You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
/api/health reports the server's build fingerprint. On launch the app compares it with its own
bundle and replaces a stale server process, so reinstalling never leaves the panel showing data
produced by the previous version.
Changed
Task state is now derived from each channel's own turn boundaries instead of a fixed 120-second
file-mtime window:
Codex task_started/task_complete, Claude Code stop_reason, DeepSeek turn/start→turn/end,
Cue agentStatus, Today wireState + last message role.
Turns that stay open with no progress fall back to an adaptive timeout
(p90 gap × 4, clamped to 600–3600 s) instead of a hard-coded value. Measured intra-turn silence
reaches p95 287 s / p99 1032 s, so the old 120 s window misreported long builds as finished.
Release builds are Universal (arm64 + x86_64) and declare macOS 12.0 as the minimum.
Fixed
Codex sessions whose last events were token_count after task_complete were reported as
unfinished; the state is now decided by event order, not timestamp comparison.
Cue sessions in waiting state were reported as completed and produced false "task finished"
notifications.
Claude Code sub-agent transcripts (agent-*.jsonl, isSidechain) and workflow-harness meta
sessions (cwd inside ~/.claude/projects, [Workflow harness …] prompts) flooded the list as
duplicate tasks. Both are filtered now.