Problem

Gemini CLI runtime (src/middleware/runtimes/gemini.ts) currently only handles text prompts. Gemini has the best multimodal support of all runtimes — it can natively handle images, audio, video, and PDF via @path inline syntax.

Current state

• Uses --output-format stream-json --prompt for invocation
• buildEnv() returns {}
• No media handling

Implementation

Inbound (images, audio, video, PDF)

• Save MediaAttachment content to temp files when params.media is present
• Inject @/path/to/media.jpg references inline in the prompt text
• Gemini CLI resolves @path references and includes the media natively
• Declare capability: acceptsInbound: ["image/", "audio/", "video/", "application/pdf"]

Outbound

• No native media emission capability
• emitsOutbound: false

Capability declaration

readonly mediaCapabilities = {
  acceptsInbound: ["image/", "audio/", "video/", "application/pdf"],
  emitsOutbound: false,
};

Temp file management

• Create temp directory before execution
• Write media attachments as temp files with appropriate extensions
• Clean up temp directory after done event (or on error)

Tests

Unit tests (src/middleware/runtimes/gemini.test.ts)

• execute() with image attachment creates temp file and injects @path in prompt
• execute() with audio attachment creates temp file and injects @path
• execute() with multiple attachments creates multiple temp files
• execute() without media uses existing prompt path (backwards compat)
• Temp files are cleaned up after execution completes
• Temp files are cleaned up on execution error
• mediaCapabilities reports correct values

Live smoke tests (src/middleware/__smoke__/gemini.live.test.ts)

• Send image attachment, verify Gemini describes the image content
• Send audio attachment, verify Gemini transcribes/describes audio
• Send text-only prompt, verify existing behavior unchanged

Depends on

• feat(middleware): add multimodal media attachments to AgentRuntime contract #385 — AgentRuntime multimodal contract

Related

• feat(runtimes): implement multimodal I/O for each CLI runtime #386 — per-runtime multimodal (parent tracking issue)
• research(auth+media): auth profiles and media understanding architecture #376 — auth/media restructuring