Run Execution

Run creation and execution are separate operations. prepare_run loads the strategy and initializes its daemon systems. After the daemon replies with run_prepared, send start_run with the returned run ID and optional initial input. Send stop_run to cancel either a prepared or running run.

Prepare A Run

{
  "type": "prepare_run",
  "runId": "run_abc123",
  "strategyPath": "./strategies/code-review.yaml",
  "modelOverride": "anthropic/claude-sonnet-4-20250514",
  "cwd": "/workspace",
  "requestId": "prepare-1"
}

For a new run, strategyPath is required and runId is optional. To prepare an existing run for continuation, send its persisted runId; strategyPath, modelOverride, cwd, and manifestPath may override the previous execution's configuration. The requesting client is automatically subscribed to the run.

Successful preparation replies directly:

{
  "type": "run_prepared",
  "runId": "run_abc123",
  "strategyName": "Code Review Pipeline",
  "agents": ["analyzer", "reviewer"],
  "flowTree": { "name": "Review Pipeline", "type": "sequential" },
  "requestId": "prepare-1",
  "ts": "2026-06-14T10:30:00.000Z"
}

The run remains pending and is not persisted until it is started. Preparation failures return a correlated error with code PREPARE_FAILED.

Start A Run

{
  "type": "start_run",
  "runId": "run_abc123",
  "input": "Review this function.",
  "requestId": "start-1"
}

runId is required and input is optional. A successful start emits strategy_started, followed by execution events. Lifecycle events echo the start_run request ID.

{
  "type": "strategy_started",
  "runId": "run_abc123",
  "strategyName": "Code Review Pipeline",
  "agents": ["analyzer", "reviewer"],
  "flowTree": { "name": "Review Pipeline", "type": "sequential" },
  "requestId": "start-1",
  "ts": "2026-06-14T10:30:01.000Z"
}

Successful execution ends with strategy_completed. Execution failures end with strategy_error.

Continue A Run

Prepare the completed run by ID before continuing it:

{
  "type": "prepare_run",
  "runId": "run_abc123",
  "requestId": "prepare-continuation-1"
}

Preparation reloads the strategy and restores each matching agent's persisted conversation context. It does not append timeline events until execution starts. Then send the new input with continue_run:

{
  "type": "continue_run",
  "runId": "run_abc123",
  "input": "Refine the result using the review feedback.",
  "requestId": "continue-1"
}

The continuation reuses the same run ID and appends another lifecycle segment to its persisted timeline. continue_run only accepts a run prepared from an existing timeline; synchronous failures return a correlated error with code CONTINUE_FAILED.

Stop A Run

{
  "type": "stop_run",
  "runId": "run_abc123"
}

Stopping a pending run releases its prepared resources. Stopping a running run aborts execution. Subscribers receive strategy_error with code CANCELLED.

Event Sequence

prepare_run -> run_prepared
start_run -> strategy_started -> step/agent events -> strategy_completed
prepare_run(existing) -> run_prepared
continue_run -> strategy_started -> step/agent events -> strategy_completed
stop_run -> strategy_error (CANCELLED)