Reticle

Fast, private code completion on your own endpoint.
Local-first FIM ghost text for VS Code—with multi-line suggestions, no hosted account, and no telemetry.

Install · MTPLX quick start · Provider guides · Configuration

Reticle connects VS Code directly to an OpenAI-compatible POST /v1/completions endpoint. It uses true fill-in-the-middle context to suggest an identifier, a line, or a bounded multi-line block at the cursor.

The selected model must actually support FIM. Reticle can send a separate OpenAI suffix or embed Qwen PSM special tokens for plain-completion servers such as MTPLX. Chat-only models are not compatible even when their server implements an OpenAI-shaped API.

Install

Install roboalchemist.reticle from the VS Code Marketplace or Open VSX:

code --install-extension roboalchemist.reticle

For an offline install, download the VSIX from the latest GitHub release, then run Extensions: Install from VSIX... in VS Code.

Quick start with MTPLX

On an Apple Silicon Mac with at least 16 GB of unified memory:

brew install roboalchemist/tap/reticle-mtplx
reticle-mtplx install
reticle-mtplx status

Then configure VS Code:

{
  "reticle.baseURL": "http://127.0.0.1:8000/v1",
  "reticle.model": "mtplx-qwen35-9b-optimized-speed",
  "reticle.fimFormat": "qwen",
  "reticle.maxLines": 8,
  "reticle.maxTokens": 64,
}

Run Reticle: Test Autocomplete Endpoint. See the complete MTPLX guide for monitoring, logs, dashboard, restart, custom models, and uninstall.

Quick start with Ollama

1. Install Ollama, then pull the FIM-trained Base checkpoint:

   ollama pull qwen2.5-coder:1.5b-base
   ollama serve
   

2. Verify the model/server pair before configuring Reticle:

   curl --silent --show-error http://127.0.0.1:11434/v1/completions \
     -H 'Content-Type: application/json' \
     -d '{
       "model": "qwen2.5-coder:1.5b-base",
       "prompt": "function add(a, b) {\n  return ",
       "suffix": "\n}\n",
       "max_tokens": 32,
       "temperature": 0,
       "stream": false
     }'
   

   choices[0].text should be a bare insertion such as a + b, without a rewritten function, explanation, or Markdown fence.

3. Install the Reticle VSIX, open VS Code Settings, and set:

   {
     "reticle.baseURL": "http://127.0.0.1:11434/v1",
     "reticle.model": "qwen2.5-coder:1.5b-base",
   }
   

4. Run Reticle: Test Autocomplete Endpoint from the Command Palette. Then type after return in a source file. Reticle shows ghost text; press Tab to accept the entire single- or multi-line suggestion. Suggestions trigger automatically after a short typing pause. To force one manually, press Option+\ on macOS or Ctrl+Alt+Space elsewhere.

Disable other inline-completion extensions while validating so their ghost text is not mistaken for Reticle's.

The compatibility litmus

Use the same request with every provider, changing only the URL, model ID, and optional authorization header:

curl --silent --show-error https://HOST/v1/completions \
  -H 'Content-Type: application/json' \
  -H "Authorization: Bearer $API_KEY" \
  -d '{
    "model": "MODEL_ID",
    "prompt": "function add(a, b) {\n  return ",
    "suffix": "\n}\n",
    "max_tokens": 32,
    "temperature": 0,
    "stream": false
  }'

Compatible output has a choices[0].text insertion like a + b. Once that passes, repeat with "stream": true; Reticle uses the streaming path. A server may accept suffix while silently ignoring it, so an HTTP 200 alone proves nothing.

Configuration

Reticle requires HTTPS and an API key for non-loopback endpoints. It never logs response bodies from HTTP errors, because providers can reflect credentials or request content.

Provider guides

• Ollama — recommended first setup; its OpenAI compatibility documents suffix.
• MTPLX — managed Apple Silicon service with health, metrics, dashboard, and Qwen PSM transport.
• llama.cpp — excellent FIM runtime, with an important /infill versus /v1/completions caveat.
• OMLX — Apple Silicon serving and the archive's fastest Mac-local model result.
• LM Studio — GUI/headless local server; verify suffix mapping for the exact version and model.
• Remote hosted provider — HTTPS, key handling, and suffix-support checks.

See the evidence-based model compatibility table before choosing a checkpoint.

Troubleshooting

• Prose, an explanation, or fenced Markdown: the model is behaving as chat, not FIM. Select a Base/FIM checkpoint and rerun the litmus. Fence stripping cannot make a model suffix-aware.
• The model repeats or rewrites the suffix: the server ignored suffix or used the wrong FIM serialization. This model/server pair is not compatible through Reticle's transport.
• Empty output: verify the exact model ID with GET /v1/models, inspect server logs, increase reticle.maxTokens, and rerun the non-streaming probe.
• Remote configuration error: use an https:// base URL and set reticle.apiKey. Reticle intentionally rejects insecure remote HTTP.
• Suggestions are slow only on the first edit: cold prompt prefill can take seconds. Keep the model loaded; cache-aware servers benefit from Reticle's stable per-model/per-document session header.
• No suggestion in one language: check the allowlist and denylist. The denylist wins.
• Wrong extension's suggestion appears: temporarily disable Copilot and other inline-completion extensions.

Development

npm ci
npm run compile
npm test
npm run lint

The live integration test is explicit and loopback-only:

RETICLE_INTEGRATION=1 \
RETICLE_INTEGRATION_MODEL=mtplx-qwen35-9b-optimized-speed \
RETICLE_INTEGRATION_FIM_FORMAT=qwen \
npm run test:integration

Run the same live path inside a real VS Code Extension Development Host:

RETICLE_E2E_LIVE=1 \
RETICLE_INTEGRATION_MODEL=mtplx-qwen35-9b-optimized-speed \
npm run test:e2e

License

MIT. The optional context engine is an original lightweight implementation; no third-party context-retrieval code is vendored.