Editing Openai/694057b6-101c-8007-9a65-c40578c7252d (section)

=== Assistant: Locked in. ===
Locked in. I’ve added this as a tooling rule (applies to both normal tools and any future heavier tools) and then here’s the next section.

==== ### ====
* The worker must treat a tool call as failed (request-killing) only when there is an actual execution problem, such as: - tool runner timeout - tool runner raises an exception - tool runner returns a non-serializable result (implementation bug) - malformed tool arguments (parse/validation failure)
* “No results found” (e.g., search returns empty) is not a failure. It must be represented as a normal successful tool result (e.g., [], { "results": [] }, or { "found": false }) and the request continues.

===== - The worker does not interpret tool output semantics (empty vs non-empty). It only enforces transport/format/execution correctness. =====
* The ToolRunner is responsible for returning a meaningful “no results” value rather than raising.

===== - Appendix G (Tool loop): failure conditions for tools are strictly execution/format problems, not domain outcomes. =====
* Appendix A (ToolRunner Protocol): “no results” must be represented in the returned JSON-serializable value, not as an exception.