fix(chat): keep ACP session recovery from failing on restored settings

Recovery observes the configuration a provider restores for an existing
session; it does not apply a new request. ACP skips the reapply entirely
and some providers omit configOptions from session/load, so the restored
model or reasoning effort can legitimately differ from the request that
originally opened the session.

The resume path nevertheless passed the stored request to
_observed_model_settings as a required confirmation. Any ACP conversation
opened with a non-default model or effort then failed to reopen after a
Chat restart: the controller marked the session resume_failed and rendered
"The previous Agent conversation could not be restored." This also turned
the daily full-public loopx-chat-acp-adapter-smoke red.

Record the restored identity on recovery instead, while the creation path
keeps the confirmation check for the value it just applied. The original
request stays recorded as requested_agent_model_id and
requested_agent_reasoning_effort provenance.

Signed-off-by: JasonBuildAI <jasonbuildai@gmail.com>

Author: JasonBuildAI

loopx/chat_runtime.py

@@ -992,8 +992,8 @@ def _ensure_adapter_locked(
                 codex_home_override=codex_profile_home(self.codex_home, session.get("codex_chat_configuration") or {}),
             )
             try:
-                observed_settings = self._observed_model_settings(adapter, session.get("requested_agent_reasoning_effort"),
-                                                                  session.get("requested_agent_model_id"))
+                # Recovery reads back the provider-restored configuration.
+                observed_settings = self._observed_model_settings(adapter)
                 if session.get("upstream_mode") == CODEX_GOAL_CHAT_MODE:
                     if not isinstance(adapter, CodexAppServerAdapter):
                         raise CodexChatAgentError("Native Goal continuation requires a Codex adapter.", error_code="native_goal_adapter_mismatch", gate=None)

tests/test_chat_reasoning_selection.py

@@ -50,8 +50,12 @@ def options():
     request = json.loads(line)
     with capture.open("a") as out: out.write(json.dumps(request)+"\n")
     method = request.get("method")
-    if method == "initialize": result = {"protocolVersion": 1, "agentCapabilities": {}}
+    if method == "initialize": result = {"protocolVersion": 1, "agentCapabilities": {"loadSession": True}}
     elif method == "session/new": result = {"sessionId": "fixture-session", "configOptions": options()}
+    elif method == "session/load":
+        # Recovery restores the provider's own configuration, not the request.
+        selected.update({"model": "first", "depth": "low"})
+        result = {"configOptions": options()}
     elif method == "session/set_config_option":
         if behavior != "ignore" or request["params"]["configId"] != "depth":
             selected[request["params"]["configId"]] = request["params"]["value"]
@@ -181,6 +185,47 @@ def test_registered_acp_discovery_and_selection_share_native_configuration(tmp_p
         runtime.close()
 
 
+def test_acp_restart_records_provider_restored_settings_instead_of_failing(tmp_path):
+    """Recovery reports the provider's configuration; that is not a failed request."""
+    script, capture = tmp_path / "agent.py", tmp_path / "capture.jsonl"
+    script.write_text(ACP_CONFIG)
+    store = ChatSessionStore(tmp_path / "runtime")
+    endpoints = AgentEndpointRegistry(store.root)
+    endpoints.upsert({"agent_id": "fixture-acp", "display_name": "Fixture ACP", "location": "local",
+                      "command": [sys.executable, str(script), str(capture), "apply"]})
+    first = ChatRuntimeController(store=store, codex_bin="unavailable-fixture-codex",
+                                  endpoint_registry=endpoints)
+    try:
+        session, was_resumed = first.open_session(
+            goal_id="fixture-restart", agent_id="fixture-acp", work_dir=tmp_path,
+            objective="Resume after a Chat restart.", mode="new",
+            agent_model_id="second", agent_reasoning_effort="high")
+        assert was_resumed is False
+        assert session["agent_model_id"] == "second"
+        assert session["requested_agent_model_id"] == "second"
+        assert session["agent_reasoning_effort"] == "high"
+    finally:
+        first.close()
+
+    second = ChatRuntimeController(store=store, codex_bin="unavailable-fixture-codex",
+                                   endpoint_registry=endpoints)
+    try:
+        restored, was_resumed = second.open_session(
+            goal_id="fixture-restart", agent_id="fixture-acp", work_dir=tmp_path,
+            objective="Resume after a Chat restart.", mode="resume_latest")
+        assert was_resumed is True
+        assert restored["session_id"] == session["session_id"]
+        # The ACP provider restores its own values.  Record that readback instead
+        # of rejecting a valid recovery for not echoing the original request.
+        assert restored["agent_model_id"] == "first"
+        assert restored["agent_reasoning_effort"] == "low"
+        assert restored["requested_agent_model_id"] == "second"
+        assert any(json.loads(line)["method"] == "session/load"
+                   for line in capture.read_text().splitlines())
+    finally:
+        second.close()
+
+
 def test_codex_catalog_failure_is_retryable_without_inventing_current_model(tmp_path, monkeypatch):
     runtime = ChatRuntimeController(store=ChatSessionStore(tmp_path / "runtime"), codex_bin="fixture")
     monkeypatch.setattr(runtime, "capabilities", lambda: [{"agent_id": "codex", "available": True}])