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
Add user-defined custom tools loaded from the tools directory in the config folder, exposed in the Settings > Agents Custom Tools panel and usable in custom-agent tools whitelists and the central tool dispatcher
Add report_llm_error/3 to AgentScheduler to allow specifying an explicit backoff duration when reporting LLM errors
Changed
Compress built-in agent system prompts (Manager, Executor, ContextExtractor, Investigator, Architect, TaskScheduler, GenesisPlanner, SkillExtractor) by 14-37% with no behavior change
Reduce rendered built-in prompt text by roughly 5% by compressing shared prompt fragments, preserving all instructions and the public API
Agents now trust subagent reports by default instead of re-investigating completed work, and test runs are scoped to each node's subtree rather than the full suite at every depth, reducing token and test-execution cost
Handle model-exhaustion errors (e.g. insufficient balance / HTTP 402) with a cancel-safe per-model scheduler backoff instead of short agent-side retries; transient failures keep the existing exponential backoff
Update dependencies: jsv 0.23.0, llm_db 2026.9.4, mint 1.10.1, and xqlite 0.12.2 with a matching NIF pin
Fixed
Fix sandbox port lifecycle: wait_for_os_pid/2 now returns the real integer PID, and close_port/1 is nil-safe and idempotent on timeout
Fix executable resolution to only return bundled tool paths when the bundled file actually exists, restoring PATH and Windows fallback lookup for non-bundled tools
Fix Windows crash (badarg) in the None sandbox shell tools by converting Port.open environment variables to charlists