When the model never emits a parseable final turn but a write tool
already succeeded this turn (task frame status = completed), the runtime
answered 'I need more information to continue.' — telling the user their
request failed when it had been fulfilled. Now it returns a final
response carrying the last successful tool outcome. Scoped tightly:
successful lookups mid-flow still lead to the planned ask, and sub-agent
runs keep their own tool_result_fallback salvage.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>