@gitawego/opencode-vision

Capability-aware vision + paste plugin for opencode.

opencode-vision makes image files referenced in your messages usable by the active model — and never wastes a call:

• Multimodal primary (e.g. MiniMax-M3, GPT-5 vision) → image paths in your message are rewritten to [Image-#N] markers and the images are attached natively. Zero delegation, zero extra tokens.
• Text-only primary → the message gets markers + a hint (or an automatic description) and the model delegates to a configured vision-capable model through a hidden opencode subagent (the vision agent), not a hand-rolled HTTP call.

Ported from the pi extension @gitawego/pi-vision.

Features

• Paste UX. Paths (absolute /…, home ~/…, relative ./…, Windows C:\…, drag-and-drop escaped spaces) → [Image-#N] markers.
  • Multimodal primary → native attachment.
  • Text-only primary → hint (zero-token nudge), auto (auto-describe, timeout protected, concurrency-bounded), or off (markers only).
• describe_image tool. Single (image_path) or batch (image_paths, up to 50) analysis with one subagent turn. Returns structured text.
• Subagent delegation. A hidden vision subagent runs the configured vision model in a child session; the images travel as data URLs so the model sees them natively. Retry (5xx/429/network, backoff), configurable fallback model, abort-aware.
• Settings menu (TUI). /vision, the command palette (Vision settings), or ctrl+shift+m opens a dialog-based panel — model picker (auto-lists vision-capable models), paste mode, marker style, local-only, cache, concurrency, and more. ctrl+shift+i is the quick model picker.
• Capability-aware by default. Auto-detects a vision-capable model when none is configured. On a multimodal primary, describe_image refuses with a redirect instead of delegating.
• Cache. Content-addressed (sha256(bytes) + prompt + model), memory + optional disk LRU. A repeat call costs zero subagent spawns.
• Audit log. Every delegation is recorded as one JSONL line in ~/.config/opencode/vision-audit.log (routing only — image bytes are never logged, only a source_hash fingerprint). Opt-out.
• Local-only mode. Image bytes never leave the machine; cache hits still work, a miss refuses with a clear error.

Install

Requires opencode ≥ 1.18 (TUI plugin support). Loads the ./server and ./tui package entries automatically. Plugins load at opencode startup — restart opencode after installing.

One command installs both entries (server + TUI settings page):

opencode plugin /path/to/opencode-vision -g

This is fully transparent: opencode surgically inserts the plugin into the plugin array of both opencode.json and tui.json (global config), preserving every other setting and comment, and skips files where the plugin is already present. The in-TUI plugin manager (/plugins → install) does the same. Re-running is a no-op.

Why both files? opencode loads the two entries from different config files:

• The server entry (image detection, describe_image, subagent delegation) loads from the plugin array in opencode.json.
• The TUI entry (the /vision settings page, ctrl+shift+m hotkeys) loads from the plugin array in tui.json (global ~/.config/opencode/tui.json, or a project-local tui.json / .opencode/tui.json).

If the plugin only appears in opencode.json, opencode starts and image handling works, but there is no settings page (/vision won't exist).

From git (the pi install git:... equivalent):

opencode plugin github:gitawego/opencode-vision -g   # install + write to global opencode.json
opencode plug github:gitawego/opencode-vision        # alias; -f to force-replace a pinned version

or list the git spec in opencode.json directly (a #tag/#commit pin works):

{ "plugin": ["github:gitawego/opencode-vision#v0.1.0"] }

The loader passes non-path specs to @npmcli/arborist (npm-package-arg), so github:user/repo, git+https://github.com/user/repo.git, bare user/repo, and #committish pins all work. This is undocumented upstream — the officially documented methods are npm packages and local files — so treat git install as experimental.

From a local checkout (dev):

// opencode.json
{ "plugin": ["/path/to/opencode-vision"] }

From npm (once published):

{ "plugin": ["@gitawego/opencode-vision"] }

Configure

Everything is configurable from the settings menu (no manual file editing needed):

Action	How
Open settings	/vision or command palette → "Vision settings" or ctrl+shift+m
Pick vision model	Settings → "Vision model", or ctrl+shift+i
Paste behavior on text-only primaries	Settings → "Paste mode" (hint/auto/off)
Local-only / audit / cache / concurrency	Settings menu
Reset	Settings → "Clear config"

Config is stored at ~/.config/opencode/vision.json. OPENCODE_VISION_CONFIG overrides the path (dev convenience).

Using describe_image

On a text-only primary the model can analyze images directly:

describe_image(image_path: "/tmp/screenshot.png", prompt: "What's in this image?")
describe_image(image_paths: ["/tmp/a.png", "/tmp/b.png"], prompt: "Compare these.")

On a multimodal primary you don't need the tool — just reference the path and the image is attached natively.

How it works

1. Server (config hook) registers the hidden vision subagent and auto-detects a vision-capable model.
2. Paste hook (chat.message) rewrites image paths into [Image-#N] markers and, for multimodal primaries, attaches them as native file parts.
3. Delegation (describe_image or auto mode) spawns a child opencode session running the vision agent with the vision model; images are attached as data-URL file parts. The description is returned as the tool result.
4. TUI provides the settings menu + commands.

describe_image is always registered (opencode plugins can't hide a tool per model). For multimodal primaries its execute detects the active model's capability at runtime and returns a "respond directly" redirect instead of delegating.

Verification checklist (after restart)

1. opencode starts with no plugin errors: grep -iE "opencode-vision|vision" ~/.local/share/opencode/log/opencode.log | grep -iE "error|fail" → empty.
2. /vision opens the settings menu; pick a vision model from the list.
3. Reference an image path in a message (e.g. /tmp/screenshot.png) on a text-only primary → the message gains markers + a hint, and the model can call describe_image.
4. Check ~/.config/opencode/vision-audit.log for delegation entries.

Project layout

src/server/index.ts   server plugin: hooks + describe_image tool
src/tui/index.tsx     TUI plugin: settings menu, commands, hotkey
src/…                 pure modules (config, paste, marker, cache, delegate, …)
test/                 node:test suites (30 tests, fake-client delegate tests)
SPEC.md               design + porting matrix + validation items

Development

npm install
npm run typecheck   # tsc --noEmit
npm test            # node --import tsx --test

Credits

Forked design from getpipher/vision → @gitawego/pi-vision (pi) — the code, config shape and behavior originate there. This is the opencode plugin API port with subagent-based delegation.

License

MIT