fix(agent-runner): forward pasted image content to vision-capable models

[lxistired]

Summary

When a user pastes an image in ChatView, the renderer creates a content block in Anthropic Messages API shape:

{ type: 'image', source: { type: 'base64', media_type, data } }

ClaudeAgentRunner in src/main/claude/agent-runner.ts only forwarded the prompt text via piSession.prompt(string), so the image bytes were silently dropped before reaching the model. The pre-existing hasImages flag was set but never acted on — vision-capable agents received only the text.

This patch:

1. Walks the current user message's content blocks and extracts any type:'image' entries.
2. Normalises the Anthropic shape (source.media_type, source.data) to the pi-coding-agent shape ({type:'image', mimeType, data}). Already-normalised blocks are accepted too for forward compat.
3. Passes the extracted images via piSession.prompt(text, { images: currentTurnImages }). The text-only path is preserved when the user didn't paste anything, so the wire format is unchanged in the common case.

Only the current user turn is processed — prior turns are untouched. The existing hasImages log line now also reports the extracted count for easier debugging.

Test plan

• Paste an image (PNG/JPEG) into ChatView, send with a question like "what's in this image?", confirm the model describes it instead of saying it can't see it.
• Send a text-only message, confirm no behaviour change (no images key in the underlying request).
• Send a message with multiple pasted images, confirm all are forwarded.
• Confirm log line shows e.g. User message contains images: 2 extracted.

[github-actions]

Findings

• No high-confidence issues found in the added/modified lines.

Summary

• Review mode: initial. No correctness, security, or regression issues were identified in src/main/claude/agent-runner.ts on the modified lines.
• Residual risk: this change adds a new image-forwarding path in src/main/claude/agent-runner.ts without focused regression coverage under tests/, so future SDK-shape drift may not be caught automatically.

Testing

• Not run (automation). Suggested follow-up: add an agent-runner regression test that asserts image blocks are normalized to { type: 'image', mimeType, data } and passed to piSession.prompt(..., { images }).

Open Cowork Bot