Treat an unfinished history tail as a continuation too

deferred_tool_results may arrive with a non-empty prompt, so the absence of a
prompt cannot be the only test for a resumption. Reproduced: resuming a live
question that way replaced its search result with the earlier-question notice
while the deferred result arrived alongside it.

_is_resumption accepts either signal — no prompt, or a history ending with a
request the model has not answered or a response whose tool calls have no
returns. A settled history ends with the previous answer, so a genuinely new
question is unaffected.

A new prompt on top of an unanswered tail is ambiguous and now counts as a
continuation: compacting costs the answer if it is one, while not compacting
only costs a larger request.

haiku_rag_slim/haiku/rag/capabilities/_base.py

@@ -123,6 +123,31 @@ def _compact_old_tool_returns(
     return compacted
 
 
+def _is_resumption(prompt: Any, messages: list[ModelMessage]) -> bool:
+    """Whether this run continues a question rather than asking a new one.
+
+    Two signals, either of which is enough, because getting this wrong hands the
+    model a notice where its own evidence should be:
+
+    - no prompt: how pydantic-ai resumes for interruptions and suspensions.
+    - an unfinished tail: the history ends with a request the model has not
+      answered, or with a response whose tool calls have no returns yet. Deferred
+      tool results may arrive *with* a prompt, so the prompt alone is not enough.
+
+    A settled history ends with the previous answer, so a genuinely new question
+    is not mistaken for a continuation. The framework's own first-new-message
+    index would be better than either signal, but it is not public here.
+    """
+    if prompt is None:
+        return True
+    if not messages:
+        return False
+    last = messages[-1]
+    if isinstance(last, ModelRequest):
+        return True
+    return any(isinstance(part, ToolCallPart) for part in last.parts)
+
+
 def _called_own_tool(messages: list[ModelMessage], tool_names: frozenset[str]) -> bool:
     """Whether the model's most recent response called one of these tools."""
     for message in reversed(messages):
@@ -157,14 +182,11 @@ class RAGCapabilityBase[StateT: BaseModel](AbstractCapability[Any]):
     async def for_run(self, ctx: RunContext[Any]) -> "RAGCapabilityBase[StateT]":
         """Start a run's own copy, and decide what counts as an earlier question.
 
-        A run carrying no prompt is continuing a question rather than asking one:
-        pydantic-ai resumes that way for deferred tool results, interruptions and
-        suspended responses. ``len(ctx.messages)`` would then count the live
-        question's own messages and hand the model a notice where its search
-        result should be, so compaction is switched off for the whole run
-        (``turn_start=0``). The absence of a prompt is the signal — the three
-        resume shapes differ in message layout, and reading the layout is what
-        broke this in the first place.
+        On a resumption (see ``_is_resumption``) compaction is switched off for the
+        whole run: ``len(ctx.messages)`` would count the live question's own
+        messages and replace its evidence with the earlier-question notice, leaving
+        the model to answer with the evidence taken away. Failing this way costs a
+        larger request; failing the other way costs the answer.
         """
         outer = getattr(ctx.deps, "state", None)
         outer_state = outer if isinstance(outer, dict) else None
@@ -181,7 +203,9 @@ class RAGCapabilityBase[StateT: BaseModel](AbstractCapability[Any]):
             search_count=0,
             request_count=0,
             grace_requests_used=0,
-            turn_start=0 if ctx.prompt is None else len(ctx.messages),
+            turn_start=(
+                0 if _is_resumption(ctx.prompt, ctx.messages) else len(ctx.messages)
+            ),
         )
         run_capability._sync_state()
         return run_capability

tests/capabilities/test_capabilities.py

@@ -988,7 +988,39 @@ def _wire_returns(sent: list[Any]) -> list[str]:
 
 @pytest.mark.asyncio
 async def test_a_new_question_compacts_the_previous_one(temp_db_path):
-    """The baseline the resume cases are contrasted against."""
+    """The baseline the resume cases are contrasted against.
+
+    A settled history — the previous question answered — is what a genuinely new
+    question follows. An unfinished tail is ambiguous instead, and treated as a
+    continuation.
+    """
+    capability = create_rag(
+        db_path=temp_db_path, config=AppConfig(), defer_loading=False
+    )
+    wire: list[list[Any]] = []
+
+    async def model(messages, _info):
+        wire.append(list(messages))
+        return ModelResponse(parts=[TextPart("answer")])
+
+    agent = Agent(FunctionModel(model), deps_type=Deps, capabilities=[capability])
+    settled = [*_in_flight_history(), ModelResponse(parts=[TextPart("first answer")])]
+
+    await agent.run("a different question", deps=Deps(), message_history=settled)
+
+    assert _wire_returns(wire[-1]) == [PRIOR_TURN_NOTICE]
+
+
+@pytest.mark.asyncio
+async def test_a_prompt_on_an_unanswered_tail_is_treated_as_a_continuation(
+    temp_db_path,
+):
+    """Ambiguous shape, resolved the safe way.
+
+    A history ending in a request the model never answered, plus a new prompt,
+    could be a fresh question or a continuation. Compacting would cost the answer
+    if it is a continuation; not compacting only costs a larger request.
+    """
     capability = create_rag(
         db_path=temp_db_path, config=AppConfig(), defer_loading=False
     )
@@ -1004,7 +1036,41 @@ async def test_a_new_question_compacts_the_previous_one(temp_db_path):
         "a different question", deps=Deps(), message_history=_in_flight_history()
     )
 
-    assert _wire_returns(wire[-1]) == [PRIOR_TURN_NOTICE]
+    assert _wire_returns(wire[-1]) == ["EVIDENCE FOR THE LIVE TURN"]
+
+
+@pytest.mark.asyncio
+async def test_a_resume_carrying_a_prompt_keeps_the_active_evidence(temp_db_path):
+    """Deferred results may arrive with a prompt, and that is still a continuation.
+
+    ``ctx.prompt`` is non-null here, so the absence of a prompt cannot be the only
+    signal: the history tail is unfinished — a response whose tool call has no
+    return yet — and the question it belongs to is still being answered.
+    """
+    capability = create_rag(
+        db_path=temp_db_path, config=AppConfig(), defer_loading=False
+    )
+    wire: list[list[Any]] = []
+
+    async def model(messages, _info):
+        wire.append(list(messages))
+        return ModelResponse(parts=[TextPart("answer")])
+
+    agent = Agent(FunctionModel(model), deps_type=Deps, capabilities=[capability])
+    history = [
+        *_in_flight_history(),
+        ModelResponse(parts=[ToolCallPart("external_tool", {}, "call-2")]),
+    ]
+
+    await agent.run(
+        "carry on",
+        deferred_tool_results=DeferredToolResults(calls={"call-2": "external result"}),
+        message_history=history,
+        deps=Deps(),
+    )
+
+    assert "EVIDENCE FOR THE LIVE TURN" in _wire_returns(wire[-1])
+    assert PRIOR_TURN_NOTICE not in _wire_returns(wire[-1])
 
 
 @pytest.mark.asyncio