You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(core): Dispatch reports success for a tool call that changed nothing #502
The quiet reason a stubbed MCP server looked like it worked, found while proving #470 and left alone by #491 because a named test pins the current behaviour.
What happens
dispatch answers a speak or a play_behavior with success even when its context has an empty roster, so nothing on screen could possibly have changed. A caller cannot tell "I said it" from "there was nobody to say it to".
That is how every MCP verification before #491 stopped at "the tool call returned success", and it is how a stdio-only Harness still behaves today, because that path builds its context from stubs.
Why it is worth fixing rather than papering over
The shipped loopback path now always has a real roster, so this no longer bites there. It still bites the stdio path, and more importantly it is the wrong contract: a tool that reports success without an effect trains everyone, human and Harness alike, to trust a meaningless answer. The shim in the sibling issue depends on a failure being visible.
A named test currently pins the empty-roster success, so changing this means deciding what the honest answer is: an error, or a success that says plainly it addressed no Instance. Either is better than today, and the test should be rewritten to pin whichever is chosen rather than deleted.
Acceptance
A speak or play_behavior that reaches no Instance does not report plain success.
Whatever it does report is legible to a Harness, and the existing test pins that rather than the old behaviour.
The shipped loopback path is unaffected, since it always has a real roster.
Related: #470, #491, #166, and the stdio shim issue beside this one.
Activity