Staffdeck API Run Agent

Run a StaffDeck agent

Prepare

1. Require STAFFDECK_BASE_URL, STAFFDECK_API_KEY, and an accessible agent_id.
2. Read references/run-api.md for routes and payloads.
3. Generate a unique, stable Idempotency-Key for each logical session or run. Reuse it only to retry the identical request.

Choose session mode

• Use a persistent session when the task may require user clarification, a pending SOP, or follow-up turns.
• Use stateless mode only for an isolated request that must not affect later conversations.
• Reuse the same session_id for every continuation of an awaiting_input run.

Execute

1. For a persistent conversation, create or recover a session with POST /agents/{agent_id}/sessions. Supply a stable external_session_id from the caller's system.
2. For an interactive answer, prefer POST /agents/{agent_id}/runs:stream and consume SSE in the same request. Retain the X-Run-ID response header.
3. For detached execution, use POST /agents/{agent_id}/runs, then consume GET /runs/{run_id}/events or poll GET /runs/{run_id} with bounded backoff.
4. Stop polling at succeeded, failed, cancelled, or awaiting_input.
5. Call /result only after succeeded.
6. Preserve reply, citations, tool_calls, task_results, session_state, and artifacts in the result passed to the caller.

Continue or cancel

• On awaiting_input, report the requested information and submit the user's answer as a new run in the same session.
• On ACTION_BUDGET_EXHAUSTED, preserve the session and continue only when authorized; do not silently loop forever.
• Cancel with POST /runs/{run_id}:cancel when the caller requests cancellation or the result is no longer needed.
• Use Last-Event-ID when reconnecting SSE.
• Append run.output.delta content in order. Replace the accumulated text when run.output.replace is emitted, and stop display streaming at run.output.completed.

Evidence and privacy

Treat public Trace as an auditable execution summary, not raw COT. Keep citation metadata attached to the claims it supports. Do not expose masked credentials, sensitive model inputs, or unrelated session data.