Hooks work here because they move the obligation to the side of the ledger that gets checked. The author's diagnosis of why the prompt line fails is exact: the task is enforced by the point of the session, and the devlog instruction is enforced by nothing [5]. Read the three hooks as a schedule, though, and they do not cover the session evenly.
The nudge does not speak until the fifth turn [7]. The stop gate trips at three or more Write, Edit or NotebookEdit calls with no matching wiki update [8]. Three edits fit inside four turns, so an agent that opens with a burst of file writes reaches the blocking gate having never seen a reminder [1]. The gap in the other direction is the one I would worry about. Two edits and no entry still closes clean, since two sits under the threshold [2]. Small changes that carry real decisions are exactly what the counter waves through.
Then there is what the gate reads. It clears on evidence that a devlog or index file was touched, not on evidence that a reason was written down [3]. An entry reading "refactored the queue" satisfies it. The entries worth having are the ones that say why retry logic gives up after three attempts instead of five, or which approach was tried first and thrown out [11], and no hook can count those. The fallback where a blocked agent must instead explain why this session needed no entry [10] is the good part of the design, because the explanation is itself a record. (An entry documenting why no entry was needed is still an entry.)
The author says the Cursor path is weaker. There is no transcript path exposed the same way, so the check degrades to whether anything in the wiki directory changed in the last two hours [9]. The vault is shared across both machines by livesync [4]. A devlog entry written on the Claude Code box twenty minutes ago therefore satisfies the Cursor box's gate for work it never inspected [4]. On one machine that is a reasonable proxy, but across two synced machines it is a timestamp some other process can set.
For any of this to transfer, your harness has to expose three things:
1. A session-start injection point that can front-load prior entries [6]. 2. A hook that can fire on a fixed turn cadence [7]. 3. A stop event a hook is allowed to refuse, with a transcript to count against [8].
Claude Code has all three, down to the literal `{"decision": "block"}` response the script returns [8]. Cursor has the first two plus a two-hour window [9].
That is the adoption cost. It is why I would build the vault before the enforcement. The vault is plain files that outlive whichever agent I am paying for this month [4]. The hooks are per-tool, and they break when the tool's hook contract changes.
