24-hour review, 2026-10-16

Requested by the owner ("24 hour review"). Window: the 24 hours to about 21:00 UTC on 2026-10-16. Source: prompts/runs/ (224 records), promptctl.py measure, runcost.py report, git log origin/main.

What happened

• 148 merges to main; 224 recorded runs. By track: research 88, ventures 57, process 32, redteam 17, explore 16, review 12.
• Success: research 88/88, explore 15/16, redteam 15/17, review 11/12, ventures 47/57, process 18/32.
• Cost (measured runs only, API-equivalent at list price): $143.83 across 129 runs. Research $49.51, process $56.44, ventures $21.75, explore $7.59, redteam $5.24, review $3.30.
• Leaked work: none. Of today's records, 143 PRs merged, 78 name no PR (mostly doc or upstream runs), and one closed unmerged (#7330). Both of #7330's changes, the help-keeper widget and the percent-aware input-bounds check, reached main through other PRs.
• Audit: shift-roster-builder, the most-changed venture (7 merges), re-run: 64 of 64 tests pass, with its disclaimer and usage counter in place.

What is off

1. Twelve reviews in one day. Most changed nothing. Already fixed at 20:09 (E28: review_every 20 → 40); this run checks the effect later.
2. The process track costs the most per success. $3.76 per successful run against $0.47-$1.11 elsewhere; 14 of 32 runs ended partial. The partials say why: backlog items too big for one run, a reverted solve, and items already done by someone else. Proposal G2 in prompts/experiments.md asks whether to halve its weight.
3. One environment problem caused a third of the friction. 66 reports named CLAUDE.md. About 20 were the same thing: python3 on PATH is not the interpreter pip or apt install into, so pytest and numpy "install" and then cannot be imported. About 14 more were commands that outlast the 2-minute tool limit.

What this review shipped

• platform/ci/pyenv.py: PY=$(python3 platform/ci/pyenv.py pytest numpy) gives a working interpreter on any image layout (7 tests). Linked from CLAUDE.md, verify-and-merge.md and research/CLAUDE.md.
• Background-it notes on the four slow commands in CLAUDE.md.
• prompts/derived/ refreshed (derive.py fetch and build).

Owner question

• G2: halve the process track's weight (0.10 → 0.05) until it passes 80% success over 20 runs? One line in prompts/routing.json.