Agent sessions & HITL

A session is one agent run. It is configured at dispatch time, scheduled onto a node in the project’s enclave, streams its work live, and keeps a full transcript for review. Because per-run configuration lives on the session (not the project), one project can run many concurrent sessions across different nodes and AI accounts.

Lifecycle

A session moves through:

• queued — waiting for a node with capacity (it stays here safely while a node is offline);
• running — executing on the node, streaming output;
• awaiting-input — paused on a blocker, waiting for you (see below);
• a terminal state — completed, failed, or abandoned.

A session can be bound to a Halyard task. When it is, status changes roll up automatically and forward-only: a task moves to started when its session begins running, and to in-review when the session completes. The rollup never demotes, so you can always re-open a task without it fighting you.

Monitoring (Fleet View)

The Fleet View shows every active agent with its real-time state and a live tail of terminal output, replayed from the transcript and then streamed live. Each session has its own live stream.

Human-in-the-loop (HITL)

When an agent needs a decision — a missing secret, an architectural choice, a clarification — the session moves to awaiting-input and surfaces in your Inbox and on the Today dashboard. Open it, answer in the HITL console, and the agent resumes from exactly where it paused. The whole exchange is kept in the session transcript, so the reasoning behind a decision stays traceable.

Context the agent receives

At dispatch, the project library is materialized into the workspace: a synthesized CLAUDE.md preamble (goal doc, recent journal entries, library index) plus design and guideline docs. Sessions can narrow or extend this with per-session context, and the agent’s changes to library files flow back into the project at session end (additive only — a runaway agent can’t delete your library).

Reviewing the result

Completed sessions produce the work plus, for the AI review flow, a draft review you read, edit, and post yourself. You remain the human in the loop for every approval — see Dispatch an agent and Traceability & compliance.