mirror of
https://github.com/stablyai/orca.git
synced 2026-09-29 16:02:50 +00:00
Merge Sim
278596a3cd
INTERMEDIATE STATE - typecheck is red by design; stage 2 reconnects the seams.
Main rewrote this lane's subsystem (a65332a8bd moved structured chat onto the
Claude Agent SDK and made the launch layer agent-generic). Every conflict is
resolved TOWARD main so the merge is coherent, which means codex stays refused
on win32 for now - main's behaviour. That is deliberate: enabling codex before
the host probe is wired would leave a window with no gate at all.
Dropped `launch-structured-codex-session.ts` and its test. Main deleted that
module; our refusal->fallback mapping lived in it, and in the merged tree every
production caller already resolved to main's replacement - the only remaining
importer was our own test. The feature was dead code with green tests.
Main's replacement is also strictly better: it classifies by WIRE CODE via
isDefinitiveAgentSessionCreateRefusal and adds a separate
StructuredAgentSessionCreateUnknownOutcomeError, so an unproven create is not
treated as a definitive refusal. Ours matched error-message substrings and would
have called every matched throw definitive - abandoning a session that may
exist. Reference implementations agree: allowlist exact codes, treat timeouts
and disconnects as ambiguous, never act destructively on an unproven refusal.
The @ts-nocheck conflict was imports-only; resolved as a union and each of the
four symbols verified used by hand, since the compiler cannot check that file.
Remaining red, all ours to reconnect in stage 2:
- isWindowsProcessStartTimeAvailable no longer on StructuredAgentSessionRuntimeDeps
- host-handoff test predates main's HostHandoffAccess.findSession + journal API