage
• tool image rehydration degrades safely to text-only when:
  • the model is not vision-capable
  • the artifact metadata is absent or malformed
  • the stored file no longer exists

Tests

• ./mvnw -Dmaven.gitcommitid.skip=true surefire:test -Dtest=Langchain4jAdapterTest,Langchain4jAdapterRetryTest,DefaultHistoryWriterTest
• ./mvnw -Dmaven.gitcommitid.skip=true pmd:check spotbugs:check -DskipTests

Coverage added

• tool image metadata is persisted on tool history messages
• vision models receive multimodal context from tool screenshots
• non-vision models stay text-only
• missing/broken artifacts do not break the LLM loop

[sonarqubecloud]

Quality Gate failed

Failed conditions
71.1% Coverage on New Code (required ≥ 80%)

See analysis details on SonarQube Cloud