Skip to content
Halyard Docs

Dispatch an agent

A session is one agent run. You configure it at dispatch time, it executes on a node, streams its work live, and keeps a full transcript for review.

Start a session

Section titled “Start a session”

From a project, choose + New Session. The dialog collects the per-run configuration:

  • Name (and optional description) — distinguishes concurrent sessions of the same project, e.g. “review #42” or “feature/preview-pane”.
  • Target node — which machine runs the code (filtered to the project’s enclave).
  • Workspace path — the absolute path on the node for the working tree.
  • AI account — which provider link the agent authenticates as (filtered to the enclave).

Defaults are prefilled from the project’s most recent session, so re-running on the same node with the same account is one click.

You can also dispatch straight from a Halyard task: the session is bound to the task and its initial prompt is synthesized from the task definition (title, description, Definition of Ready / Done, and testing plan), so the work item travels with the session.

What the agent sees

Section titled “What the agent sees”

At dispatch, Halyard materializes the project library into the workspace — a synthesized CLAUDE.md preamble (goal doc first, then recent journal entries and a “what changed since last time” summary, then the library index) plus the design and guideline docs under .halyard/library/. The agent reads the why before the how.

Monitor and unblock

Section titled “Monitor and unblock”

Watch progress live in the Fleet View — real-time state and streaming terminal output. If the agent hits a blocker it moves to awaiting-input and surfaces in your Inbox and on Today; answer in the human-in-the-loop console and it resumes from where it paused. See Agent sessions & HITL.

Review with traceability

Section titled “Review with traceability”

When a session bound to a task completes, the task is automatically moved to in-review. Review the work and the generated PR-review draft, then post it yourself — you stay the human in the loop for every approval. Every session traces back through task → issue → requirement → PRD, so nothing ships without a line back to why it exists. See Traceability & compliance.