[nasser1941]

Release 0.62.0: what the automatic checks tell the model, plus two fixes found on 0.61.0's gate run.

• Don't call a test run that stopped before its first test a failure of the change #250: a test run that stops before its first test no longer goes to the model, unless its output names a file the request changed. Checks run from the project's root, and a check that hits its time limit no longer counts as a failure.
• Don't count a failing test run piped into tail as passed #252: a failing pytest -q | tail no longer counts as passed. lcode then runs the checks itself.
• Count a SPARQL query it never ran like code it never ran #254: a SPARQL query the model never ran counts like code it never ran (lcode says it's done without running what it wrote #242).
• Make lcode bench --compare say when attempts are missing #253: lcode bench --compare warns when attempts are missing, and no longer treats one attempt as certain.
• Say how much disk a dev-slice run needs while it runs #251: the benchmark docs say how much disk a run needs, and not to clear uv's cache during one.
• Bump the python-dependencies group across 1 directory with 3 updates #227: test and lint tools updated.

#250, #252 and #254 change when the checks run and what the model is told, so this release goes through the gate (Releases in the contributing guide). The release candidate is this branch, built from its checkout, on the full dev slice at 32K with qwen3.6-35b, three attempts per task. It's compared with 0.60.0's full night on 4–5 November and with 0.61.0's runs on 5–6 November:

Within the noise against both. All 75 attempts ran, on the full slice with torch-tensor-parallelism back in. The run started at 08:29 on 7 November, in the daytime, at the maintainer's request. The model ran at 72–76 tokens/s every hour.

pass@1 went from 45% to 51% against 0.60.0, over the 25 tasks both ran (+5 points, noise ±10). Against 0.61.0 it went from 50% to 53%, over 24 tasks (+3 points, noise ±11).

	0.60.0	0.61.0	0.62.0
Attempts that passed	34/75	33/62	38/75
Checks failed at least once	27	—	9
Stopped after replies that went nowhere	9	1	1
Attempts out of time	5	10	9
Model speed, median attempt	76 tokens/s	68	73

lcode bench --compare exits 0 against 0.60.0. Against 0.61.0, it was run with that release's two runs combined. No attempt crashed inside lcode.

• Don't call a test run that stopped before its first test a failure of the change #250: 15 attempts saw a check stop before its first test and get the message instead of a note to the model. Checks that failed at least once fell from 27 to 9.
• Count a SPARQL query it never ran like code it never ran #254: the note fired in all 3 sparql-university attempts, and 1 passed (0/3 on 0.60.0 and 0.61.0). The note also fired in 4 attempts on other tasks.
• Tasks that moved, against 0.60.0: pypi-server 0/3 → 3/3, rstan-to-pystan 0/3 → 2/3, sparql-university 0/3 → 1/3, xarray-6721 and django-14787 2/3 → 3/3, django-12774 1/3 → 2/3, django-11555 2/3 → 0/3, openssl-selfsigned-cert 1/3 → 0/3, pytest-10051 1/3 → 0/3, and build-cython-ext 3/3 → 2/3.
• Out of time: 3 of the 9 are torch-tensor-parallelism, which now runs its code and needs most of its 15 minutes for that. 3 more are polyglot-rust-c (The model deliberates in circles instead of acting, and lcode gives up #241).

Three things on the machine differ from the earlier runs, and the comparison has to allow for them: Ollama is now 0.40.15 (lcode update upgraded it), swap is now 32 GB instead of 8 GB, and the Linux kernel was updated.

Real tests before the night

All with qwen3.6-35b at 32K. Daytime speed limited them, so they show how lcode behaves, not scores or time.

Test	Result
#250: django-12774, pytest-10051, build-cython-ext, 1 attempt each	All 3 reached a check that stopped before its first test. The message appeared, and no note reached the model.
#250: real terminal, 3 small projects	The checks run from the root after cd tests, and a change that really breaks the project still reaches the model
#252: a project with one failing test, and the model running pytest -q 2>&1 | tail -5	0.61.0 doesn't run the checks. This code runs them, and the failure goes back to the model.
#254: sparql-university, 2 attempts	1/2. Both ran the query on their own, so the new note hasn't fired with a real model yet.

pytest, ruff check, ruff format --check, basedpyright and mkdocs build --strict passed on each change. The same on this branch are running now.

This bumps __version__, the quickstart banner and the CHANGELOG (dated section and compare links).

🤖 Generated with Claude Code