A run needs reliable boundaries.
Session, prompt, task, and worktree events make it possible to reconstruct where work started, what intent entered, which unit completed, and what workspace was touched.
Deep implementation
At fleet scale, lifecycle events become the control plane.
Autonomous engineering becomes governable when the system can observe, enforce, preserve, and reconstruct every meaningful boundary around the agent.
Read the original on XThe evidence card
The final note is the implementation map: 29 Claude Code lifecycle events organized into session, prompt, tool, turn, compaction, subagent, permission, environment, task, and worktree boundaries. Citadel turns those events into an operating layer rather than leaving them as callbacks.
The model
The visual is an editorial compression of the source essay: enough structure to orient a reader, without pretending that the model replaces the underlying evidence.
The argument
The site version keeps the source's operational insight while making its sequence easier to inspect, share, and connect to the rest of the archive.
Session, prompt, task, and worktree events make it possible to reconstruct where work started, what intent entered, which unit completed, and what workspace was touched.
Pre-tool gates, post-tool checks, failure signals, and stop-time verification turn repository rules into a living operating surface.
Compaction, subagent, permission, environment, and MCP events surface the state that otherwise disappears when context compresses or an agent silently stalls.
The boundary
Most agent systems focus on prompting. Citadel focuses on lifecycle.
A control plane improves observability and operating discipline; it does not make the underlying agent wise or remove the need for review.