fix(soul): repair orphan tool_calls when replaying history

CHANGELOG.md

@@ -11,6 +11,8 @@ Only write entries that are worth mentioning to users.
 
 ## Unreleased
 
+- Soul: Sessions killed mid-turn (e.g. under memory pressure) now resume correctly — `normalize_history` injects placeholder responses for orphan `tool_call_id`s so the next provider request stops failing with `400 ... tool_call_ids did not have response messages`
+
 ## 1.47.0 (2026-06-05)
 
 - Shell: Guide users to the new standalone Kimi Code — adds a `/upgrade` command that installs it (migrating your config & sessions automatically), a welcome-screen nudge, and a once-per-day tip shown on exit

src/kimi_cli/soul/dynamic_injection.py

@@ -5,7 +5,7 @@
 from dataclasses import dataclass
 from typing import TYPE_CHECKING
 
-from kosong.message import Message
+from kosong.message import Message, TextPart
 
 from kimi_cli.notifications import is_notification_message
 
@@ -55,6 +55,61 @@ async def on_afk_changed(self, enabled: bool) -> None:
         return None
 
 
+# Placeholder body for synthetic tool responses inserted to repair an
+# assistant ``tool_calls`` message whose matching ``tool`` response was lost
+# (e.g. when the previous session was killed mid-turn under memory pressure).
+# Kept short so it adds minimal context tokens on resume.
+_ORPHAN_TOOL_PLACEHOLDER = (
+    "(tool result unavailable: the previous session was interrupted before "
+    "this tool call completed)"
+)
+
+
+def _repair_orphan_tool_calls(history: list[Message]) -> list[Message]:
+    """Insert placeholder tool responses for any orphan ``tool_call_id``.
+
+    A persisted assistant message can carry ``tool_calls`` without the
+    matching ``tool`` role responses ever being flushed -- e.g. the CLI
+    process was killed mid-turn or the worker crashed before the tool
+    result was written. OpenAI-compatible providers then reject the next
+    request with ``400 ... tool_call_ids did not have response messages``,
+    permanently breaking conversation resume (regression #2336).
+
+    Scan the history for assistant messages with ``tool_calls``, gather
+    their declared ids, and check the immediately following ``tool`` role
+    messages. For any id without a response, insert a synthetic placeholder
+    so the API call shape stays valid. The stored history itself is left
+    untouched -- only the sequence sent to the provider is patched.
+    """
+    if not history:
+        return history
+    result: list[Message] = []
+    for i, msg in enumerate(history):
+        result.append(msg)
+        if msg.role != "assistant" or not msg.tool_calls:
+            continue
+        expected_ids = [tc.id for tc in msg.tool_calls if tc.id]
+        if not expected_ids:
+            continue
+        seen: set[str] = set()
+        for follower in history[i + 1 :]:
+            if follower.role != "tool":
+                break
+            if follower.tool_call_id:
+                seen.add(follower.tool_call_id)
+        for expected_id in expected_ids:
+            if expected_id in seen:
+                continue
+            result.append(
+                Message(
+                    role="tool",
+                    content=[TextPart(text=_ORPHAN_TOOL_PLACEHOLDER)],
+                    tool_call_id=expected_id,
+                )
+            )
+    return result
+
+
 def normalize_history(history: Sequence[Message]) -> list[Message]:
     """Merge adjacent user messages to produce a clean API input sequence.
 
@@ -64,6 +119,10 @@ def normalize_history(history: Sequence[Message]) -> list[Message]:
     Only ``user`` role messages are merged. Assistant and tool messages
     are never merged because their ``tool_calls`` / ``tool_call_id``
     fields form linked pairs that must stay intact.
+
+    Orphan ``tool_calls`` whose tool responses were lost (e.g. mid-turn
+    crash) are patched with placeholder ``tool`` messages so the provider
+    request stays well-formed; see ``_repair_orphan_tool_calls``.
     """
     if not history:
         return []
@@ -81,4 +140,4 @@ def normalize_history(history: Sequence[Message]) -> list[Message]:
             result[-1] = Message(role="user", content=merged_content)
         else:
             result.append(msg)
-    return result
+    return _repair_orphan_tool_calls(result)

tests/core/test_normalize_history.py

@@ -2,7 +2,7 @@
 
 from __future__ import annotations
 
-from kosong.message import ContentPart, Message, TextPart
+from kosong.message import ContentPart, Message, TextPart, ToolCall
 
 from kimi_cli.soul.dynamic_injection import normalize_history
 
@@ -12,6 +12,10 @@ def _text(part: ContentPart) -> str:
     return part.text
 
 
+def _tool_call(call_id: str, name: str = "Shell") -> ToolCall:
+    return ToolCall(id=call_id, function=ToolCall.FunctionBody(name=name, arguments="{}"))
+
+
 def test_empty_history() -> None:
     assert normalize_history([]) == []
 
@@ -123,3 +127,127 @@ def test_notification_messages_not_merged_with_user_messages() -> None:
     ]
     result = normalize_history(msgs)
     assert len(result) == 2
+
+
+# ---------------------------------------------------------------------------
+# Orphan tool_call repair (regression #2336)
+# ---------------------------------------------------------------------------
+
+
+def test_orphan_tool_call_synthesized_when_followed_by_user() -> None:
+    """A persisted assistant message whose tool response was lost must
+    not break the next API call -- a placeholder tool message is inserted."""
+    msgs = [
+        Message(role="user", content=[TextPart(text="run it")]),
+        Message(
+            role="assistant",
+            content=[TextPart(text="ok")],
+            tool_calls=[_tool_call("Shell:206")],
+        ),
+        Message(role="user", content=[TextPart(text="any update?")]),
+    ]
+    result = normalize_history(msgs)
+
+    assert [m.role for m in result] == ["user", "assistant", "tool", "user"]
+    assert result[2].tool_call_id == "Shell:206"
+    assert "interrupted" in _text(result[2].content[0])
+
+
+def test_orphan_tool_call_synthesized_at_history_tail() -> None:
+    """Assistant with tool_calls at the very end of history (no follower)
+    should still get a placeholder so the next /resume turn is valid."""
+    msgs = [
+        Message(role="user", content=[TextPart(text="run it")]),
+        Message(
+            role="assistant",
+            content=[TextPart(text="ok")],
+            tool_calls=[_tool_call("Shell:1")],
+        ),
+    ]
+    result = normalize_history(msgs)
+
+    assert [m.role for m in result] == ["user", "assistant", "tool"]
+    assert result[2].tool_call_id == "Shell:1"
+
+
+def test_complete_tool_response_not_duplicated() -> None:
+    """A well-formed assistant+tool pair must round-trip unchanged."""
+    msgs = [
+        Message(role="user", content=[TextPart(text="hi")]),
+        Message(
+            role="assistant",
+            content=[TextPart(text="")],
+            tool_calls=[_tool_call("t1")],
+        ),
+        Message(role="tool", content=[TextPart(text="result")], tool_call_id="t1"),
+    ]
+    result = normalize_history(msgs)
+
+    assert [m.role for m in result] == ["user", "assistant", "tool"]
+    assert result[2].tool_call_id == "t1"
+    assert _text(result[2].content[0]) == "result"
+
+
+def test_partial_orphan_only_missing_ids_synthesized() -> None:
+    """When parallel tool_calls have a mix of responded/missing ids,
+    only the missing ones get placeholders."""
+    msgs = [
+        Message(
+            role="assistant",
+            content=[TextPart(text="")],
+            tool_calls=[_tool_call("t1"), _tool_call("t2"), _tool_call("t3")],
+        ),
+        Message(role="tool", content=[TextPart(text="r1")], tool_call_id="t1"),
+        Message(role="tool", content=[TextPart(text="r3")], tool_call_id="t3"),
+        Message(role="user", content=[TextPart(text="next")]),
+    ]
+    result = normalize_history(msgs)
+
+    tool_msgs = [m for m in result if m.role == "tool"]
+    assert sorted(m.tool_call_id for m in tool_msgs if m.tool_call_id) == ["t1", "t2", "t3"]
+
+    synth = next(m for m in tool_msgs if m.tool_call_id == "t2")
+    assert "interrupted" in _text(synth.content[0])
+
+
+def test_multiple_assistant_tool_call_groups_independent() -> None:
+    """Repair must operate per assistant message; an earlier orphan
+    must not consume responses meant for a later assistant message."""
+    msgs = [
+        Message(
+            role="assistant",
+            content=[TextPart(text="")],
+            tool_calls=[_tool_call("a1")],
+        ),
+        # a1 orphan
+        Message(role="user", content=[TextPart(text="continue")]),
+        Message(
+            role="assistant",
+            content=[TextPart(text="")],
+            tool_calls=[_tool_call("a2")],
+        ),
+        Message(role="tool", content=[TextPart(text="ok")], tool_call_id="a2"),
+    ]
+    result = normalize_history(msgs)
+
+    assert [m.role for m in result] == [
+        "assistant",
+        "tool",
+        "user",
+        "assistant",
+        "tool",
+    ]
+    assert result[1].tool_call_id == "a1"
+    assert "interrupted" in _text(result[1].content[0])
+    assert result[4].tool_call_id == "a2"
+    assert _text(result[4].content[0]) == "ok"
+
+
+def test_assistant_without_tool_calls_untouched() -> None:
+    msgs = [
+        Message(role="user", content=[TextPart(text="hi")]),
+        Message(role="assistant", content=[TextPart(text="hello")]),
+        Message(role="user", content=[TextPart(text="bye")]),
+    ]
+    result = normalize_history(msgs)
+    assert [m.role for m in result] == ["user", "assistant", "user"]