Tom Boucher 80923244b1 fix(#2124): harden parsePhaseFromProse per orthogonal review (ReDoS + coercion) ...

Orthogonal security review of the Phase 1 surface found two issues; both fixed
and regression-tested:

- MEDIUM ReDoS: the name-extraction regexes /\(([^)]+)\)/ and
  /—\s*([^(\n]+?).../ backtrack O(n^2) on a crafted STATE.md field value with a
  long unterminated "(" / "—" run (reviewer measured ~38s at 320k chars).
  Length-bound both quantifiers to {1,200} -> linear (320k now ~100ms). A real
  phase name is far shorter than the cap.
- LOW: parsePhaseFromProse threw on non-string truthy input, unlike its three
  sibling #2121 functions. Coerce via String(value) up front.

The identical ReDoS regexes are copied verbatim from the pre-existing
state.cts:parseProsePhaseField; per the no-defer rule that surfaced defect is
fixed inline there too (Phase 2 / #2125 later supersedes that function by
delegating to the bounded phase-id.cts parser).

Adds a behavioral bound-guard regression test (a >200-char parenthetical is not
extracted) and a non-string-coercion test.

Refs #2124, #2121

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

2026-07-09 16:33:39 -04:00

28 KiB

Raw Blame History

View Raw