Skip to content
Orcho Docs

Observe a run

This is the MCP way to observe a run live: instead of reading the terminal, a client polls typed state. Observation is separate from execution — an MCP client can lose its connection and reconnect by reading durable run state.

For the first setup path, read Start with MCP. For the economic reason to spend more tokens on client-side lifecycle control, read LLM captain mode. For the full client-captain workflow, read Expert MCP control loop.

Section titled “Recommended loop”
  1. Start the run with orcho_run_start.
  2. Read orcho_run_status for the run id.
  3. Tail events with orcho_run_events_tail.
  4. When the run stops, call orcho_run_diagnose.
  5. If evidence is needed, call orcho_run_evidence.
  6. If a decision is needed, use the typed ready call returned by the tool.

Do not scrape raw logs first

Section titled “Do not scrape raw logs first”

Raw logs are useful for forensic debugging. They should not be the primary control loop.

Use typed slices first:

  • status for state;
  • events for progress;
  • evidence for proof;
  • diagnosis for next action;
  • delivery gate for release/correction decisions.