feat: builtin media tools — read_image / transcribe_audio / read_video

[localai-bot]

What

Three builtin agent tools that let the model read a media file from the workspace and get text back — the programmatic counterpart to the user-facing attachments from #48. Each is one specialist call via the LocalAI API, so any agent model (even text-only) can use them:

Tool	Delegate	How
read_image(path, question?)	specialist.Describe	chat with an image_url part → vision model → text
transcribe_audio(path)	specialist.Transcribe	/audio/transcriptions → parakeet/whisper → transcript
read_video(path, question?)	new specialist.DescribeVideo	chat with a native video_url part → video-capable model → text

read_video needs no ffmpeg — LocalAI parses video_url content parts natively (core/http/middleware/request.go), so it mirrors read_image.

How

• specialist.DescribeVideo mirrors Describe, sending a base64 video_url part; reuses DataURI.
• Config gains VideoModel; the Session gains videoModel/workingDir; per-tool model selection (read_image→visionModel, read_video→videoModel, transcribe_audio→transcribeModel; empty ⇒ LocalAI default). The visionModel field reserved in feat: multimodal attachments — file→text, transcription, native image/audio/video (+ CLI @path//attach) #48 is now consumed.
• Tools live in chat/mediatools.go, each holding an injected delegate (the askUserTool pattern → unit-testable with fakes). Paths resolve against WorkingDir via resolveWorkspacePath (roots relative paths, matching the host read tool — not a sandbox). Errors are returned as the tool-result string so the model reads and reacts to them.
• Gating: each tool is individually gated by s.toolEnabled(name) in the BuiltinTools allowlist and added to readOnlyTools (observational → auto-approve in default prompt mode).
• Media calls run on the per-turn context, so Ctrl+C interrupts an in-flight description/transcription.

Testing

DescribeVideo httptest (asserts a data:video/ video_url part is sent), each tool's Run with fake delegates (arg threading, empty-path guard, delegate-error → error result), resolveWorkspacePath table test, and IsReadOnly membership. go build ./... + go test ./chat/ ./specialist/ green.

Built subagent-driven with per-task spec+quality reviews and a final whole-branch review, which caught and this PR fixes the media delegates capturing the session context instead of the per-turn context (so in-flight calls now honor interrupt).

Follow-ups

Optional per-tool config docs; frame-sampling fallback for models lacking native video input (only if a text-only-video workflow is ever needed). The desktop UI is a separate nib-desktop effort.

🤖 Generated with Claude Code