Deliver an async-tool result on its own once the user is answered

A result that lands after the model has already answered is delivered
alone; the guidance used to have it re-answer the user first, which had
the cascade frontend re-tell a joke before relaying the weather.

Author: kompfner

changelog/+async-tool-own-run.fixed.md

@@ -0,0 +1 @@
+- Fixed the async-tool guidance given to LLMs (`ASYNC_TOOL_INSTRUCTIONS` and the final-result message) so a result that lands after the model has already answered the user is delivered on its own, without the model repeating its previous reply first. The guidance now distinguishes a result arriving while the user has an unanswered request from one arriving on a run of its own.

src/pipecat/processors/aggregators/async_tool_messages.py

@@ -96,30 +96,32 @@
 
 # Standing guidance composed into the system instruction whenever an async tool is
 # registered. The per-result message says the same thing, but it arrives buried in a
-# context whose most recent turn is the user asking for something else; a model
+# context whose most recent turn may be the user asking for something else; a model
 # weighing the two follows the nearer, louder request. This states the policy before
 # any result exists, so it is in force when one arrives.
 ASYNC_TOOL_INSTRUCTIONS = """ASYNC TOOLS:
 Some of your tools keep running after you have replied. Their results arrive later as \
 messages in the conversation, on whatever turn happens to be in progress by then.
 
 A result that has arrived is owed to the user, whatever the conversation has moved on to. \
-Answer what the user just said first, then add the result at the end of that same reply — \
-never before your answer, and never as a reply of its own. State a short result outright; \
-for a long one, say what came back and offer the details. Say it once, and do not repeat \
-it in later replies."""
+If the user has said something you have not answered yet, answer that first, then add the \
+result at the end of that same reply — never before your answer. If you have already \
+answered everything the user said, deliver just the result — do not repeat or rephrase \
+your earlier reply. State a short result outright; for a long one, say what came back and \
+offer the details. Say it once, and do not repeat it in later replies."""
 
 # Description shipped on the final-result message.
 _FINAL_DESCRIPTION = (
     "This is the final result for the asynchronous task associated with this "
     "tool_call_id. The task has completed. No further results will arrive for "
     "this tool_call_id. You must convey this result to the user, even if the "
-    "conversation has moved on. Never leave it unsaid. First finish responding "
-    "to whatever the user is talking about now, then deliver the result at the "
-    "end of your response. How you deliver it depends on its size: if the "
-    "result is short, simply state it; if it is long or complex, name what has "
-    "come back and offer the details. Convey it once; do not repeat it in "
-    "later responses."
+    "conversation has moved on. Never leave it unsaid. If the user has said "
+    "something you have not answered yet, first finish responding to that, "
+    "then deliver the result at the end of your response; if you have already "
+    "answered, deliver just the result and do not repeat your earlier reply. "
+    "How you deliver it depends on its size: if the result is short, simply "
+    "state it; if it is long or complex, name what has come back and offer "
+    "the details. Convey it once; do not repeat it in later responses."
 )