How A Run Works

1. Resolve bundle - resolveExecutionBundle loads the skill version, model, rendered composition, and tools (runPrompt.ts).
2. Create run - insert ai_runs with input_snapshot, status: running, and provider/model ids.
3. Execute - provider calls and optional tool steps; significant events log ai_run_steps (tool_call, final_response).
4. Validate - output is checked against the output schema; validation and diagnostics are collected.
5. Complete - update output_snapshot, token usage, status, and return the response payload.

HTTP POST /api/skills/:id/run wraps runPrompt and returns skillRunResponseSchema.

See also

• Failure Model
• What Gets Persisted