Investigation: can python-vibe be everyday laptop work?

Answer: no — not this 0.5B LoRA. Comfortable daily explore / edit / run needs a larger model, tool-use training, and a local editor connected to Ollama. Keep python-vibe-0.5b as a cheap draft + harness.

Related: research-vibe-review · local loop vs hosted agents · what to improve · issues #8, #9.

What everyday laptop work means

You open a repo and talk. The model explores, edits files, runs tests, and keeps a short plan. That is comfortable daily work. The weights need to be large enough to emit tool calls.

python-vibe today is a 400 MB style prior plus scripts:

The 0.5B model misses Action: lines. agent.py only feels daily-usable when --model is something like llama3.1:8b.

Measured gap

Held-out laptop tasks (LoRA + harness): weekday name, count .md, jsonl reader, tiny docstring apply — all failed (wrong main(), month-as- weekday, filtered the word "bad", junk docstring). Base qwen2.5-coder:0.5b failed the same class.

OpenSRE: 100 smallest first-party files (200–2500 bytes) → 100× “no issues”, 0 applied. That is not a review.

What will not work

• More steps on the 0.5B run (already overfit after step 100).
• More short stdlib pairs only (issue #9: this is a style prior).
• Asking the 0.5B weights to plan a repo.

What to do

1. This week. scripts/run/agent.py defaults to llama3.1:8b. Local editor: local-editor.md. scripts/run/openai_compat.py proxies /v1/chat/completions. scripts/weights/export_ollama.py --create names python-vibe-everyday.
2. Your model. scripts/weights/build_agent_data.py writes seed tool traces (data/agent-loop). scripts/weights/train.py --everyday is the 7B-class LoRA. Append redacted explore / edit / run sessions before claiming 2k traces. Fuse/GGUF: export_ollama.py --from-gguf.
3. Eval. scripts/measure/eval_everyday.py (offline in CI). --live must beat untuned 8B on parse rate before anyone says everyday-ready.

0.5B stays public for download, CI, and the harness demo. It is not the everyday brain.

Shipped in this repo (laptop path)

• scripts/run/agent.py defaults to llama3.1:8b. --tiny is the sidecar.
• scripts/run/openai_compat.py + local-editor.md for a local OpenAI-compatible editor.
• Seed tool traces + --record → data/agent-loop/extra.jsonl (gitignored).
• scripts/weights/train.py --everyday (7B-class MLX). export_ollama.py --create names the stand-in; GGUF of your LoRA is --from-gguf.
• scripts/measure/eval_everyday.py: gold weekday + count-md /run, ≥1 KB NameError fixture, Action: parse fixtures. --live on this machine (29 Aug 2026): llama3.1:8b parsed 2 / 3 prompts (above the 50% floor). 5 September 2026, same machine: harness parse 11 / 15 vs clean 0 / 15; harness failed the ≥1 KB logic fix 0 / 3 (two writes were tests only). After #229 (same evening): parse 10 / 15 vs 0 / 15; fix still 0 / 3, writes [] × 3. After #238 (refuse explore): parse 11 / 15 vs 0 / 15; fix still 0 / 3, writes [] × 3. After #246 (same evening): parse 9 / 15 vs 0 / 15; harness fix 3 / 3 with no model turns (the harness wrote the sum) while a clean one-shot passed 3 / 3. A whole-line return 0 on a named sum is compiler-class, like subtotl and page=. That ≥1 KB cell is retired as a model job. Do not remasure it. The live cell is clip in eval/fixtures/everyday_live (filter vs clamp), scored only when turns are non-empty. After #254 (same evening): parse 8 / 15 vs 0 / 15; harness fix 0 / 3 (steps × 2, done × 1; writes [] × 3; turns non-empty) while a clean one-shot passed 3 / 3. The model ran. It did not write clip. Everyday-ready is still not met. Replay: PYTHONPATH=src python scripts/measure/eval_everyday_bar.py.

Live agent.py + llama3.1:8b loops on this machine (29 Aug 2026):

1. NameError fixture copy: read → Find: return tota → tests OK → done (4 steps).
2. This repo: patched scripts/run/agent.py docstring to python3.13.
3. Failed: full-file edit wiped tests/test_agent_tools.py (20% length guard was too weak). Guard is now 2/3 of original; file restored by hand.
4. This repo: patched resolve_project_file to allow .md.
5. This repo: patched README agent example to python3.13.

The loop works on scoped patch tasks. It is comfortable daily work on small, well-scoped jobs — not a full-repo rewrite.

Small vs large (29 Aug 2026)

Same CLI, two briefs (no extra model):

• Small (≤40 first-party text files, ≤200 KB): inject the file list. Questions → read → Action: done. Bugs → patch → run. This is the everyday laptop path.
• Large: inject top-level counts, require Action: map, --scope, and truncated grep. Do not ask the 8B to read the whole tree.

PYTHONPATH=src python3.13 scripts/run/agent.py --project /path/to/app --brief prints the mode without calling Ollama.