Development record
The product thinker wants a self-optimisation function built into abcd: it monitors CPU utilisation, token use and everything else that can be monitored, and optimises the setup over time. Today every tuning knob is a fixed number a person writes (the preflight slot count, the sub-agent ceiling, the pace window, fix rounds, the review set), and nothing measures the outcomes those knobs drive, so each value is a guess that never improves. The 2026-10-24 test run, seven sessions on one 10-core machine, showed the cost: the load average held near 25, tests timed out, a nitpick lane waited over an hour at the load gate, and the sub-agent cap was negotiated by hand between sessions. iss-2610241214563687 (an adaptive preflight slot count) and iss-2610140652230444 (reviewer catch rates measured per model) are two instances of this one capability.
The product thinker wants a self-optimisation function built into abcd: it monitors CPU utilisation, token use and everything else that can be monitored, and optimises the setup over time. Today every tuning knob is a fixed number a person writes (the preflight slot count, the sub-agent ceiling, the pace window, fix rounds, the review set), and nothing measures the outcomes those knobs drive, so each value is a guess that never improves. The 2026-10-24 test run, seven sessions on one 10-core machine, showed the cost: the load average held near 25, tests timed out, a nitpick lane waited over an hour at the load gate, and the sub-agent cap was negotiated by hand between sessions. iss-2610241214563687 (an adaptive preflight slot count) and iss-2610140652230444 (reviewer catch rates measured per model) are two instances of this one capability.
Frontmatter
| id | iss-2610241215547722 |
|---|---|
| type | issue |
| lifecycle | open |
| severity | minor |
.abcd/work/issues/open/iss-2610241215547722-the-product-thinker-wants-a.md
open on GitHub ↗ · commit history ↗
mentions2
- iss-2610140652230444Every review should run as a pair of models from two classes, with the top class only an occasional advisor. The classes are vendor-neutral: top, middle, the class below it, and smallest (the scale itd-2610100704270254 uses). For one vendor today, Anthropic, they are Fable, Opus, Sonnet and Haiku. The rule is stated in classes so it survives model churn and holds for any vendor whose models map onto them.
- iss-2610241214563687The product thinker wants the number of preflights a machine runs at once to follow the evidence, not stay fixed. On 2026-10-24 they set load.preflights to 2 in ~/.abcd.noindex/config.json, which applies to every abcd-managed repository on the machine, and asked that load monitoring keep it right: raised when the machine has room, lowered when it is overloaded. Today the count is a fixed number a person writes (default 1, at most 2), and the gate measures load only to decide whether to wait, never to learn: nothing records how long stages waited or ran, whether tests timed out, or what the load was, so no one can tell whether 2 is too many or too few.