fix(gateway): pass full transcript to compressor instead of filtered messages (#58551)

Both gateway compression entry points (session-hygiene auto-compress in
run.py; manual /compress in slash_commands.py) filtered the transcript
to user/assistant-only, content-bearing messages before calling
_compress_context. That starved the compressor:

- tool results are usually the bulk of the context, and
  _prune_old_tool_results never saw them
- short filtered histories tripped the protect-first/last early-return,
  so compression became a no-op even on huge sessions
- assistant tool_calls stubs (content=None) were dropped, so even the
  summary lost the tool activity

Pass user/assistant/tool messages through intact, matching what the
agent loop itself feeds _compress_context.

Port of PR #3854 onto current main (the manual-compress handler moved
from run.py to slash_commands.py since the PR branched); regression test
asserts tool messages reach the compressor.

Authored-by: David Zhang <david.d.zhang@gmail.com> (@Git-on-my-level)

Co-authored-by: David Zhang <david.d.zhang@gmail.com>

gateway/run.py

@@ -10794,11 +10794,18 @@ class GatewayRunner(GatewayAuthorizationMixin, GatewayKanbanWatchersMixin, Gatew
                             user_config=_hyg_data if isinstance(_hyg_data, dict) else None,
                         )
                         if _hyg_runtime.get("api_key"):
+                            # Pass the FULL transcript (tool results included).
+                            # Filtering to user/assistant-only starved the
+                            # compressor: tool results are usually the bulk of
+                            # the context, _prune_old_tool_results never saw
+                            # them, and short filtered histories tripped the
+                            # protect-first/last early-return so nothing was
+                            # compressed at all (#3854). The agent loop passes
+                            # its full message list to _compress_context — the
+                            # gateway now matches.
                             _hyg_msgs = [
-                                {"role": m.get("role"), "content": m.get("content")}
-                                for m in history
-                                if m.get("role") in {"user", "assistant"}
-                                and m.get("content")
+                                m for m in history
+                                if m.get("role") in {"user", "assistant", "tool"}
                             ]
 
                             if len(_hyg_msgs) >= 4:

gateway/slash_commands.py

@@ -3140,10 +3140,14 @@ class GatewaySlashCommandsMixin:
             if not runtime_kwargs.get("api_key"):
                 return t("gateway.compress.no_provider")
 
+            # Pass the FULL transcript (tool results included) — same
+            # rationale as the session-hygiene auto-compress in
+            # gateway/run.py (#3854): filtering to user/assistant-only
+            # starves the compressor's tool-result pruning and can trip the
+            # protect-first/last early-return on short filtered histories.
             msgs = [
-                {"role": m.get("role"), "content": m.get("content")}
-                for m in history
-                if m.get("role") in {"user", "assistant"} and m.get("content")
+                m for m in history
+                if m.get("role") in {"user", "assistant", "tool"}
             ]
 
             # Boundary-aware split: only the head is summarized; the most

tests/gateway/test_compress_command.py

@@ -483,3 +483,51 @@ async def test_compress_command_overrides_stale_resolver_identity():
     # Source-derived identity overrides the stale resolver values, passed once.
     assert kwargs["platform"] == "telegram"
     assert kwargs["gateway_session_key"] == runner._session_key_for_source(_make_source())
+
+
+@pytest.mark.asyncio
+async def test_compress_command_passes_tool_messages_to_compressor():
+    """Tool results must reach _compress_context (#3854).
+
+    Filtering the transcript to user/assistant-only starved the
+    compressor's tool-result pruning — tool messages are usually the bulk
+    of the context.
+    """
+    history = [
+        {"role": "user", "content": "run it"},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [{"id": "t1", "type": "function",
+                            "function": {"name": "x", "arguments": "{}"}}],
+        },
+        {"role": "tool", "content": "BIG RESULT " * 50, "tool_call_id": "t1"},
+        {"role": "assistant", "content": "done"},
+        {"role": "user", "content": "thanks"},
+        {"role": "assistant", "content": "np"},
+    ]
+    runner = _make_runner(history)
+    agent_instance = MagicMock()
+    agent_instance.shutdown_memory_provider = MagicMock()
+    agent_instance.close = MagicMock()
+    agent_instance._cached_system_prompt = ""
+    agent_instance.tools = None
+    agent_instance.context_compressor.has_content_to_compress.return_value = True
+    agent_instance.session_id = "sess-1"
+    agent_instance._compress_context.return_value = (list(history), "")
+
+    with (
+        patch("gateway.run._resolve_runtime_agent_kwargs", return_value={"api_key": "test-key"}),
+        patch("gateway.run._resolve_gateway_model", return_value="test-model"),
+        patch("run_agent.AIAgent", return_value=agent_instance),
+        patch("agent.model_metadata.estimate_request_tokens_rough", return_value=100),
+    ):
+        await runner._handle_compress_command(_make_event())
+
+    args, _kwargs = agent_instance._compress_context.call_args
+    passed = args[0]
+    roles = [m.get("role") for m in passed]
+    assert "tool" in roles, f"tool messages filtered out: {roles}"
+    # Assistant tool_calls stubs (content=None) must survive too, or the
+    # tool message would dangle without its call.
+    assert any(m.get("tool_calls") for m in passed), "assistant tool_calls stub dropped"