Attached images degrade to a bare text placeholder when the model lacks vision — agent has no way to read them

Summary

When a user pastes images into a chat and the active model does not support vision (e.g. a text-only model like deepseek-v4-flash), the image file parts are replaced by a plain text placeholder before being sent to the model. The agent then has no way to access the actual image content — the placeholder carries no reference to how the image can be retrieved.

Current behavior

In packages/opencode/src/session/message-v2.ts (around line 213):

if (options?.stripMedia && isMedia(part.mime)) {
  userMessage.parts.push({
    type: "text",
    text: `[Attached ${part.mime}: ${part.filename ?? "file"}]`,
  })
}

So the model sees only [Attached image/png: image-2.png]. There is no session/message/part id, no API reference, no data URL — nothing the agent could use to fetch the image itself. The only way to recover the image is to query the local opencode.db (the part table stores {type:"file", mime, filename, url} with session_id/message_id) or hit the local serve API, which is undocumented for this purpose and requires guessing the current session.

Why it matters

This forces workarounds: to read a pasted image, an agent (or a skill like an "image recognition" skill) has to:

1. guess the current session (no env var exposes session_id),
2. query the sqlite DB / serve API to find the most recent image part,
3. forward the data: URL to a separate vision model.

That's fragile (parallel sessions race) and entirely non-obvious from the placeholder alone.

Proposal (any of these would help)

1. Include a retrievable reference in the placeholder, e.g. [Attached image/png: image-2.png] (session: …, message: …, part: …) or a URL like opencode://session/{sessionID}/message/{messageID}/part/{partID} the agent can dereference.
2. Expose the current session id to agents (e.g. OPENCODE_SESSION_ID env var when running tool calls) so tools/skills can scope queries precisely.
3. A config option to keep sending lightweight media metadata (mime, filename, session/message/part ids) to text-only models instead of dropping it.

Environment

• opencode version: 1.18.18 (OpenChamber distribution, opencode serve)
• Model: deepseek-v4-flash (no vision)
• OS: macOS

Happy to help with a PR if a direction is agreed.