Skip to content

fix: boundary hook parses real-world boundary bullets; doctor accepts n/a (0.15.1) - #10

Merged
stid merged 1 commit into
mainfrom
fix/boundary-parse-and-doctor-na
Aug 13, 2026
Merged

stid merged 1 commit into
mainfrom
fix/boundary-parse-and-doctor-na

Conversation

@stid

@stid stid commented Aug 13, 2026

Copy link
Copy Markdown
Owner

Two fail-safe defects found by dogfooding LCD on a real work-item (private decision log D-023) — both hit while using the plugin, not while reading it.

1. The edit-boundary hook read a boundary differently than people write one

lcd_journal_boundary emitted each EDIT BOUNDARY bullet as a single entry, and the hook matched that whole string against the edited path. So both of these denied every edit to the files they list:

- `CLAUDE.md`, `README.md`, `.gitignore`     ← three paths on one bullet
- `PROPOSALS.md` (historical note only)      ← path plus a trailing note

The deny message names the boundary, so the author sees their own path in the list and is told it isn't in the list — with no hint that the format was the problem. It denied two legitimate edits in the work-item that found it.

Entries are now split on commas, and a trailing parenthesised note is dropped. Two things deliberately survive:

  • a brace glob ({src,lib}/**) keeps its commas — there they are pattern, not separator;
  • a (group) path segment (Next.js route groups) is untouched, because a note is only recognised after whitespace.

Not addressed here: bash case can't match a brace pattern at all, so {src,lib}/** as a boundary entry still matches nothing. That's a pre-existing matching gap, separate from this parsing fix.

2. lcd:doctor treated n/a as a mistake

A project with no test suite fills the test keys with an explicit n/a (the shipped doctor fixture already does this for eval:). The placement-vs-glob heuristic ran on those strings anyway and emitted a WARN that no edit could clear — permanent noise for docs/prose/decision-record projects. n/a, with or without a reason after it, now counts as filled in and skips the heuristic.

Testing

Tests first, then the fix. New fixture cases in tests/test-boundary-check.sh (multi-path bullet, annotated entry, brace-glob passthrough) and tests/test-doctor.sh (a fully n/a docs-only project → 0 FAIL, 0 WARN). Both failed before the change.

tests/run.sh → 10/10 · claude plugin validate . → passed.

Patch bump to 0.15.1 in plugin.json + marketplace.json (lockstep), CHANGELOG section added.

…cepts n/a (0.15.1)

The edit-boundary hook matched EDIT BOUNDARY entries as whole bullet strings, so a
bullet listing several comma-separated paths, or a path followed by a parenthesised
note, matched nothing and denied every edit to those files. Entries are now split on
commas with a trailing note dropped; brace globs keep their commas and '(group)' path
segments are untouched.

lcd-doctor ran its placement-vs-glob heuristic on explicit 'n/a' values, giving any
project without a test suite a permanent WARN it could not clear.

Both found by dogfooding LCD on a real work-item (D-023). Tests first: each fix has a
fixture case in tests/test-boundary-check.sh / tests/test-doctor.sh.
@stid
stid merged commit 0998412 into main Aug 13, 2026
3 checks passed
@stid
stid deleted the fix/boundary-parse-and-doctor-na branch August 13, 2026 06:00
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant