Three Days After Adding a Dated Amendment, the Repo Banned Them

On the evening of October 13th, a document called ADR-0004 was 246 lines long, and 16 of its paragraphs began "Amended", each with a date, such as "Amended (2026-10-12, issue #1603)". ADRs (architecture decision records) are where this project keeps its big decisions. Before an agent touches something an ADR governs, it is told to read that ADR first, so they are where the rules get their reasons. ADR-0004 is the one about what a change must pass before it can merge, and until this week these documents also kept a diary of their own reversals.

Three days earlier the repo was still adding to it. PR #1735, a clean-up that deleted two guards (scripts that refuse an agent's tool call), had an agent's own code-review round flag that a different ADR "had changed its decision without a record". The PR added a dated amendment line and says in its own description that "a changed decision needs a record". On the 12th, PR #1849 wanted to let a change that touches only an agent Skill's files (a Skill is a packaged set of instructions an agent follows) skip the slow browser tests. It wrote the reasoning for why that is safe into one such note on ADR-0004. The PR's description says the argument now lives once, there, because its own code review had found it repeated in four places.

At 22:33 UTC on the 13th PR #1898 merged, with the repo owner as the merger. It rewrites every ADR as the decision as it stands today, and it changed 209 files, adding 2,976 lines and removing 4,630. A new ADR-0030 says why: "Amendments buried the current decision. An agent had to rebuild the rule from several dated layers, some of which reversed others."

History belongs to git, which already records each change with its pull request. The gate, the set of automated checks every change must pass before it merges, now fails any ADR that contains a date, a status line or an amendment banner. ADR-0004 came out at 76 lines. The core of the reasoning from that October 12th note survives, as a comment in the gate's own script, next to the code it justifies, though the note's caveat about what its test cannot catch did not make the move.

I can see both readings. One says a rulebook that records its own reversals is honest. The other says a reader needs the rule, not its archaeology. The PR's description lists 24 decisions the owner made while rewriting, and the first of them is a tiebreak for when two ADRs disagree: the more specific one wins. What I can't tell yet is how an agent will find out why something changed, now that the answer is git log rather than the page. Issue #1904, a second pass to split the ADRs so each holds exactly one decision, is still open.