Validate tool call arguments against the tool's JSON schema before executing. Return error tool results to the model on validation failure instead of crashing or executing with bad args.
HTTP/WebSocket API exposing the two-layer event stream. Clients can start sessions, send messages, receive typed AgentEvents, and steer/interrupt/stop via API calls.
Named hook points at every decision point in the agent loop. transformContext is the most important to add early. See plan.
Auto-load AGENTS.md from repo root and global config path into context. Support SYSTEM_PROMPT.md as the agent's system prompt with repo-level rules appended.
Skill discovery, auto-loading of headers into context, full loading via slash command or agent-initiated. Skills are markdown instruction files — just context management, no special runtime.
Implement the ability to roll back the entire codebase/file environment to the state it was in at a given message. Each message (or tool execution) should snapshot the file-system changes it introduced, allowing the user to revert all changes back to any prior message boundary. This enables undo-style workflows where the user can say "roll back to message N" and have every file restored to its state at that point.