Feed tool errors back to the agent for recovery

Goal

Turn tool failures into structured feedback so the agent can correct course instead of claiming success.

Steps

1. Register a PostToolUse matcher, read input_data["tool_response"], and detect an error state. [hooks]
2. Return systemMessage, reason, and hookSpecificOutput.additionalContext with concise, non-sensitive recovery hints. [hooks]
3. Test with one reproducibly failing tool and one successful tool; still inspect whether the final ResultMessage succeeded. [hooks]

How to verify

An error response adds context while a successful response does not; the application never treats a tool error as task completion.

Troubleshooting

If the model repeats failures, cap retries and define a stop condition; ensure detection does not flag normal output.