Fix expression error in child flow

Real production case: Contoso Timesheet child flow failing intermittently.

What the portal showed

• Parent flow: succeeded
• Child flow: Failed
• Error: ExpressionEvaluationFailed — no further detail visible

The portal does not expand into child flow actions or show which expression failed or what input caused it.

What the agent did

1. Listed recent runs of the parent flow

Agent: list_live_flows → found "Contoso Timesheet" parent flow
Agent: get_live_flow_runs → found 2 recent runs: one succeeded, one failed

2. Compared the two runs

The agent retrieved action outputs for both the successful and failed runs using get_live_flow_run_action_outputs. This is the key step — the portal only shows one run at a time and hides action-level detail behind click-through.

3. Traced into the child flow

The failed run showed the parent flow called a child flow ("Contoso Timesheet Child"). The agent followed the child flow's run ID and retrieved its action outputs.

4. Found the failing action inside a loop

Inside the child flow, a Select action inside an Apply to each loop was using this expression:

contains(string(item()?['serviceExceptionJson']), 'error')

The expression crashed when serviceExceptionJson was a nested object instead of a string. On successful runs, the field was either null or a flat string. On the failing run, the Power BI refresh returned a nested JSON structure that string() could not serialize the same way.

5. Root cause

The Power BI data source had an intermittent SQL timeout. When the timeout occurred, the serviceExceptionJson field contained a nested retry structure instead of a flat error string. The expression assumed a flat string.

What the agent could not see without MCP

Via Graph API alone, the agent only saw ExpressionEvaluationFailed as a status code. A human could click through the portal and find all of this, but the agent had no way to:

• See which expression failed or what input caused it
• Retrieve the nested JSON structure from the action inputs
• Compare the successful and failed run inputs side by side
• Trace the root cause upstream to the PBI data source timeout

Tools used

• list_live_flows — find the flow
• get_live_flow_runs — get recent run history
• get_live_flow_run_action_outputs — retrieve action-level inputs and outputs
• get_live_flow_run_error — top-level error (insufficient on its own)