Using posttrainllm with Continue.dev / Cline / Aider

posttrainllm exposes an Ollama-compatible HTTP surface so it drops straight into Continue.dev, Cline, and Aider configs as a local provider. Setup + caveats.

Using posttrainllm with Continue.dev / Cline / Aider

posttrainllm serve exposes both an OpenAI-compatible surface (for lm-evaluation-harness, langchain, etc.) and an Ollama-compatible surface (for the dev-tool ecosystem). The Ollama surface uses the same generation core but emits NDJSON streaming chunks per the Ollama protocol — Continue.dev, Cline, and Aider all configure against provider: ollama and “just work” with posttrainllm as the backend.

Quickstart

1. Start the server on Ollama’s conventional port:

   posttrainllm serve <model.tinygpt> --port 11434
   # → listening on http://127.0.0.1:11434

2. Sanity check the Ollama surface:

   curl -s http://127.0.0.1:11434/api/tags
   # → {"models":[{"name":"posttrainllm:latest", ...}]}
   
   curl -s http://127.0.0.1:11434/api/generate \
        -d '{"model":"posttrainllm","prompt":"Hello","stream":false}'
   # → {"model":"posttrainllm:latest","response":"...","done":true,...}

3. Point a client at it (see per-tool sections below).

Continue.dev

Add to ~/.continue/config.json:

{
  "models": [
    {
      "title": "posttrainllm",
      "provider": "ollama",
      "model": "posttrainllm:latest",
      "apiBase": "http://127.0.0.1:11434"
    }
  ]
}

Reload Continue (Cmd-Shift-P → “Continue: Reload”) and the posttrainllm model appears in the model picker. Chat, autocomplete, and edit all route through posttrainllm’s local model now — code never leaves device.

Cline (VS Code)

Cline’s settings → API Provider → “Ollama”:

That’s it. Cline’s Plan-mode dialogue + Act-mode tool-call enforcement both work against the posttrainllm local model.

Aider

aider --model ollama/posttrainllm:latest --openai-api-base http://127.0.0.1:11434

Or in ~/.aider.conf.yml:

model: ollama/posttrainllm:latest
openai-api-base: http://127.0.0.1:11434

Aider’s “architect mode” can use posttrainllm for either the planner or the editor; configure two model entries if you want different specialists for each role.

Endpoints exposed

Plus the OpenAI surface on the same port:

• GET /v1/models, POST /v1/chat/completions, POST /v1/completions

Same server, two compatible wire protocols.

Wire format details

Streaming = on by default for /api/chat and /api/generate (matches Ollama’s behavior). Pass "stream": false for one-shot responses.

When streaming, the wire format is NDJSON — one JSON object per line, terminated by a chunk with "done": true:

{"model":"posttrainllm:latest","created_at":"...","message":{"role":"assistant","content":"Hello"},"done":false}
{"model":"posttrainllm:latest","created_at":"...","message":{"role":"assistant","content":" there"},"done":false}
{"model":"posttrainllm:latest","created_at":"...","message":{"role":"assistant","content":""},"done":true,"done_reason":"stop"}

This is not SSE (no data: prefix). The OpenAI surface on /v1/chat/completions is SSE — pick the right port + endpoint for your client.

Tool calling

Today: posttrainllm’s Ollama surface does not emit tool-call chunks (message.tool_calls). Continue/Cline tool calling requires the OpenAI surface (/v1/chat/completions with tools: [...]) or the agent runtime path (posttrainllm agent --tools tools.json).

Roadmap: once the tool-call extractor (mini-router) ships (see docs/roadmap/north_star_refined.md), the Ollama surface will embed tool calls in the streamed message.content per Ollama’s own format. Tracked as a Wave 2.6 follow-on.

Tradeoffs vs the cloud providers

You probably came here because you wanted to keep code on-device. Honest framing:

The right mental model: posttrainllm handles routine asks (rename a symbol, write a small function, explain a snippet) instantly + privately; escalate to cloud (via posttrainllm’s --cloud-escalate flag, or by switching providers in Continue) for the hard ones.

Limits + bugs to know about

• posttrainllm serve is single-threaded for inference (one model, one request at a time). Multiple concurrent Continue requests queue.
• Tokenizer mismatch: Continue may count tokens with its own tokenizer and over-/under-shoot budgets. Use generous num_predict budgets on the client side.
• --port 11434 collides with a real Ollama install. Either stop Ollama (ollama serve shutdown) or pick a different port + set apiBase accordingly in the client config.
• Tool calls don’t round-trip yet (see above).
• Streaming cancellation works (close the socket → generation stops; see commit c11265b for the SIGPIPE-safe wiring).

What’s next

Once the tool-call extractor (mini-router) lands, this page becomes the canonical “use posttrainllm in your editor” walkthrough — with tool-call support, it’s the first real-user-visible product surface for the project. See docs/progress.md for shipping status.