Kuzz007 34c43c8a9d fix: address the automated review round on PR #6154 ...

- Replace parseGeodataFile's full proto.Unmarshal with a protowire-based
  scan that reads only each entry's Code, skipping every Domain/CIDR
  payload without allocating it -- the actual bulk of a real
  geoip.dat/geosite.dat. Also caps the file read at 256 MiB.
- Hold geodataMu across the full scan-and-maybe-parse in
  GetGeodataCategories instead of releasing it around the parse, so
  concurrent cache misses (e.g. several browser tabs) can't all
  independently re-parse every file; clone the cached slices before
  returning them so a caller mutating its result can't corrupt the cache.
- Gate useGeodataCategories on the rule editor's own `open` state instead
  of firing on every visit to the Routing tab.
- formatGeodataSuggestion now compares filenames with strings.EqualFold,
  matching scanGeodataFiles' own case-insensitive match -- a file that IS
  the default one on a case-insensitive filesystem (e.g. Windows) no
  longer gets the long ext: form.
- Fix a real bug the review's hypothesis led to: Select mode="tags" only
  commits the search text on Enter/comma, so clicking Save right after
  typing (a blur, not an Enter) silently dropped the value entirely, with
  no domain/ip key at all in the saved rule. Wrap it in a small
  TagsAutocomplete that also commits on blur. Same autocomplete now
  applies to sourceIP, which accepts geoip:/ext: too.
- Guard useGeodataCategories' fetch per-field with Array.isArray instead
  of a single top-level `?? EMPTY_CATEGORIES`, since parseMsg returns the
  original unvalidated obj (not null) on a schema mismatch.
- Test fixes: exact slices.Equal instead of slices.Contains-only
  assertions, t.Run subtests, a cache-hit-skips-reparse test (via a
  test-only parse counter), a returns-independent-slices test, a
  file-size-cap test, and four new frontend tests covering the tags
  round-trip including the blur-commit regression above.
- GeodataCategories now goes through the same generated-example path as
  every other response type (StructAllow + example: tags + responseSchema
  in endpoints.ts) instead of a hand-written response string. The
  existing hand-written GeodataCategoriesSchema in schemas/routing.ts is
  unrelated to this and is left alone -- CLAUDE.md is explicit that Zod
  schemas under src/schemas/ are the source of truth and only the
  generated example/openapi path comes from Go example: tags.
- Drop the two PR-illustration screenshots from media/ -- nothing in the
  repo referenced them; they only ever needed to exist in the PR
  description itself.

Not changed: leaving geodataFileKind's leak into generated/{types,zod}.ts
as-is. internal/web/service's openapigen request has no AliasAllow at
all, so every non-struct type in the package already leaks this way
(e.g. staticEgressResolver, transportBits predate this PR) -- scoping an
AliasAllow for the whole package is a real cleanup but a separate, wider
change than this PR's own footprint, and needs checking nothing already
depends on those existing generated aliases first.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

2026-07-30 01:18:00 +03:00

329 KiB

Raw Blame History