Regenerate app-server protocol schemas for updated approval types and refresh VT100 settings snapshots to reflect the gpt-5.3-codex default model label.
Choose larger-context retry models from remote model metadata only, and ignore deprecated gpt-4.1* candidates.
When context windows tie, prefer the higher-priority model from the remote model list.
Add GPU device path exceptions for /dev/dri, /dev/kfd, and /dev/nvidia* when present so sandboxed commands can access GPU devices without granting broad /dev write access.
Include deterministic unit regression tests for GPU path discovery to keep behavior stable on hosts with and without GPUs.
fixes#546
Keep cycle-safe symlink traversal but drop scan depth and directory-count
ceilings that could omit valid skills in large/deep repositories.
Add regression tests to verify deep and high-cardinality skill trees are
fully discovered.
Add Codex-compatible skill discovery for repo and user .agents paths,
keep legacy .codex compatibility, and include optional /etc/codex/skills.
Follow symlinked skill directories for repo/user/admin roots with cycle-safe
bounded traversal, and add regression tests for roots, precedence, symlinks,
and legacy compatibility.
fixes#540
Fix /model selector rendering on small terminals by introducing a scrollable body window. Add regression test to verify the selected model remains visible when offscreen.
Align websocket beta header and stream completion handling with upstream behavior.
Emit completion on response.done/response.completed, surface wrapped websocket
errors without status, fail fast on response.failed, and abort websocket reader
tasks after parser exit to avoid lingering sockets. Update wire compatibility
fallback floor to 0.101.0 with regression coverage.
Introduce stdio/websocket transport abstraction for app-server, including routed outbound handling and websocket parity coverage.
Also extends protocol/config plumbing for initialize capabilities, web search external access control, and ChatGPT auth-token login support.
James Peter