Agent nodes: Anthropic models with structured output fail ('Invalid returnData shape in AI action task callback') — and each failed attempt bills 0.1 data credits

[TenSpy-ai]

Summary

Workflow agent nodes configured with an Anthropic model (agentModel: claude-haiku-4-5) and a structured outputSchema fail at result delivery on every run with:

Invalid returnData shape in AI action task callback

The identical node configuration with an OpenAI model (gpt-5.4-nano) runs clean. And each failed Anthropic attempt still bills 0.1 data credits — paying for guaranteed-zero output, which compounds while debugging (every retry bills).

Reproduction (verified 2026-08-06, reproduced twice)

1. Build a workflow with an agent node that has a structured outputSchema (simple object schema, e.g. {pass, persona_bucket, reasoning}).
2. Set agentModel: claude-haiku-4-5.
3. Run the workflow → the node fails with the callback-shape error above.
4. Check the credit ledger: 0.1 data credits consumed per failed attempt.
5. Change only the model to gpt-5.4-nano → same node succeeds (1.0 data + 1 action credit).

Expected

• Anthropic models deliver structured output like OpenAI models do — or the config surface rejects the unsupported model+outputSchema combination up front instead of failing at run time.
• Runs that fail inside the AI action callback (producing no output) should not bill data credits.

Built/observed via the Clay plugin's workflow MCP (edit_node) + CLI on Claude Code. Happy to provide more detail.