What to improve

The product gap is not closable. The harness gap is.

A hosted IDE agent has native tools, extra servers, a browser, and a large context window. python-vibe is a local loop: twenty typed Actions, a write limit, twenty steps, 700 tokens, project text files only (Python plus a few config suffixes; no secrets). Improving this project means making the 8B loop reliable on small Python and platform trees — not growing a browser Action or pretending 8B is a frontier model.

Related: local loop vs hosted agents · harness comparison.

What to copy, what to refuse

Published harness notes in this repo already said the quiet part: edit format and context assembly move small-model pass rates. A free shell tool does not transfer to an 8B on a laptop working tree.

Closest score on the same jobs

Score is “would a daily user get the same outcome,” not model size. 0–5. “After harness” is the recommended local work, not a new weight.

Work already in the tree

Ship these before training another model.

1. Design loop. Wired. After each one-split edit the harness re-scans. done is refused while findings remain. Review tasks may edit. Live 8B still aims at the god module; refuse_god_target now blocks that path before the draft runs.
2. Auto-pick review-design, refactor-split, and readable-layout. Thin-review refuse is in the done handler.
3. Verify writes. Add / bug / rename / refactor / a design-loop write: inject tests or run. done is refused until a passing unittest.
4. Deeper small-file reads. Files under 12 KB are read whole. Larger files still truncate at 3500 characters plus a tail.
5. Measure bigger local models. 29 Aug 2026, this laptop: 8B first Action was done on complete (thin summary), patch on add-multiply (tests file first, not impl), read on a dirty design review (skills were auto-picked; prelude asked for edit). The on-disk 30B coder timed out at the 180s Ollama cap. 7B coder is not pulled. Default stays 8B.
6. Raise the live parse floor. eval/action_prompts.jsonl has ten rows. Live 8B parse is 7/10. Everyday-ready still means beating an untuned 8B on parse and a real ≥1 KB fix.
7. AAA tests. write-tests now requires test_<unit>_<result> and got = …. One-line assertEqual(fn(), n) is refused.
8. Editor drop-in. python -m harness editors vscode|continue|cursor copies tasks / Continue / local MCP. serve speaks /v1/chat/completions. Hosted-editor chat override of loopback is out of scope (no public tunnel).
9. Everyday laptop Python. Skills write-script, call-http, analyze-data, write-algorithm. Prelude now quotes the skill Path: (pkg/weekday_name.py, not pkg/<noun>.py). Drafts that emit curl / wget / os.system are refused.
10. Compiler oracle. Undefined-name scan (scan/names.py). done is refused if a bugfix file still has subtotl. Tests in an impl file are refused. A rename that still defines the old name is refused. See small models, classic development.
11. Platform paths. Skill write-paths. The limit includes .toml / .yml / .json. Drafts that use os.path.join, a hardcoded home or /tmp, or a POSIX-only venv path are refused. See platform engineering.
12. Traces, then a 7B LoRA. Only after a live design loop reaches no structure findings. --record into data/agent-loop/extra.jsonl (gitignored). Thirty seed rows are not enough. Decision write-up: fine-tune or harness.
13. Named-file review quotes the compiler. Wired. review src/orders.py no longer asks for a patch, then refuses it. Undefined names finish the run with no generate. Measured: same jobs, same evening.
14. New functions stay with related names. Wired. pick_module no longer prefers the largest file (that was the controller). Prelude pins Path:. After the def exists the harness writes the AAA test. done is refused until def <symbol> exists. A second orders.py is refused.

0.5B stdout, estimated

The 18-script MLX pair is 0.5B sample-and-run. Four drafts plus one repair is 9 / 18. The later loop (datetime prepend, “stdout is wrong,” one 8B hint) is 12 / 18. Zero of those twelve were a hint-repair. Greedy LoRA is 0 / 54.

Daily run stays an 8B. Do not spend the week on more 0.5B pairs or on another 8B-hint cell for this board.

What not to spend a week on

• More 0.5B train steps. The adapter is a style prior. Held-out vibe tasks failed. It misses Action: lines. Greedy LoRA scored 0 / 54 on the 18-script exec eval.
• Training python-vibe-8b on the thirty seed rows and calling it everyday-ready.
• A bash tool, a browser Action, or extra-tool bridges. Those make the laptop write limit weaker and do not move the measured jobs.
• Raising --steps as a substitute for a review → one-split → review loop.

Two success bars