codex-agent-framework 0.1.60__tar.gz → 0.1.61__tar.gz

{codex_agent_framework-0.1.60/codex_agent_framework.egg-info → codex_agent_framework-0.1.61}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-agent-framework
-Version: 0.1.60
+Version: 0.1.61
 Summary: A lightweight event-driven Codex agent runtime.
 Author: Baptiste
 License-Expression: MIT

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/builtin_plugins/context/__init__.py

@@ -37,6 +37,7 @@ Tool result wrappers expose stable `wrapper_id` values in their XML representati
             vanished_wrapper_messages=status.get("vanished_wrapper_messages", 0),
             total_wrapper_messages=status.get("total_wrapper_messages", 0),
             base_context_tokens=status.get("base_context_tokens", status.used_tokens),
+            provider_context_tokens=status.get("provider_context_tokens", 0),
             active_wrapper_tokens=status.get("active_wrapper_tokens", 0),
             vanished_wrapper_tokens=status.get("vanished_wrapper_tokens", 0),
             compactable_turns=status.get("compactable_turns", 0),
@@ -57,7 +58,7 @@ Tool result wrappers expose stable `wrapper_id` values in their XML representati
         return "\n".join([
             "Context status:",
             f"- tokens: {status.used_tokens}/{status.input_token_limit} ({status.percent:.2f}%)",
-            f"- token split base/active_wrappers/vanished_wrappers: {status.base_context_tokens}/{status.active_wrapper_tokens}/{status.vanished_wrapper_tokens}",
+            f"- token split base/vanished_wrappers/active_wrappers/providers: {status.base_context_tokens}/{status.vanished_wrapper_tokens}/{status.active_wrapper_tokens}/{status.provider_context_tokens}",
             f"- messages sent/total: {status.sent_session_messages}/{status.total_session_messages}",
             f"- persistent/temporary context messages: {status.persistent_session_messages}/{status.temporary_session_messages}",
             f"- wrappers active/vanished/total: {status.active_wrapper_messages}/{status.vanished_wrapper_messages}/{status.total_wrapper_messages}",

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/config.py

@@ -12,6 +12,8 @@ from .utils import runtime_join, text_content
 
 DEFAULT_SYSTEM_PROMPT_PATH = os.path.join(os.path.dirname(__file__), "prompts", "system_prompt.txt")
 DEFAULT_SYSTEM_PROMPT = text_content(DEFAULT_SYSTEM_PROMPT_PATH)
+DEFAULT_COMPACTION_PROMPT_PATH = os.path.join(os.path.dirname(__file__), "prompts", "compaction_prompt.txt")
+DEFAULT_COMPACTION_PROMPT = text_content(DEFAULT_COMPACTION_PROMPT_PATH)
 
 
 def parse_config_value(value):
@@ -82,6 +84,7 @@ class AgentConfig(Config):
 
     model = "gpt-5.4"
     compaction_model = "gpt-5.4-mini"
+    compaction_instructions = DEFAULT_COMPACTION_PROMPT
     system = DEFAULT_SYSTEM_PROMPT
     auto_proceed = True
     openai_api_key = None

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/context.py

@@ -14,6 +14,7 @@ class ContextManager:
     def __init__(self, agent):
         self.agent = agent
         self.token_count_coef = 1.0
+        self.last_provider_tokens = 0
 
     def provider_context_wrapper(self, provider_name, content):
         return f'<context_provider name="{provider_name}">\n{content}\n</context_provider>'
@@ -85,6 +86,8 @@ class ContextManager:
         system = [self.agent.config_manager.system_message()]
         response_tools = self.agent.tools_manager.get_response_tools()
         providers_msgs = self.get_providers_messages()
+        provider_tokens = self.response_input_token_count(providers_msgs, response_tools=[])
+        self.last_provider_tokens = provider_tokens
         max_images = self.agent.config.get("max_images", 10)
 
         if self.agent.config.prune_orphaned_tool_outputs_on_context_build:
@@ -139,7 +142,7 @@ class ContextManager:
         active_context_wrappers = [msg for msg in history if isinstance(msg, ToolResultWrapper)]
         active_wrapper_tokens = wrapper_token_count(active_context_wrappers)
         vanished_wrapper_tokens = wrapper_token_count(vanished_wrappers)
-        base_context_tokens = max(0, current_count - active_wrapper_tokens)
+        base_context_tokens = max(0, current_count - provider_tokens - active_wrapper_tokens)
         if self.agent.session:
             compaction_plan = self.agent.session.compaction_plan(
                 max_input_tokens=max_input_tokens,
@@ -174,6 +177,7 @@ class ContextManager:
             vanished_wrapper_messages=len(vanished_wrappers),
             total_wrapper_messages=len(wrappers),
             base_context_tokens=base_context_tokens,
+            provider_context_tokens=provider_tokens,
             active_wrapper_tokens=active_wrapper_tokens,
             vanished_wrapper_tokens=vanished_wrapper_tokens,
             compactable_turns=compactable_turns,

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/message.py

@@ -259,7 +259,7 @@ class CompactionMessage(DeveloperMessage):
     name = "compaction"
     content = "Previous conversation context was compacted by the backend."
     response_id = ""
-    summaries = modict.factory(list)
+    output = modict.factory(list)
     usage = modict.factory(modict)
     compacted_message_count = 0
 
@@ -268,8 +268,12 @@ class CompactionMessage(DeveloperMessage):
         super().__init__(*args, **kwargs)
         if output_items is None:
             output_items = self.pop("output_items", None)
-        if output_items is not None and not self.get("summaries"):
-            self.summaries = self.summary_items(output_items)
+        legacy_summaries = self.pop("summaries", None)
+        if not self.get("output"):
+            if output_items is not None:
+                self.output = list(output_items or [])
+            elif legacy_summaries is not None:
+                self.output = list(legacy_summaries or [])
 
     @staticmethod
     def summary_items(items):
@@ -278,8 +282,12 @@ class CompactionMessage(DeveloperMessage):
             if item.get("type") == "compaction_summary"
         ]
 
+    @property
+    def summaries(self):
+        return self.summary_items(self.get("output") or [])
+
     def to_response_format(self):
-        return list(self.get("summaries") or [])
+        return list(self.get("output") or [])
 
     def observed_token_count(self):
         usage = self.get("usage") or {}
@@ -306,7 +314,7 @@ class CompactionMessage(DeveloperMessage):
 
     def token_count_payload(self):
         sanitized = []
-        for item in self.get("summaries") or []:
+        for item in self.get("output") or []:
             entry = dict(item)
             if entry.get("type") == "compaction_summary" and entry.get("encrypted_content"):
                 entry["encrypted_content"] = "<encrypted_compaction_summary>"
@@ -320,17 +328,19 @@ class CompactionMessage(DeveloperMessage):
         return self.format_as_string_attrs(attrs)
 
     def as_string_body(self):
-        summary = next(
-            iter(self.get("summaries") or []),
-            None,
-        )
-        if not summary:
+        summaries = self.summary_items(self.get("output") or [])
+        if not summaries:
             return self.content
+        summary_lines = "\n".join(
+            f"<compaction_summary id={summary.get('id', '')!r}>opaque backend summary</compaction_summary>"
+            for summary in summaries
+        )
+        output_count = len(self.get("output") or [])
         return dedent(
             f"""{self.content}
-            <compaction_summary id={summary.get('id', '')!r}>
-            opaque backend summary
-            </compaction_summary>"""
+            <compaction_output items={output_count}>
+            {summary_lines}
+            </compaction_output>"""
         )
 
 

codex_agent_framework-0.1.61/codex_agent/prompts/compaction_prompt.txt

@@ -0,0 +1,17 @@
+You are the standalone compaction model for an agentic coding assistant conversation.
+
+Goal: replace the provided conversation history with the smallest useful compacted history that still lets the agent continue the task safely and coherently.
+
+Instructions:
+- Reduce the visible content of each retained message as aggressively as possible.
+- Keep only information that is useful for future continuation: user goals, decisions, constraints, file paths, commands/results that matter, bugs found, fixes applied, validation status, unresolved questions, TODOs, and exact identifiers such as versions, commits, tags, job ids, paths, APIs, function/class names, and error messages.
+- Remove chatter, repeated confirmations, verbose tool output, incidental logs, redundant reasoning, and obsolete branches of investigation.
+- Preserve chronology and causality when it affects future decisions.
+- The history to compact will often begin with output from a previous compaction. Treat that previous compacted context as important long-range background: carry forward its still-relevant facts, decisions, constraints, unresolved TODOs, and exact identifiers into the new final compaction, further compressed if needed, so repeated compactions keep a durable trace of distant context instead of replacing it with only the newest turns.
+- Preserve the distinction between facts observed from tools/source-of-truth and inferences or hypotheses.
+- Do not invent missing facts. If something is uncertain, mark it as uncertain.
+- If tool outputs were summarized, keep the actionable conclusions and any exact values needed later, not the full raw output.
+- The encrypted compaction summary should be dense, explicit, and highly informative. It must help the agent resume without relying on external memory/RAG.
+- Prefer compact bullet-like factual wording over prose.
+
+Return the compacted output in the format required by the compaction endpoint.

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/sessions.py

@@ -183,7 +183,7 @@ class AgentSession(modict):
             usage = dict(usage)
         message = CompactionMessage(
             response_id=result.id,
-            summaries=CompactionMessage.summary_items(result.output),
+            output=list(result.output or []),
             usage=usage or {},
             compacted_message_count=len(compacted_messages),
             compacted_turn_count=plan.compacted_turn_count,
@@ -832,6 +832,8 @@ class SessionsManager:
             model=model,
             max_input_tokens=max_input_tokens,
         )
+        if instructions is None:
+            instructions = self.agent.config.get("compaction_instructions")
         if instructions is not None:
             compact_kwargs.instructions = instructions
         if max_compacted_turns is not None:

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/status.py

@@ -33,6 +33,7 @@ class StatusManager:
             vanished_wrapper_messages=context_status.get("vanished_wrapper_messages", 0),
             total_wrapper_messages=context_status.get("total_wrapper_messages", 0),
             base_context_tokens=context_status.get("base_context_tokens", context_status.used_tokens),
+            provider_context_tokens=context_status.get("provider_context_tokens", 0),
             active_wrapper_tokens=context_status.get("active_wrapper_tokens", 0),
             vanished_wrapper_tokens=context_status.get("vanished_wrapper_tokens", 0),
             compactable_turns=context_status.get("compactable_turns", 0),
@@ -89,8 +90,9 @@ class StatusManager:
                 msg_copy.content = truncate(msg_copy.content, max_tokens=max_input_tokens, model=self.agent.config.model)
             truncated_others.append(msg_copy)
 
+        provider_tokens = int(getattr(context_manager, "last_provider_tokens", 0) or 0)
         fixed_tokens = context_manager.response_input_token_count(system + images, response_tools=response_tools)
-        current_count = fixed_tokens
+        current_count = fixed_tokens + provider_tokens
         history_start = len(truncated_others)
         for i, msg in enumerate(reversed(truncated_others)):
             msg_count = context_manager.response_input_token_count([msg], response_tools=[])
@@ -131,7 +133,7 @@ class StatusManager:
         active_context_wrappers = [msg for msg in included_others if isinstance(msg, ToolResultWrapper)]
         active_wrapper_tokens = wrapper_token_count(active_context_wrappers)
         vanished_wrapper_tokens = wrapper_token_count(vanished_wrappers)
-        base_context_tokens = max(0, current_count - active_wrapper_tokens)
+        base_context_tokens = max(0, current_count - provider_tokens - active_wrapper_tokens)
         if self.agent.session:
             compaction_plan = self.agent.session.compaction_plan(
                 max_input_tokens=max_input_tokens,
@@ -151,6 +153,7 @@ class StatusManager:
             input_token_limit=input_limit,
             percent=percent,
             fixed_tokens=fixed_tokens,
+            provider_context_tokens=provider_tokens,
             tool_spec_tokens=context_manager.response_tools_token_count(response_tools),
             token_count_source="local_estimate",
             visible_history_messages=visible_history_count,
@@ -213,6 +216,7 @@ class StatusManager:
                 vanished_wrapper_messages=estimated.get("vanished_wrapper_messages", 0),
                 total_wrapper_messages=estimated.get("total_wrapper_messages", 0),
                 base_context_tokens=estimated.get("base_context_tokens", estimated.used_tokens),
+                provider_context_tokens=estimated.get("provider_context_tokens", 0),
                 active_wrapper_tokens=estimated.get("active_wrapper_tokens", 0),
                 vanished_wrapper_tokens=estimated.get("vanished_wrapper_tokens", 0),
                 compactable_turns=estimated.get("compactable_turns", 0),
@@ -247,6 +251,7 @@ class StatusManager:
             vanished_wrapper_messages=estimated.get("vanished_wrapper_messages", 0),
             total_wrapper_messages=estimated.get("total_wrapper_messages", 0),
             base_context_tokens=estimated.get("base_context_tokens", estimated.used_tokens),
+            provider_context_tokens=estimated.get("provider_context_tokens", 0),
             active_wrapper_tokens=estimated.get("active_wrapper_tokens", 0),
             vanished_wrapper_tokens=estimated.get("vanished_wrapper_tokens", 0),
             compactable_turns=estimated.get("compactable_turns", 0),
@@ -282,7 +287,8 @@ class StatusManager:
             active_wrapper_messages=context_status.get("active_wrapper_messages", 0),
             vanished_wrapper_messages=context_status.get("vanished_wrapper_messages", 0),
             total_wrapper_messages=context_status.get("total_wrapper_messages", 0),
-            base_context_tokens=max(0, input_tokens - context_status.get("active_wrapper_tokens", 0)),
+            base_context_tokens=max(0, input_tokens - context_status.get("provider_context_tokens", 0) - context_status.get("active_wrapper_tokens", 0)),
+            provider_context_tokens=context_status.get("provider_context_tokens", 0),
             active_wrapper_tokens=context_status.get("active_wrapper_tokens", 0),
             vanished_wrapper_tokens=context_status.get("vanished_wrapper_tokens", 0),
             compactable_turns=context_status.get("compactable_turns", 0),

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent/tui/status_bar.py

@@ -26,25 +26,50 @@ class StatusBar:
 
     @classmethod
     def _context_token_segments(cls, status, used, limit):
-        if "base_context_tokens" not in status and "active_wrapper_tokens" not in status and "vanished_wrapper_tokens" not in status:
+        if (
+            "base_context_tokens" not in status
+            and "provider_context_tokens" not in status
+            and "active_wrapper_tokens" not in status
+            and "vanished_wrapper_tokens" not in status
+        ):
             return ((f"{used}/{limit} ktok", "#8b949e on #161b22"),)
-        return (
-            (cls._format_ktok(status.get("base_context_tokens") or 0), "#c9d1d9 on #161b22"),
-            (f"+{cls._format_ktok(status.get('active_wrapper_tokens') or 0)}", "italic #f0883e on #161b22"),
-            (f"+{cls._format_ktok(status.get('vanished_wrapper_tokens') or 0)}", "italic #6e7681 on #161b22"),
+        base_tokens = status.get("base_context_tokens") or 0
+        provider_tokens = status.get("provider_context_tokens") or 0
+        active_wrapper_tokens = status.get("active_wrapper_tokens") or 0
+        vanished_wrapper_tokens = status.get("vanished_wrapper_tokens") or 0
+        active_total = cls._format_ktok(base_tokens + provider_tokens + active_wrapper_tokens)
+        segments = [
+            (cls._format_ktok(base_tokens), "#c9d1d9 on #161b22"),
+        ]
+        segments.extend([
+            (f"(+{cls._format_ktok(vanished_wrapper_tokens)})", "italic #6e7681 on #161b22"),
+            (f"+{cls._format_ktok(active_wrapper_tokens)}", "italic #f0883e on #161b22"),
+        ])
+        if provider_tokens:
+            segments.append((f"+{cls._format_ktok(provider_tokens)}", "italic #58a6ff on #161b22"))
+        segments.extend([
+            ("=", "#8b949e on #161b22"),
+            (active_total, "#c9d1d9 on #161b22"),
             (f"/{limit} ktok", "#8b949e on #161b22"),
-        )
+        ])
+        return tuple(segments)
 
     @classmethod
     def _message_segments(cls, status):
         sent = int(status.get("sent_session_messages") or 0)
         total = int(status.get("total_session_messages") or 0)
         if "active_wrapper_messages" in status or "vanished_wrapper_messages" in status:
+            persistent = int(status.get("persistent_session_messages") or 0)
+            active_wrappers = int(status.get("active_wrapper_messages") or 0)
+            vanished_wrappers = int(status.get("vanished_wrapper_messages") or 0)
+            active_total = persistent + active_wrappers
             return (
-                (str(int(status.get("persistent_session_messages") or 0)), "#c9d1d9 on #161b22"),
-                (f"+{int(status.get('active_wrapper_messages') or 0)}", "italic #f0883e on #161b22"),
-                (f"+{int(status.get('vanished_wrapper_messages') or 0)}", "italic #6e7681 on #161b22"),
-                (f"/{total}", "#c9d1d9 on #161b22"),
+                (str(persistent), "#c9d1d9 on #161b22"),
+                (f"(+{vanished_wrappers})", "italic #6e7681 on #161b22"),
+                (f"+{active_wrappers}", "italic #f0883e on #161b22"),
+                ("=", "#8b949e on #161b22"),
+                (str(active_total), "#c9d1d9 on #161b22"),
+                (f"/{total}", "#8b949e on #161b22"),
             )
         if "persistent_session_messages" in status or "temporary_session_messages" in status:
             return (

codex_agent_framework-0.1.61/codex_agent/version.py

@@ -0,0 +1 @@
+__version__ = "0.1.61"

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61/codex_agent_framework.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: codex-agent-framework
-Version: 0.1.60
+Version: 0.1.61
 Summary: A lightweight event-driven Codex agent runtime.
 Author: Baptiste
 License-Expression: MIT

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/codex_agent_framework.egg-info/SOURCES.txt

@@ -89,6 +89,7 @@ codex_agent/get_text/__init__.py
 codex_agent/get_text/default_gitignore
 codex_agent/get_text/get_text.py
 codex_agent/get_text/simpler_get_text.py
+codex_agent/prompts/compaction_prompt.txt
 codex_agent/prompts/system_prompt.txt
 codex_agent/scripts/install-system-dependencies.sh
 codex_agent/server/__init__.py

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "codex-agent-framework"
-version = "0.1.60"
+version = "0.1.61"
 description = "A lightweight event-driven Codex agent runtime."
 readme = "README.md"
 requires-python = ">=3.11"

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_agent_context.py

@@ -103,7 +103,10 @@ def test_agent_session_compact_purges_compacted_prefix_before_remainder(tmp_path
     assert session.messages == [message, current]
     assert [msg.id for msg in AgentSession.load(session.session_id).messages] == [message.id, current.id]
     assert "output_items" not in message
+    assert "summaries" not in message
+    assert message.output == FakeCompactionResult.output
     assert message.summaries == [FakeCompactionResult.output[1]]
+    assert message.to_response_format() == FakeCompactionResult.output
     assert message.usage == FakeCompactionResult.usage
     assert message.count_tokens(model="gpt-5") == 12000
     assert message.compacted_message_count == 2
@@ -358,7 +361,7 @@ def test_agent_get_response_auto_compact_failure_continues_backend_call(monkeypa
     assert not any(isinstance(message, CompactionMessage) for message in agent.session.messages)
 
 
-def test_agent_compact_session_uses_mini_model_by_default():
+def test_agent_compact_session_uses_mini_model_and_instructions_by_default():
     agent = Agent(session="new")
     visible = DeveloperMessage(content="visible compactable context")
     done = AssistantMessage(content="done")
@@ -370,7 +373,8 @@ def test_agent_compact_session_uses_mini_model_by_default():
 
     assert isinstance(message, CompactionMessage)
     assert client.params["model"] == "gpt-5.4-mini"
-    assert "instructions" not in client.params
+    assert "smallest useful compacted history" in client.params["instructions"]
+    assert "encrypted compaction summary" in client.params["instructions"]
 
 
 def test_agent_compact_session_respects_explicit_model_override():
@@ -401,6 +405,21 @@ def test_agent_compact_session_respects_explicit_instructions_override():
     assert client.params["instructions"] == "Summarize faithfully."
 
 
+def test_agent_compact_session_can_disable_default_instructions():
+    agent = Agent(session="new")
+    agent.config.compaction_instructions = None
+    visible = DeveloperMessage(content="visible compactable context")
+    done = AssistantMessage(content="done")
+    current = DeveloperMessage(content="current unfinished turn")
+    agent.session.messages = [visible, done, current]
+    client = FakeCompactionClient()
+
+    message = agent.compact_session(client=client)
+
+    assert isinstance(message, CompactionMessage)
+    assert "instructions" not in client.params
+
+
 def test_slash_compact_reports_compaction_failure(monkeypatch):
     agent = Agent(session="new")
     monkeypatch.setattr(agent, "compact_session", lambda **_kwargs: modict(status="failed", error="boom"))
@@ -472,18 +491,24 @@ def test_agent_tool_result_messages_append_directly_without_pending():
 def test_agent_get_context_starts_at_latest_compaction():
     agent = Agent(session="new")
     old = DeveloperMessage(content="old hidden context")
-    previous = CompactionMessage(output_items=[{"type": "compaction_summary", "id": "cs_123"}])
+    previous_output = [
+        {"type": "message", "role": "developer", "content": [{"type": "input_text", "text": "compacted visible anchor"}]},
+        {"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"},
+    ]
+    previous = CompactionMessage(output=previous_output)
     current = DeveloperMessage(content="current visible context")
     agent.session.messages = [old, previous, current]
 
     context = agent.context_manager.get_context()
     context_text = "\n".join(str(message.get("content", "")) for message in context)
+    payload = agent.context_manager.response_input_token_payload(context)
 
     context_ids = [message.id for message in context]
     assert previous.id in context_ids
     assert current.id in context_ids
     assert old.id not in context_ids
     assert "old hidden context" not in context_text
+    assert payload.input[:2] == previous_output
 
 
 def test_agent_local_context_estimate_does_not_count_encrypted_compaction_payload():

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_agent_plugins.py

@@ -387,6 +387,7 @@ def test_context_plugin_exposes_monitor_control_tools_and_wrapper_pruning():
     assert status.temporary_session_messages >= 3
     monitor = agent.providers.context_status_monitor()
     assert "Context status:" in monitor
+    assert "token split base/vanished_wrappers/active_wrappers/providers" in monitor
     assert "context_prune_wrappers" in monitor
 
     wrappers = agent.tools.context_list_wrappers()

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_agent_response.py

@@ -204,6 +204,7 @@ def test_agent_get_response_auto_compacts_before_backend_call(monkeypatch):
     assert calls == [{
         "model": agent.config.compaction_model,
         "max_input_tokens": agent.config.max_input_tokens,
+        "instructions": agent.config.compaction_instructions,
     }]
     assert any(isinstance(message, CompactionMessage) for message in agent.ai.params["messages"])
 

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_chat.py

@@ -285,7 +285,22 @@ def test_agent_context_status_counts_active_and_vanished_wrappers():
     assert status.active_wrapper_tokens > 0
     assert status.vanished_wrapper_tokens > 0
     assert status.base_context_tokens >= status.fixed_tokens
-    assert status.used_tokens == status.base_context_tokens + status.active_wrapper_tokens
+    assert status.provider_context_tokens == 0
+    assert status.used_tokens == status.base_context_tokens + status.provider_context_tokens + status.active_wrapper_tokens
+
+
+def test_agent_context_status_reuses_last_provider_tokens_in_visible_total():
+    agent = Agent(session="new", input_token_limit=10_000)
+    agent.context_manager.last_provider_tokens = 1_234
+    agent.session.messages = [DeveloperMessage(content="persistent context")]
+
+    status = agent.status_manager.context_status()
+
+    assert status.provider_context_tokens == 1_234
+    assert status.vanished_wrapper_tokens >= 0
+    assert status.used_tokens == status.base_context_tokens + status.provider_context_tokens + status.active_wrapper_tokens
+    assert status.used_tokens < status.base_context_tokens + status.provider_context_tokens + status.active_wrapper_tokens + status.vanished_wrapper_tokens or status.vanished_wrapper_tokens == 0
+    assert status.percent == status.used_tokens / 10_000 * 100
 
 
 def test_agent_status_preserves_persistent_and_temporary_message_counts():
@@ -557,6 +572,7 @@ def test_chat_status_bar_formats_context_and_message_wrapper_splits():
         context_limit_tokens=128_000,
         context_percent=9.609,
         base_context_tokens=10_000,
+        provider_context_tokens=3_000,
         active_wrapper_tokens=2_000,
         vanished_wrapper_tokens=4_000,
         sent_session_messages=43,
@@ -574,15 +590,15 @@ def test_chat_status_bar_formats_context_and_message_wrapper_splits():
     rendered = StatusBar.render(status)
     output = rendered.plain
 
-    assert "ctx 10.0+2.0+4.0/128.0 ktok" in output
-    assert "msg 33+7+12/93" in output
+    assert "ctx 10.0(+4.0)+2.0+3.0=15.0/128.0 ktok" in output
+    assert "msg 33(+12)+7=40/93" in output
     assert "cmp 17" in output
     assert any(
         "#f0883e" in str(span.style) and output[span.start:span.end] in {"+2.0", "+7"}
         for span in rendered.spans
     )
     assert any(
-        "#6e7681" in str(span.style) and output[span.start:span.end] in {"+4.0", "+12"}
+        "#6e7681" in str(span.style) and output[span.start:span.end] in {"(+4.0)", "(+12)"}
         for span in rendered.spans
     )
 

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_messages.py

@@ -214,18 +214,19 @@ def test_server_tool_response_message_as_string_includes_item_metadata():
     assert '"type": "web_search_call"' in text
 
 
-def test_compaction_message_replays_only_backend_summary_items():
-    output_items = [
+def test_compaction_message_replays_full_backend_output():
+    output = [
         {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "old"}]},
         {"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"},
     ]
     message = CompactionMessage(
         response_id="resp_123",
-        output_items=output_items,
+        output=output,
         compacted_message_count=4,
     )
 
-    assert message.to_response_format() == [output_items[1]]
+    assert message.to_response_format() == output
+    assert message.summaries == [output[1]]
     text = message.as_string()
     assert text.startswith("<CompactionMessage ")
     assert "response_id='resp_123'" in text
@@ -244,13 +245,14 @@ def test_compaction_message_migrates_legacy_output_items_without_persisting_them
     message = CompactionMessage({"output_items": output_items})
 
     assert "output_items" not in message
+    assert message.output == output_items
     assert message.summaries == [output_items[1]]
-    assert message.to_response_format() == [output_items[1]]
+    assert message.to_response_format() == output_items
 
 
 def test_compaction_message_prefers_output_usage_token_count():
     message = CompactionMessage(
-        output_items=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
+        output=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
         usage={"input_tokens": 42897, "output_tokens": 12000, "total_tokens": 54912},
     )
 
@@ -260,7 +262,7 @@ def test_compaction_message_prefers_output_usage_token_count():
 
 def test_compaction_message_ignores_input_usage_as_future_context_size():
     message = CompactionMessage(
-        output_items=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
+        output=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
         usage={"input_tokens": 50000, "max_input_tokens": 60000},
     )
 

{codex_agent_framework-0.1.60 → codex_agent_framework-0.1.61}/tests/test_token_counter.py

@@ -147,7 +147,7 @@ def test_token_counter_prefers_observed_usage_for_compaction_message():
     ai = FakeAI()
     counter = TokenCounter(ai, default_model="gpt-5")
     message = CompactionMessage(
-        output_items=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
+        output=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
         usage={"input_tokens": 400, "output_tokens": 120, "total_tokens": 520},
     )
 
@@ -162,7 +162,8 @@ def test_token_counter_falls_back_to_compaction_request_when_no_observed_usage()
     ai = FakeAI()
     counter = TokenCounter(ai, default_model="gpt-5")
     message = CompactionMessage(
-        output_items=[{"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
+        output=[{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "old"}]},
+                {"type": "compaction_summary", "id": "cs_123", "encrypted_content": "opaque"}],
     )
 
     count = counter.count_object(message)
@@ -170,11 +171,14 @@ def test_token_counter_falls_back_to_compaction_request_when_no_observed_usage()
     assert count == 303
     assert ai.calls == [{
         "model": "gpt-5",
-        "input": [{
-            "type": "compaction_summary",
-            "id": "cs_123",
-            "encrypted_content": "<encrypted_compaction_summary>",
-        }],
+        "input": [
+            {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "old"}]},
+            {
+                "type": "compaction_summary",
+                "id": "cs_123",
+                "encrypted_content": "<encrypted_compaction_summary>",
+            },
+        ],
     }]
     assert message.token_count["gpt-5"]["count"] == 303
 

codex_agent_framework-0.1.60/codex_agent/version.py

@@ -1 +0,0 @@
-__version__ = "0.1.60"