Repository navigation
v0.1.30
Summary
Introduces vectl checkpoint — a machine-readable, token-efficient JSON snapshot of the plan state, designed for LLM context compaction and agent-to-agent handoff.
New Features
- CLI:
vectl checkpoint(outputs JSON). - MCP:
vectl_checkpointtool (returns structured dict). - Features:
- Deterministic Focus: Selects a single "focus step" to guide the next agent (prevents drift).
- Bounded Output: Limits
next_steps,active_steps, andguidancerefs to save tokens. - Agent Aware:
--agent <name>prioritizes steps claimed by that agent. - Parity: CLI and MCP return the exact same schema (v1).
Use Cases
- Context Compaction: Inject the checkpoint JSON as "Ground Truth" when summarizing long conversation history.
- Handoff: Pass the
focus.step_idto a sub-agent to strictly scope its work.