Workflow doctor

Workflow Doctor

Diagnose and repair workflows. The value is in correlation — don't just say "this workflow has a low success rate," say which node fails and why, then offer the specific fix.

Steps

1. Scope. check_connection, then list_sub_accounts if multi-workspace and ask which one. list_workflows to get the inventory (name, trigger, active, execution count).

2. Run the audit. workflow_health_audit for the account-level view, then for each concerning workflow:

   • get_workflow_performance — success rate, avg duration, throughput.
   • get_workflow_failure_stats — what's erroring and how often.
   • get_workflow_node_analytics / get_workflow_step_analytics — the key correlation: which specific node drops or fails contacts.
   • get_workflow_details — the actual config of the failing node.

3. Classify each problem into one of:

   • Broken — errors at a node (bad config, missing integration, invalid template variable). Correlate the error to the node.
   • Stalled — active but hasn't fired recently; usually the trigger no longer matches (e.g. a tag/form that was renamed) or the source integration disconnected (check_connection on that account).
   • Leaky — fires fine but contacts drop at a wait/condition node that's too strict, or a send step with a low open/click.
   • Dormant — inactive draft that should probably be live.

4. Propose a fix per problem — concrete and specific: "Node 3 'Send Email' fails with invalid_template_variable {first_name} — should be {lead.first_name}." Use list_template_variables, describe_trigger, describe_condition to get the correct shapes.

5. Apply on approval. Only after the user says yes: update via the appropriate write tool (or create_workflow for a rebuild), then always test_workflow (simulated — nothing sent to real people) and report the per-node trace. Never activate a workflow the user didn't ask to activate.

Output

• Health summary — X workflows, Y healthy, Z need attention.
• Findings — a ranked list, worst first. Each: workflow name → the exact node → the cause → the one-line fix. Rank by impact (contacts affected × how central the workflow is).
• Proposed fixes — numbered, each with the precise change. Ask which to apply.

Guardrails

• Read + diagnose freely; never write without an explicit yes naming the workflow(s) to change.
• Confirm the workspace before any change (the MCP enforces this via a confirmation envelope — relay it).
• After any change, test in simulation before suggesting activation.
• If a fix needs something outside Inflowave (reconnect Instagram, verify a domain), say so plainly instead of pretending the workflow alone can fix it.