refactor(sanitizer): extract shared _classify_tool_call_orphans to eliminate drift

agent/agent_runtime_helpers.py

@@ -2368,6 +2368,46 @@ def _strip_tool_suffix(s: str) -> str | None:
 
 
 
+def _classify_tool_call_orphans(
+    messages: List[Dict[str, Any]],
+) -> tuple:
+    """Return (surviving_call_ids, result_call_ids, orphaned_results, missing_results).
+
+    Collects tool-call IDs from both assistant ``tool_calls`` (using the
+    ``call_id || id`` resolution that ``_get_tool_call_id_static`` uses)
+    and from ``tool`` messages, then classifies orphans:
+
+    - ``orphaned_results``: tool messages whose ``tool_call_id`` has no
+      matching assistant tool_call in the window.
+    - ``missing_results``: assistant tool_calls whose results were dropped
+      (no matching ``tool`` message found).
+
+    This is the single source of truth for orphan detection. The caller
+    decides whether to insert stubs (pre-API sanitizer) or strip orphans
+    (context compressor) — but the *detection* logic is shared so dedup
+    and id-resolution rules never drift between the two sites (#58357).
+    """
+    surviving_call_ids: set = set()
+    for msg in messages:
+        if msg.get("role") == "assistant":
+            for tc in msg.get("tool_calls") or []:
+                cid = _ra().AIAgent._get_tool_call_id_static(tc)
+                if cid:
+                    surviving_call_ids.add(cid)
+
+    result_call_ids: set = set()
+    for msg in messages:
+        if msg.get("role") == "tool":
+            cid = (msg.get("tool_call_id") or "").strip()
+            if cid:
+                result_call_ids.add(cid)
+
+    orphaned_results = result_call_ids - surviving_call_ids
+    missing_results = surviving_call_ids - result_call_ids
+    return surviving_call_ids, result_call_ids, orphaned_results, missing_results
+
+
+
 def sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]]:
     """Fix orphaned tool_call / tool_result pairs before every LLM call.
 
@@ -2471,23 +2511,9 @@ def sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]
             elif isinstance(tc, dict):
                 tc["function"] = {"name": _EMPTY_NAME_SENTINEL, "arguments": "{}"}
 
-    surviving_call_ids: set = set()
-    for msg in messages:
-        if msg.get("role") == "assistant":
-            for tc in msg.get("tool_calls") or []:
-                cid = _ra().AIAgent._get_tool_call_id_static(tc)
-                if cid:
-                    surviving_call_ids.add(cid)
-
-    result_call_ids: set = set()
-    for msg in messages:
-        if msg.get("role") == "tool":
-            cid = (msg.get("tool_call_id") or "").strip()
-            if cid:
-                result_call_ids.add(cid)
+    _sv, _rs, orphaned_results, missing_results = _classify_tool_call_orphans(messages)
 
     # 1. Drop tool results with no matching assistant call
-    orphaned_results = result_call_ids - surviving_call_ids
     if orphaned_results:
         messages = [
             m for m in messages
@@ -2499,7 +2525,6 @@ def sanitize_api_messages(messages: List[Dict[str, Any]]) -> List[Dict[str, Any]
         )
 
     # 2. Inject stub results for calls whose result was dropped
-    missing_results = surviving_call_ids - result_call_ids
     if missing_results:
         patched: List[Dict[str, Any]] = []
         for msg in messages:

agent/context_compressor.py

@@ -2187,23 +2187,11 @@ def _sanitize_tool_pairs(self, messages: List[Dict[str, Any]]) -> List[Dict[str,
         silently dropped by the repair pass, re-exposing the original orphans.
         Stripping at the source avoids this entire class of mismatch.
         """
-        surviving_call_ids: set = set()
-        for msg in messages:
-            if msg.get("role") == "assistant":
-                for tc in msg.get("tool_calls") or []:
-                    cid = self._get_tool_call_id(tc)
-                    if cid:
-                        surviving_call_ids.add(cid)
+        from agent.agent_runtime_helpers import _classify_tool_call_orphans
 
-        result_call_ids: set = set()
-        for msg in messages:
-            if msg.get("role") == "tool":
-                cid = msg.get("tool_call_id")
-                if cid:
-                    result_call_ids.add(cid)
+        _sv, _rs, orphaned_results, missing_results = _classify_tool_call_orphans(messages)
 
         # 1. Remove tool results whose call_id has no matching assistant tool_call
-        orphaned_results = result_call_ids - surviving_call_ids
         if orphaned_results:
             messages = [
                 m for m in messages
@@ -2213,10 +2201,7 @@ def _sanitize_tool_pairs(self, messages: List[Dict[str, Any]]) -> List[Dict[str,
                 logger.info("Compression sanitizer: removed %d orphaned tool result(s)", len(orphaned_results))
 
         # 2. Strip orphaned tool_calls from assistant messages whose results
-        #    were dropped.  Stripping is preferred over inserting stub results
-        #    because stubs can be dropped by downstream repair_message_sequence
-        #    when call_id != id (Codex Responses API format), re-exposing orphans.
-        missing_results = surviving_call_ids - result_call_ids
+        #    were dropped.
         if missing_results:
             for msg in messages:
                 if msg.get("role") != "assistant":

scripts/release.py

@@ -64,6 +64,7 @@
     "wyuebei@gmail.com": "wyuebei-cloud",  # PR #56640 salvage (hermes journey: replace GNU-only %-d strftime with dt.day for Windows)
     "yingwaizhiying@gmail.com": "msh01",  # PR #58250 salvage (telegram: wall-clock init timeout via daemon-thread deadline + abandon the shielded initialize task on timeout so the retry ladder advances instead of hanging on attempt 1/8 under s6 supervision; #58236). Also covers PR #58276 salvage (compression: preserve a real user turn after compaction; #55677).
     "danilo@falcao.org": "danilofalcao",  # PR #56674 salvage (update: skip unsupported platform.matrix lazy refresh on native Windows — python-olm has no Windows wheel)
+    "ishengeqi@163.com": "isheng-eqi",
     "huanshan5195@users.noreply.github.com": "huanshan5195",  # PR #57601 salvage (custom-provider: emit reasoning_effort at the live CustomProfile path so GLM-5.2/ARK/vLLM/Ollama endpoints receive it; + "max" reasoning level)
     "infinitycrew39@gmail.com": "infinitycrew39",  # PR #56431 salvage (honor live vLLM context limits on local endpoints)
     "jonathan.kovacs999@gmail.com": "CocaKova",  # PR #57692 salvage (cron: run jobs under the profile secret scope so get_secret does not fail-close with UnscopedSecretError under profile isolation)

tests/run_agent/test_message_sequence_repair.py

@@ -770,3 +770,64 @@ def test_sanitize_preserves_populated_tool_calls():
     out = sanitize_api_messages(list(messages))
     assistant = [m for m in out if m.get("role") == "assistant"][0]
     assert [tc["id"] for tc in assistant["tool_calls"]] == ["call_Z"]
+
+
+# ── _classify_tool_call_orphans ─────────────────────────────────────────
+
+def test_classify_orphans_empty():
+    from agent.agent_runtime_helpers import _classify_tool_call_orphans
+    sv, rs, orphaned, missing = _classify_tool_call_orphans([])
+    assert sv == set()
+    assert rs == set()
+    assert orphaned == set()
+    assert missing == set()
+
+
+def test_classify_orphans_clean_pair():
+    from agent.agent_runtime_helpers import _classify_tool_call_orphans
+    messages = [
+        {"role": "assistant", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "f", "arguments": "{}"}}]},
+        {"role": "tool", "tool_call_id": "call_1", "content": "ok"},
+    ]
+    sv, rs, orphaned, missing = _classify_tool_call_orphans(messages)
+    assert sv == {"call_1"}
+    assert rs == {"call_1"}
+    assert orphaned == set()
+    assert missing == set()
+
+
+def test_classify_orphans_detects_orphaned_result():
+    from agent.agent_runtime_helpers import _classify_tool_call_orphans
+    messages = [
+        {"role": "tool", "tool_call_id": "orphan_1", "content": "no matching call"},
+    ]
+    sv, rs, orphaned, missing = _classify_tool_call_orphans(messages)
+    assert orphaned == {"orphan_1"}
+    assert missing == set()
+
+
+def test_classify_orphans_detects_missing_result():
+    from agent.agent_runtime_helpers import _classify_tool_call_orphans
+    messages = [
+        {"role": "assistant", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "f", "arguments": "{}"}}]},
+    ]
+    sv, rs, orphaned, missing = _classify_tool_call_orphans(messages)
+    assert orphaned == set()
+    assert missing == {"call_1"}
+
+
+def test_classify_orphans_mixed():
+    from agent.agent_runtime_helpers import _classify_tool_call_orphans
+    messages = [
+        {"role": "assistant", "tool_calls": [
+            {"id": "call_A", "type": "function", "function": {"name": "f", "arguments": "{}"}},
+            {"id": "call_B", "type": "function", "function": {"name": "g", "arguments": "{}"}},
+        ]},
+        {"role": "tool", "tool_call_id": "call_A", "content": "ok"},
+        {"role": "tool", "tool_call_id": "call_C", "content": "orphan"},
+    ]
+    sv, rs, orphaned, missing = _classify_tool_call_orphans(messages)
+    assert sv == {"call_A", "call_B"}
+    assert rs == {"call_A", "call_C"}
+    assert orphaned == {"call_C"}
+    assert missing == {"call_B"}