Use python-vibe from an editor

Three easy paths. All stay on 127.0.0.1 unless you choose otherwise.

python3 scripts/run/install.py then source .venv/bin/activate so python-vibe is on PATH (macOS often has no pip). Activate in every new terminal or the shell says command not found. --project defaults to the current folder. Files land in .vscode/, .continue/, or .cursor/ inside your app. This repo already ships .cursor/mcp.json.

Drop-in sources: editors/.

1. Pull the everyday brain

ollama pull llama3.1:8b
# or: ollama pull qwen2.5-coder:7b
# or: ollama pull qwen2.5-coder:14b

2. Easiest: tasks in the integrated terminal

python-vibe editors vscode --project /path/to/your/app

Then Run Task and type a task, for example:

• what does compute_total return?
• write a weekday script from argv
• fetch json from the HTTP API
• tally counts by key from a csv
• implement binary search

The same tasks.json works in VS Code and in other editors that read .vscode/tasks.json.

3. OpenAI-compatible chat (brain only)

Ollama already exposes:

http://127.0.0.1:11434/v1/chat/completions

A localhost proxy that defaults to the everyday model (and warns if you pick 0.5B):

PYTHONPATH=src python scripts/run/openai_compat.py
# http://127.0.0.1:8081/v1/chat/completions

Or let the write limit apply to chat (writes off unless --allow-writes):

python-vibe serve --project /path/to/your/app
# GET  http://127.0.0.1:8090/v1/models
# POST http://127.0.0.1:8090/v1/chat/completions

In the editor’s OpenAI-compatible settings:

• Base URL: http://127.0.0.1:8081/v1 (proxy) or http://127.0.0.1:8090/v1 (harness)
• API key: ollama (any non-empty string)
• Model: llama3.1:8b

Some hosted editors send the OpenAI request from a remote backend. Those cannot see 127.0.0.1. Do not open a public tunnel to it. Use tasks or the local MCP instead.

4. Cursor / local MCP (write limit, no tunnel)

python-vibe editors cursor --allow-writes

Cursor launches python3 -m harness mcp --project ${workspaceFolder}. Tools: ask (read-only) and run (writes if you passed --allow-writes). Stdout is JSON-RPC only. Step-by-step: Cursor.

This is the editor calling python-vibe. It is not an Action the 8B may emit.

5. CLI (same write limit, no editor)

python-vibe run /path/to/your/app "write tests for apply_discount"
python-vibe run /path/to/your/app --scope src "what does apply_source refuse?"

--tiny / --engine mlx is smoke only.

What python-vibe is good at

Kit skills for everyday laptop Python (stdlib, AAA tests):

Each write is followed by write-tests (test_<unit>_<result>, Act into got).

Optional: a Hub GGUF that Ollama does not ship

OpenCoder 8B and SWE-agent-LM 7B are on Hugging Face, not in the Ollama library. Import the Q4_K_M file, then pass --model:

python3 scripts/weights/import_hf_ollama.py --name opencoder
python3 scripts/weights/import_hf_ollama.py --name swe-agent-lm
python-vibe --model opencoder:8b run "add a function clamp and a unit test"

Default stays llama3.1:8b. Detail: Hub models.

Optional: your LoRA as GGUF / Ollama

Stand-in (this week): export_ollama.py --create is FROM llama3.1:8b plus the agent system prompt. That is not a trained python-vibe-8b.

After you fuse a 7B-class MLX adapter to a folder:

1. Convert with llama.cpp convert_hf_to_gguf.py (not in this repo).
2. PYTHONPATH=src python scripts/weights/export_ollama.py --from-gguf fused/everyday.gguf --create

Do not call this everyday-ready until scripts/measure/eval_everyday.py --live beats untuned 8B on Action: parse rate.