Repository navigation
1.1.0-rc.3 — Private Editor history and resilient HTTP responses
Pre-releaseFixes for HTTP disconnect handling and hidden Editor history isolation. API contract 1, native protocol 16 and the existing 31 MCP tools are unchanged.
- Handle connection aborts/resets during JSON headers, response bodies and artifact downloads. Do not send a second response to a closed client or replay native work. Other I/O errors remain visible.
- Redirect only the owned Editor's Recent Files registry subtree, including remembered open documents, to a private volatile instance key before startup. Remove that private key after the exact owned process exits. Existing normal history is preserved; this is not full user-profile virtualization.
- Retain startup failure phase, Win32 code, remote result, bootstrap code and bounded log tails in private startup-failure evidence. Include this evidence in document-validation failures even after worker reset.
- Allow a continuous 15-second dispatch-marker grace interval instead of rejecting a newly created document window immediately.
Validation: 131 Python tests and ten native test executables passed. Two real hidden workers each opened, saved and generated/compiled a private map with zero compiler errors. Native history writes were observed in their private keys; normal history and source bytes remained unchanged. Both owned workers and private keys were removed. Observation recorded zero visible owned windows, foreground samples and input-desktop changes. A deliberately missing startup DLL retained phase load-local-bootstrap / Win32 126 and its suspended owned child exited. Frozen HTTP and MCP checks passed; packaged source/native hashes were verified.
The separate intermittent startup complaint still needs its original machine's logs; these local checks do not identify its cause. Terrain appearance was not changed, and no new Terrain-quality or full Data-audit acceptance is claimed.
Replace the complete toolkit package and restart its owned worker to apply this update. Existing running workers and separately pinned MCP integrations are not automatically upgraded. Stable 1.0.0 remains available.
See docs/COMPLAINT_FIXES.md for details, reproduction commands and limits.