v1.3.190

v1.3.190

Features

• Shell-to-native tool suggestions: When the agent uses shell commands (cat, grep, find, git log, etc.) that native tools handle better, a one-time hint suggests the native equivalent for richer structured output.
• Post-completion commit hint gate: After all completion gates pass, if the agent made code changes but hasn't staged/committed them, a system message reminds it to do so -- matching Aider/Claude Code auto-commit behavior.
• Session-level approval memory: After 3 consecutive approvals of the same (tool, path-pattern) combination in supervised mode, future matching calls are auto-approved -- eliminating approval fatigue.
• Sliding-window compounding failure detection: Detects repeated failure patterns in a sliding window and triggers strategy reset to break out of failure loops.

Fixes

• Session timeout increased to 600 minutes: The 30-minute autopilot timeout was too short for complex autonomous tasks. Now 600 minutes (10 hours) as an ultimate backstop.
• Session timeout warnings no longer pollute LLM context: 80%/95% timeout warnings are now emitted as system events visible only to the user, not injected into LLM context.
• Visible system message on timeout stop: When the agent is stopped due to timeout, a visible system event is now emitted so the user knows why.
• Sub-agent and teammate todo file isolation: Sub-agents and swarm teammates no longer share the parent agent's todo file. Each gets its own session-scoped todo file, auto-cleaned on exit.