Measurement culture

The engineering rules for this repository live in AGENTS.md, and they are not opinions. Each one names the probe, the version and the number that produced it, and the test that keeps it true. The evidence itself, including the dates, the superseded corrections and the probes that produced each verdict, lives in docs/agents-history.md, and every rule carries an (h #gNN) marker pointing at the section that holds it.

That file is deliberately not a page on this site. It is the maintainer’s internal log, it is addressed to maintainers and to agents working in the repository rather than to users, and rendering it here would mean rewriting it. Read it on GitHub.

Three rules, quoted as they stand, to show the shape:

An abort that lands while a turn is being PREPARED must stop it, and the only thing that can see it is a watch created before the prologue’s first await (watchTurnAbort, src/turn-abort.ts): addEventListener("abort") on a signal that already aborted never fires, so the stream’s own handler, which cannot exist until the prologue has finished, missed every stop in that window and the turn spawned, wrote and billed anyway (measured: a turn alive at 5,874 ms against a 5,000 ms signal). (h #g182)

A finish’s usage is context occupancy to opencode, so its input side is the turn’s LAST real API call, never result.usage, which sums every call: a 14-call turn at 180K real context reported 2,050,806 cache read and opencode auto-compacted. (h #g169)

LRU eviction must never take a process that is mid-turn: evictIfNeeded walks insertion order for the first with turnInFlight !== true, and when all are busy evicts nothing and warns. Do not “restore” the one-liner. (h #g62, #g179)

The practical consequence for anyone reading this site: where a page states a number, a version or a date, it is there because someone measured it on that version on that date. Where a page says something is not verified, that is not hedging, it means nobody has run it yet.