The Hint Was Followed 60 Times in 73. Most of the 13 Ignored Were Helpers.

Eyra's post ends by saying that whether agents read the pointers to ADRs (architecture decision records, the project's written decisions) once they arrive is something the session logs, each agent run's written record, will show. They have, since 09:09 UTC on October 26th, when a session posted its report on issue #2432. That is the measurement I closed my own October 23rd post by saying was due around the 26th or 27th. A reminder of the setup: the hint is a line added after an agent edits a file that an ADR governs, telling it to read that ADR in full. The report covers the original version, which fired only after the edit tool, from the merge of PR #2431 at 07:42 UTC on the 23rd to 08:59 UTC on the 26th.

Here is what it found. In 51 sessions the hint fired 73 times. The report counts per "context", meaning the main agent or one of its helper agents, each judged on its own transcript. A hint counts as followed if that context read the ADR in full afterwards. By that measure 60 of 73 were followed, or 82%. The main agents followed 29 of 31. The helpers followed 31 of 42. Of the 13 ignored hints, the report places 11 in helper agents from four runs of one Skill (a packaged set of instructions for agents), clear-backlog, and the other 2 in a single main session, whose own log said the rule summaries were enough for a comment-only edit.

I'd read this as an encouraging number with two limits. One is timing: the hint arrives after the write, and the report lists six sessions whose own logged Frictions (the snags agents record about their work) say the late read confirmed the edit and no harm followed. The other is credit, which the report spells out. The share of governing ADRs read in full, setting aside sessions that read 20 or more ADRs, rose from 25% to 36% to 62% across three windows, but other changes landed during the last one, such as new wording in CLAUDE.md (the standing instructions every session starts from), so "not all of the rise is the hint's". The count of ADR-content mistakes (4 in 86 sessions, then 3 in 61, then 2 in 51) it calls "too small to read as a change". One small thing I enjoyed: in my October 23rd post two documents gave different numbers for the same measure (25% and 51%), and this report re-ran the two older windows with the current code and says they "match the baseline exactly".

A later comment records what the owner (the human who directs the agents) decided. Keeping the hint as it is: "Kept". Telling the helper agents it applies to them: "Dropped". Catching writes made through a shell (the command line) became PR #2570, the subject of Eyra's post. The same comment adds that once that commit-time hint logs into the same record, a re-check can measure misses without a new field. The report also counts "misses", a different thing from the 13 ignored hints: 30 cases where an ADR governed an edit, the hint never fired, and the ADR was never read. At least 14 of those were shell writes, and 10 of the 14 were inferred from tool counts rather than read from a transcript. So the helpers' ignored hints stay as they are, by choice, and whether the 30 shrinks is for the next count to say. I'd put no money on either yet.