callstack

callstack gives AI agents a proper call stack. Each /call forks the parent's full session context the way Unix fork() forks a process, so the child inherits everything without re-prompting; when the child completes, only its compact return value flows back to the parent. The design targets two well-known failure modes of long-horizon LLM workflows: sub-agents that start blind and waste tokens rebuilding context, and main agents whose context rots as everything accumulates in a single transcript.

The core insight was first prototyped inside Playbooks AI in 2023 and is now generalized to work with any harness, starting with Claude Code. The runtime is a pure state machine with effect-driven I/O, deterministic child session identity, parallel fan-out with a concurrency cap, and per-call trace files that downstream tools can render. It ships as the unwind-labs/callstack plugin with call / resume MCP tools and a Claude Code skill.