refactor(turn): collapse preflight estimator branch, build_api_messages pops, docstrings; add turn_context_compaction tests

agent/turn_context.py

@@ -1,9 +1,10 @@
 """Per-turn setup for ``run_conversation`` (the turn prologue).
 
 ``build_turn_context`` runs the once-per-turn setup (stdio guard, sanitization, prompt
-restore-or-build, session row, preflight compression, pre_llm_call hook, prefetch,
+restore-or-build, session row, idle/preflight compaction — see
-persistence) mutating ``agent`` exactly as the inline code did, and returns a
+``turn_context_compaction`` — pre_llm_call hook, prefetch, persistence), mutating
-``TurnContext`` carrying only the locals the loop reads back."""
+``agent`` as the loop expects, and returns a ``TurnContext`` carrying only the locals
+the loop reads back. ``build_api_messages`` builds the wire copy for one API call."""
 
 from __future__ import annotations
 
@@ -64,17 +65,11 @@ def _preflight_request_tokens(
             "using generic transcript estimate",
             exc_info=True,
         )
-    if _agent_stale_thinking_on_wire(agent):
     return estimate_request_tokens_rough(
         messages,
         system_prompt=system_prompt or "",
         tools=tools,
-        )
+        charge_stale_thinking=_agent_stale_thinking_on_wire(agent),
-    return estimate_request_tokens_rough(
-        messages,
-        system_prompt=system_prompt or "",
-        tools=tools,
-        charge_stale_thinking=False,
     )
 
 
@@ -148,10 +143,8 @@ def extract_api_content_sidecar(msg: Mapping[str, Any]) -> Optional[str]:
 
 
 def consume_gateway_turn_context_notes(agent: Any) -> str:
-    """Pop the gateway's per-turn must-deliver notes off the agent (one-shot).
+    """Pop the gateway's per-turn must-deliver notes off the agent (one-shot, so the
-
+    system prompt stays byte-stable and a cached agent never replays a stale note)."""
-    Staged on ``agent._gateway_turn_context_notes``; consuming them keeps the system
-    prompt byte-stable and prevents a cached agent replaying a stale note."""
     notes = getattr(agent, "_gateway_turn_context_notes", "") or ""
     if hasattr(agent, "_gateway_turn_context_notes"):
         try:
@@ -1066,8 +1059,6 @@ def build_api_messages(
     from agent.agent_runtime_helpers import fill_empty_non_final_wire_payload
     from agent.conversation_loop import _clone_message_for_send
 
-    _ext_prefetch_cache = ext_prefetch_cache
-    _plugin_user_context = plugin_user_context
     api_messages = []
     for idx, msg in enumerate(messages):
 
@@ -1098,9 +1089,7 @@ def build_api_messages(
             else:
                 # Callers that bypass the prologue stamping: compose live.
                 _composed = compose_user_api_content(
-                    api_msg.get("content", ""),
+                    api_msg.get("content", ""), ext_prefetch_cache, plugin_user_context
-                    _ext_prefetch_cache,
-                    _plugin_user_context,
                 )
                 if _composed is not None:
                     api_msg["content"] = _composed
@@ -1118,13 +1107,10 @@ def build_api_messages(
         # This ensures multi-turn reasoning context is preserved
         agent._copy_reasoning_content_for_api(msg, api_msg)
 
-        # Remove 'reasoning' field - it's for trajectory storage only
+        # 'reasoning' is trajectory-only (copied to 'reasoning_content' above);
-        # We've copied it to 'reasoning_content' for the API above
+        # finish_reason is rejected by strict APIs (e.g. Mistral).
-        if "reasoning" in api_msg:
+        api_msg.pop("reasoning", None)
-            api_msg.pop("reasoning")
+        api_msg.pop("finish_reason", None)
-        # Remove finish_reason - not accepted by strict APIs (e.g. Mistral)
-        if "finish_reason" in api_msg:
-            api_msg.pop("finish_reason")
         # Fill empty non-final user/assistant wire copies so the pre-call sanitizer
         # stops re-healing and flooding errors.log; durable history is untouched.
         # After the reasoning copy so thinking-only turns keep payload (#96870).

tests/agent/test_turn_context_compaction.py

@@ -0,0 +1,80 @@
+"""Unit tests for ``agent.turn_context_compaction`` (turn-start compaction extracted
+from ``build_turn_context``)."""
+
+from types import SimpleNamespace
+from unittest.mock import MagicMock, patch
+
+from agent.turn_context_compaction import (
+    CompactionOutcome,
+    _codex_native_auto_compaction,
+    _rearm_uncompressed_overflow_warn,
+    run_turn_start_compaction,
+)
+
+
+def _agent(**kw):
+    compressor = SimpleNamespace(
+        protect_first_n=3, protect_last_n=3, threshold_tokens=1_000, context_length=8_000,
+        summary_target_ratio=0.5,
+    )
+    base = dict(
+        compression_enabled=False, context_compressor=compressor, session_id="s1",
+        model="m", _clear_context_overflow_warn=MagicMock(),
+    )
+    base.update(kw)
+    return SimpleNamespace(**base)
+
+
+def test_codex_native_auto_compaction_gate():
+    assert _codex_native_auto_compaction(
+        SimpleNamespace(api_mode="codex_app_server", codex_app_server_auto_compaction="native")
+    )
+    assert _codex_native_auto_compaction(
+        SimpleNamespace(api_mode="codex_app_server", codex_app_server_auto_compaction="OFF")
+    )
+    assert not _codex_native_auto_compaction(
+        SimpleNamespace(api_mode="codex_app_server", codex_app_server_auto_compaction="hermes")
+    )
+    assert not _codex_native_auto_compaction(SimpleNamespace(api_mode="chat_completions"))
+
+
+def test_disabled_compression_rearms_overflow_warn_when_under_window():
+    agent = _agent()
+    msgs = [{"role": "user", "content": "hi"}]
+    out = run_turn_start_compaction(
+        agent, messages=msgs, system_message=None, active_system_prompt="sys",
+        conversation_history=None, current_turn_user_idx=0, user_message="hi",
+        effective_task_id="t",
+    )
+    assert isinstance(out, CompactionOutcome)
+    assert out.messages is msgs and out.current_turn_user_idx == 0
+    assert out.compressed is False and out.blocked is False
+    agent._clear_context_overflow_warn.assert_called_once()
+    assert agent._turn_received_provider_response is False
+    assert agent._turn_preflight_display_snapshot is None
+
+
+def test_multimodal_content_forces_real_estimate():
+    agent = _agent()
+    msgs = [{"role": "user", "content": [{"type": "text", "text": "x"}]}]
+    with patch(
+        "agent.turn_context._preflight_request_tokens", return_value=9_999
+    ) as est:
+        _rearm_uncompressed_overflow_warn(agent, msgs, "sys")
+    est.assert_called_once()
+    agent._clear_context_overflow_warn.assert_not_called()
+
+
+def test_preflight_gate_skips_small_transcripts():
+    agent = _agent(compression_enabled=True)
+    agent.context_compressor.should_compress = MagicMock()
+    msgs = [{"role": "user", "content": "hi"}]
+    with patch("agent.turn_context._preflight_request_tokens") as est:
+        out = run_turn_start_compaction(
+            agent, messages=msgs, system_message=None, active_system_prompt="sys",
+            conversation_history=None, current_turn_user_idx=0, user_message="hi",
+            effective_task_id="t",
+        )
+    est.assert_not_called()
+    agent.context_compressor.should_compress.assert_not_called()
+    assert out.messages is msgs