Local Code Assistant Harness MVP

Minimal local-first OpenAI-compatible harness for VS Code Continue.

Continue talks to this local server:

For local: http://127.0.0.1:8000/v1
For LAN:   http://serverip:8000/v1

The harness rebuilds a bounded prompt for every request, stores project memory locally, and forwards each request to a stateless OpenAI-compatible Devstral MaaS endpoint.

Quick Start

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp config.example.yaml config.yaml

Edit config.yaml:

• maas.base_url: your Devstral OpenAI-compatible /v1 URL
• maas.api_key: preferably ${DEVSTRAL_API_KEY}
• maas.default_model: your Devstral model name
• workspace.root: local project path to expose read-only to the harness

Then run:

export DEVSTRAL_API_KEY="..."
export DEVSTRAL_MODEL="devstral"
bash scripts/run.sh

Continue Config

Use continue/config.yaml.template as the Continue config starting point.

Set Continue's model API base to:

http://127.0.0.1:8000/v1 
http://serverip:8000/v1

Use any placeholder API key for Continue, for example local, because the harness owns the MaaS key.

Native Harness Endpoints

GET  /health
GET  /v1/models
POST /v1/chat/completions
POST /harness/projects
GET  /harness/projects/{project_id}
PUT  /harness/projects/{project_id}/memory
POST /harness/conversations
GET  /harness/conversations/{conversation_id}
PUT  /harness/conversations/{conversation_id}/memory
POST /harness/debug/context

Project/session metadata can be supplied through:

X-Harness-Project-ID
X-Harness-Conversation-ID

If headers are absent, the harness uses a default project and conversation for development.

MVP Feature Check

This MVP includes the three core pieces:

• SessionManager: resolves project_id and conversation_id, using defaults when Continue/Copilot Chat do not send harness headers.
• MemoryManager: loads hierarchical memory for the request.
• ContextPacker: rebuilds a fresh bounded prompt for every stateless MaaS call.

Memory hierarchy:

project memory
  -> conversation memory
      -> recent raw messages
          -> explicit files from workspace/raw

Raw Source Files

Put original read-only materials in:

./raw

The harness can include PDF and Word files as explicit context:

{
  "messages": [{"role": "user", "content": "Summarize the requirements."}],
  "context_files": ["raw/requirements.pdf", "raw/spec.docx"]
}

The MVP supports .pdf and .docx. Legacy .doc is intentionally not supported; convert it to .docx.

Tests

pytest

Without pytest installed, a basic route smoke test can be run after the server starts:

bash scripts/smoke.sh

The first tests focus on:

• context overflow prevention
• workspace path isolation
• OpenAI-compatible route behavior with a mocked MaaS gateway