2026-10-15, OCX at a1faf72aee (branch agent-tools): after runs 1, 2 and 3 and the faults they found, among them read's text now in a fenced block, exact, wiki_edit in place of wiki_write, and test running every test when given no name. Three models, both variants, 3 trials of each of the eight tasks.

Passed

What it says

• The fenced block took gemma4:31b's add-function from 0/6 to 6/6; wiki_edit took qwen2.5-coder:7b's wiki-status from 2/10 to 5/6.
• ok and next_step over all four runs, qwen2.5-coder:7b: 28 of 73 trials with them, 25 of 73 without; the larger models alike either way. No measurable effect; they stay, ok as the pass or fail the maintainer asked for and next_step at no cost (HISTORY.md decision 68).
• The 7B model cannot do the three harder tasks; the 30B models do all eight.

The records, every exact request, tool answer and reply of each trial, are attached as one tar.