Repository navigation
pre-edit recall: recorded non-blocking observations from the #358 reviews #375
Description
Activity
One more for this list, measured during the final #358 review (same behaviour since the path-suffix rule was introduced, so not a regression):
A bidi control between a
/and the basename turns a path-style mention into a bare one and bypasses the path-suffix rule. Editingother/CLAUDE.md:INJECT see docs/<LRM>CLAUDE.md here <- another file's path, accepted reject see docs/CLAUDE.md here <- same text without the LRM, correctly rejectedbeforeis read as one raw UTF-16 unit, so an LRM there is neither/(the path branch) nor in the ASCII-only before-class, and the function falls through to the bare-mention accept. After #358 the same character is transparent AFTER the name and a delimiter BEFORE it. The completion is to step over a transparent bidi run backwards too, sobeforebecomes the/— taking care that the backward path-token walk starts from the right index.
Kept so the next review does not rediscover them. None is a wrong injection.
https://github.com/<o>/<r>/blob/main/CLAUDE.mdwhile editing rootCLAUDE.md) does not confirm: the path token starts at//github.com/…, which is a suffix of nothing. Missed recall only.CLAUDE.md𝐛.bak) still confirms, because the after-name class is ASCII-only for letters — the same documented decision that keepsCLAUDE.md然後recallable.file_pathis missing or not a string the hook returns before the guard pass. Correct for Edit/Write as they exist today; worth revisiting if the hook is ever registered for a tool that carries content under another key (see the matcher issue).error, file not marked seen). Accepted trade-off of treating anobservationsfault as a broken database; a per-candidate skip with a recorded count would be gentler.round N, finding Mcoordinates that resolve to nothing a reader can open. The substance of each comment is worth keeping; the coordinates could go in one mechanical pass.