Repository navigation
M3: Uri.EscapeDataString TFM divergence — refuse, normalize, or tolerate (R2-3) #32
Description
From the second verified review of PR #26 (finding R2-3, CONFIRMED). Policy decision needed.
Finding. Uri.EscapeDataString behaves differently across the locked TFM matrix: net472/netstandard2.0 throw UriFormatException for inputs over 32766 chars or containing lone surrogates, while net8+ silently substitutes U+FFFD. Same input: untyped crash outside the SDK exception taxonomy on one TFM, silently altered wire data on another. Reaches user input via SessionListRequest.Search, server-echoed cursors (OpenCodeRoutes.cs), and path ids; no guard normalizes or validates first.
Options to weigh: validate at the route boundary (refuse lone surrogates/oversize with a typed ArgumentException on every TFM — fail-closed and uniform); normalize deliberately (document U+FFFD substitution as a recorded tolerance); or wrap the net472 UriFormatException into the taxonomy. The fail-closed posture favors uniform refusal.
Trigger: adversarial or truncated user-supplied Search text on the net472 leg; must seal before any GA. No legitimate traffic hits it today.
Milestone: M3 (rides the net472 cluster with the ServicePointManager item).
Activity
docs: seal Q92 (ADR-0010) and open M3 — simplicity-first construction…
Blind-Striker commentedon Aug 15, 2026
Sealed as uniform route-boundary refusal (research log Session 24): lone surrogates and oversize inputs refuse with a typed ArgumentException on every TFM before escaping. Executes in the M3 net472 cluster (plan Arc 5), which the ADR-0010 singleton posture promotes to a GA gate.
Blind-Striker commentedon Aug 20, 2026
Implemented in b261014 and verified by hosted CI run 32350952168 on Linux, Windows, and macOS. Every generated path value and query/deep-object value now refuses inputs over 32,766 UTF-16 code units and lone surrogates before Uri.EscapeDataString. Existing path blank/dot-segment guards and empty opaque cursor behavior remain unchanged. Boundary, valid-pair, exact-limit, generator snapshot, and generated-source compilation evidence are green.