Import AITURK IDE 1.0.0-beta.1 from Hermes 63279301; preserve MIT license

tests/gateway/test_replay_entry_fields.py

@@ -0,0 +1,145 @@
+"""Tests for ``gateway.run._build_replay_entry``.
+
+The gateway rebuilds ``agent_history`` from the persisted transcript on every
+turn (unlike the CLI, which keeps the live in-memory message list).  When a
+pure-text assistant turn (no ``tool_calls``) is replayed, the simple-text
+branch in ``run_sync`` used to whitelist only three reasoning fields:
+``reasoning``, ``reasoning_details``, ``codex_reasoning_items``.
+
+That whitelist predated three fields the DB now persists:
+``reasoning_content``, ``codex_message_items``, and ``finish_reason``.  The
+unrecovered drop of ``codex_message_items`` in particular kills prefix-cache
+hits for OpenAI Codex Responses API users — OpenAI's docs require the
+``phase`` field be replayed on every assistant message.
+
+These tests pin the expanded whitelist so it doesn't regress.
+"""
+from __future__ import annotations
+
+
+from gateway.run import _ASSISTANT_REPLAY_FIELDS, _build_replay_entry
+
+
+class TestBuildReplayEntry:
+    def test_user_message_has_only_role_and_content(self):
+        entry = _build_replay_entry(
+            "user",
+            "hello",
+            {"role": "user", "content": "hello", "reasoning": "leak", "extra": "drop"},
+        )
+        assert entry == {"role": "user", "content": "hello"}
+
+    def test_tool_message_has_only_role_and_content(self):
+        # Tool messages aren't routed through this helper in production
+        # (they take the rich-passthrough branch), but the helper itself
+        # must not leak reasoning fields onto non-assistant roles even if
+        # someone calls it incorrectly.
+        entry = _build_replay_entry(
+            "tool",
+            "result",
+            {"role": "tool", "content": "result", "reasoning": "leak"},
+        )
+        assert entry == {"role": "tool", "content": "result"}
+
+
+    def test_assistant_drops_falsy_reasoning(self):
+        """Empty/None reasoning fields stay dropped (matching PR #2974
+        behaviour) — empty strings/lists for these fields carry no info."""
+        msg = {
+            "role": "assistant",
+            "content": "answer",
+            "reasoning": "",
+            "reasoning_details": [],
+            "codex_reasoning_items": [],
+            "codex_message_items": [],
+            "finish_reason": "",
+        }
+        entry = _build_replay_entry("assistant", "answer", msg)
+        assert entry == {"role": "assistant", "content": "answer"}
+
+
+    def test_assistant_preserves_all_six_fields_together(self):
+        details = [{"type": "reasoning.summary", "summary": "s"}]
+        codex_items = [{"type": "reasoning", "encrypted_content": "b"}]
+        msg_items = [
+            {
+                "type": "message",
+                "role": "assistant",
+                "phase": "final_answer",
+                "content": [{"type": "output_text", "text": "x"}],
+            }
+        ]
+        msg = {
+            "role": "assistant",
+            "content": "answer",
+            "reasoning": "thinking",
+            "reasoning_content": "structured",
+            "reasoning_details": details,
+            "codex_reasoning_items": codex_items,
+            "codex_message_items": msg_items,
+            "finish_reason": "stop",
+        }
+        entry = _build_replay_entry("assistant", "answer", msg)
+        assert entry["reasoning"] == "thinking"
+        assert entry["reasoning_content"] == "structured"
+        assert entry["reasoning_details"] == details
+        assert entry["codex_reasoning_items"] == codex_items
+        assert entry["codex_message_items"] == msg_items
+        assert entry["finish_reason"] == "stop"
+
+
+    def test_replay_fields_constant_is_stable(self):
+        """Pin the whitelist explicitly so accidental renames are caught."""
+        assert _ASSISTANT_REPLAY_FIELDS == (
+            "reasoning",
+            "reasoning_content",
+            "reasoning_details",
+            "codex_reasoning_items",
+            "codex_message_items",
+            "finish_reason",
+        )
+
+    def test_unrelated_keys_are_ignored(self):
+        """Random keys on the message must not leak into the replay entry."""
+        msg = {
+            "role": "assistant",
+            "content": "answer",
+            "timestamp": 12345.6,
+            "internal_marker": "should not flow",
+            "tool_call_id": "should not be set on simple-text branch",
+        }
+        entry = _build_replay_entry("assistant", "answer", msg)
+        assert "timestamp" not in entry
+        assert "internal_marker" not in entry
+        assert "tool_call_id" not in entry
+
+
+class TestReplayEntryApiContentSidecar:
+    """The api_content sidecar (persist-what-you-send) must survive the
+    gateway's transcript→agent_history rebuild, or the whole prompt-cache
+    fix is inert on gateway platforms — but only when this pipeline did not
+    rewrite the content (a rewrite means different bytes must replay)."""
+
+    def test_user_forwards_api_content_when_content_unchanged(self):
+        msg = {"role": "user", "content": "hi", "api_content": "hi\n\nCTX"}
+        entry = _build_replay_entry("user", "hi", msg)
+        assert entry["api_content"] == "hi\n\nCTX"
+        assert entry["content"] == "hi"
+
+    def test_assistant_forwards_api_content_when_content_unchanged(self):
+        msg = {"role": "assistant", "content": "a", "api_content": "a <memory-context>"}
+        entry = _build_replay_entry("assistant", "a", msg)
+        assert entry["api_content"] == "a <memory-context>"
+
+
+class TestGatewayHistoryBuildForwardsSidecar:
+    def test_end_to_end_history_build_keeps_sidecar(self):
+        from gateway.run import _build_gateway_agent_history
+
+        history = [
+            {"role": "user", "content": "hi", "api_content": "hi\n\nCTX", "timestamp": 123.0},
+            {"role": "assistant", "content": "hello"},
+        ]
+        agent_history, _obs = _build_gateway_agent_history(history)
+        assert agent_history[0]["api_content"] == "hi\n\nCTX"
+