feat(plugin): support images in custom tool responses

[leeweisern]

Summary

Custom tools can now return images alongside text, allowing the LLM to actually see and analyze visual data.

• Custom tools return { content: [{ type: "text", text }, { type: "image", mimeType, data }] }
• Backwards compatible: plain string returns still work
• Images converted to attachments for model vision input

Changes

• Add ToolContentPart and ToolResult types to plugin
• Add content normalization helper with validation (max 10 images, ~5MB each, PNG/JPEG/GIF/WebP)
• Handle rich content in tool registry
• Add attachments field to tool.execute.after hook
• Add 8 unit tests with 100% coverage on content module
• Update docs with usage examples and limits

Verification

• All 8 content tests pass: bun test test/tool/registry.test.ts --test-name-pattern "tool.content"
• 100% function and line coverage on content.ts

Fixes #9539

[github-actions]

The following comment was made by an LLM, it may be inaccurate:

No duplicate PRs found

[leeweisern]

@rekram1-node Could you help add the discussion and enhancement labels? This is a new feature that extends the custom tool API to support image returns.

[rekram1-node]

this seems overcomplicated, we already have the attachments notion I think someone else has a similar pr that accomplishes this

[leeweisern]

this seems overcomplicated, we already have the attachments notion I think someone else has a similar pr that accomplishes this

you mean it can support returning images in custom tool?

[nathanl]

I think this PR would solve a problem I'm having. I'm having Claude (via opencode) write Playwright tests, but when a test fails, Claude can't see the screenshots in the trace. I've written a command-line tool that supports commands like pwtrace screenshot --step 1 --base64 path/to/trace, so Claude is now able to get the base64 image data from any step, but it can't "see" what's in the image as it can if I manually drag in a file to upload. I want to take myself out of that loop so that Claude can iterate on failing tests.