Aevatar Workflow Engine Probe
Use this to confirm the workflow engine can start a run, execute steps, and stream the result back
into chat — without depending on any pre-registered workflow. It uses aevatar_start_workflow,
which accepts an inline workflow_yamls bundle and runs it self-contained.
You (the agent) start the probe workflow yourself and check the returned output. The workflow just echoes a token, so the answer is deterministic. A start/run error is a finding — quote it verbatim, never fabricate a result.
How to run it
Call aevatar_start_workflow with:
workflow_id:probe_echoinputs.prompt:workflow-probe-7421(the token to echo back)wait:complete(block until the run finishes so you can read the output)workflow_yamls: a one-element bundle containing exactly this inline workflow:
name: probe_echo
description: Inline engine probe — echoes the run input as the run output.
when_to_use: Engine liveness probe.
steps:
- id: echo
type: assign
parameters:
target: result
value: "${input}"
PASS if the run finishes and its output contains workflow-probe-7421. FAIL if the start is
rejected, the run errors, or the output is empty / different.
Output
One report: did the run start? did it reach a terminal finished state? did the output echo the
token? Then a verdict: WORKFLOW ENGINE OK or WORKFLOW ENGINE BLOCKED: <verbatim error / where it stalled>.
Interpretation (report, don't fix):
workflow_yamls are not accepted inside a managed workflow runtime context→ you are already inside a workflow run; run this probe from a plain chat turn instead.- start rejected for scope / authorization → the caller's NyxID scope or token is unavailable.
- run starts but never finishes (no terminal output) → the engine accepted but the run stalled; note that the engine ingress works but execution/return did not complete.
Guardrails
- Only start the bundled inline echo workflow; never start arbitrary or side-effecting workflows from this probe.
- Use
wait: completeso you actually observe the terminal output; never report a pass from anacceptedack alone. - Report verbatim; never claim the engine works without an observed echoed token.
- Never ask the user for tokens — your NyxID-brokered tools handle all credentials.