Skip to content

Your lightweight open-source personal agent.

test(agent): guard workspace context prompt prefixes

Xubin Renda02daaa1d1c
diff --git a/docs/configuration.md b/docs/configuration.md
index 97fdc0363..8263cce5d 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -2088,6 +2088,10 @@ context. This uses the same policy resolver as the tools, including restrictions
a configured exec sandbox. It does not grant access or change OS permissions.
The context is stored with that request, not rewritten in earlier messages. New requests
refresh it after a project switch or context compaction. It is hidden from the chat view.
+This adds a small amount of input text per user request. Tool continuations and transient
+retries reuse the stored context; they do not add another copy. Access-mode changes do not
+rewrite the system prompt, tool definitions, or previous messages. Changing the project or
+compacting history can change the prompt prefix. Cache hits still depend on the provider.
To diagnose a path error, the agent can call `my(action="check", key="workspace_sandbox")`.
This read-only result shows the current tool directory and workspace restriction, including
diff --git a/tests/agent/tools/test_runtime_control.py b/tests/agent/tools/test_runtime_control.py
index fd88b5724..b89dc58ab 100644
--- a/tests/agent/tools/test_runtime_control.py
+++ b/tests/agent/tools/test_runtime_control.py
@@ -23,6 +23,7 @@ from nanobot.bus.events import InboundMessage
from nanobot.bus.queue import MessageBus
from nanobot.config.schema import ToolsConfig
from nanobot.providers.base import GenerationSettings, LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.openai_compat_provider import OpenAICompatProvider
from nanobot.runtime_context import RUNTIME_CONTEXT_END, RUNTIME_CONTEXT_TAG
from nanobot.security.workspace_access import (
bind_workspace_scope,
@@ -60,6 +61,26 @@ def _my_tool(loop: AgentLoop) -> MyTool:
return tool
[email protected](params=["chat", "responses"])
+def wire_request(request):
+ """Use real provider conversion without contacting a model service."""
+ provider = OpenAICompatProvider(api_key="test-key")
+ build = provider._build_kwargs if request.param == "chat" else provider._build_responses_body
+
+ def convert(call):
+ body = build(
+ messages=call["messages"], tools=call["tools"], model="gpt-4o",
+ max_tokens=100, temperature=0.7, reasoning_effort=None,
+ tool_choice=call.get("tool_choice"),
+ )
+ # Keep the protocol's real item shape, but use one name for prefix comparisons.
+ if "input" in body:
+ body["messages"] = body.pop("input")
+ return body
+
+ return convert
+
+
def test_agent_loop_assembles_my_tool_with_runtime_control(tmp_path: Path) -> None:
loop = _make_loop(tmp_path)
tool = _my_tool(loop)
@@ -349,7 +370,7 @@ async def test_workspace_inspection_is_isolated_between_concurrent_turns(tmp_pat
@pytest.mark.asyncio
-async def test_workspace_inspection_reaches_model_from_selected_chat(tmp_path: Path) -> None:
+async def test_workspace_inspection_reaches_model_from_selected_chat(tmp_path: Path, wire_request) -> None:
project = tmp_path / "project"
project.mkdir()
(project / "marker.txt").write_text("selected project", encoding="utf-8")
@@ -392,6 +413,11 @@ async def test_workspace_inspection_reaches_model_from_selected_chat(tmp_path: P
assert "'restrict_to_workspace': True" in results["inspect"]
assert "selected project" in results["read"]
assert calls[0].kwargs["tools"] == calls[1].kwargs["tools"]
+ first_wire, continuation = [wire_request(call.kwargs) for call in calls]
+ assert json.dumps(
+ continuation["messages"][:len(first_wire["messages"])], sort_keys=True,
+ ) == json.dumps(first_wire["messages"], sort_keys=True)
+ assert continuation["tools"] == first_wire["tools"]
@pytest.mark.asyncio
@@ -444,7 +470,7 @@ async def test_workspace_context_does_not_require_my_tool(tmp_path: Path) -> Non
@pytest.mark.asyncio
-async def test_access_changes_refresh_only_new_message_and_preserve_prefix(tmp_path: Path) -> None:
+async def test_access_changes_refresh_only_new_message_and_preserve_prefix(tmp_path: Path, wire_request) -> None:
project = tmp_path / "project"
project.mkdir()
other = tmp_path / "other"
@@ -468,6 +494,13 @@ async def test_access_changes_refresh_only_new_message_and_preserve_prefix(tmp_p
assert json.dumps(str(other.resolve())) in third[-1]["content"]
assert json.dumps(str(project.resolve())) not in third[-1]["content"]
assert all(call.kwargs["tools"] == calls[0].kwargs["tools"] for call in calls)
+ before, after = [wire_request(call.kwargs) for call in calls[:2]]
+ assert json.dumps(
+ after["messages"][:len(before["messages"])], sort_keys=True,
+ ) == json.dumps(before["messages"], sort_keys=True)
+ assert {key: value for key, value in before.items() if key != "messages"} == {
+ key: value for key, value in after.items() if key != "messages"
+ }
assert all(message["content"] == "Check access" for message in session.get_history(
include_runtime_context=False,
) if message["role"] == "user")
@@ -504,34 +537,97 @@ async def test_injected_channel_input_reports_active_tool_scope(tmp_path: Path)
@pytest.mark.asyncio
[email protected]("trigger", ["idle", "new"])
-async def test_workspace_context_compaction_reuses_serialized_prefix(tmp_path: Path, trigger: str) -> None:
[email protected]("trigger", ["idle", "new", "pressure"])
+async def test_workspace_context_compaction_reuses_serialized_prefix(
+ tmp_path: Path, trigger: str, wire_request,
+) -> None:
loop = _make_loop(tmp_path)
provider = loop.provider
- provider.generation = GenerationSettings()
+ provider.generation = GenerationSettings(max_tokens=100)
provider.estimate_prompt_tokens.return_value = (100, "test")
- provider.chat_stream_with_retry = AsyncMock(return_value=LLMResponse(content="done"))
+ provider.can_resume_conversation_state.return_value = False
+ provider.chat_stream_with_retry = AsyncMock(side_effect=[
+ LLMResponse(content=None, tool_calls=[ToolCallRequest(
+ id="inspect", name="my", arguments={"action": "check", "key": "workspace_sandbox"},
+ )]),
+ LLMResponse(content="done"),
+ LLMResponse(content="Access checked."),
+ LLMResponse(content="done"),
+ ])
+ loop.schedule_background = lambda coro: coro.close()
await loop.process_direct("Check access", session_key="cli:context")
first = provider.chat_stream_with_retry.await_args.kwargs
session = loop.sessions.get_or_create("cli:context")
if trigger == "idle":
await loop.consolidator.compact_idle_session(session.key, runtime=loop.llm_runtime())
- else:
+ elif trigger == "new":
await loop.consolidator.archive_session(
session, archive_end=len(session.messages), runtime=loop.llm_runtime(),
)
- compact = provider.chat_stream_with_retry.await_args.kwargs
+ else:
+ loop.set_runtime_context_window(1_624)
+ loop.consolidator._SAFETY_BUFFER = 0
+
+ def estimate(messages, _tools, _model):
+ if "SNIP" in str(messages[-1].get("content")):
+ return 300, "test"
+ if any(message.get("role") == "tool" for message in messages):
+ return 600, "test"
+ return 100, "test"
+
+ provider.estimate_prompt_tokens.side_effect = estimate
+ await loop.process_direct("Check again", session_key=session.key)
+ compact = provider.chat_stream_with_retry.await_args_list[2].kwargs
first_wire = LLMProvider._sanitize_empty_content(first["messages"])
compact_wire = LLMProvider._sanitize_empty_content(compact["messages"])
assert len(compact_wire) > len(first_wire)
- assert json.dumps(compact_wire[:len(first_wire)]) == json.dumps(first_wire)
+ assert compact_wire[:len(first_wire)] == first_wire
assert compact["tools"] == first["tools"]
- await loop.process_direct("Check again", session_key=session.key)
+ before, after = wire_request(first), wire_request(compact)
+ # JSON object field order is not model input; preserve every value and array order.
+ assert json.dumps(
+ after["messages"][:len(before["messages"])], sort_keys=True,
+ ) == json.dumps(before["messages"], sort_keys=True)
+ assert after["tools"] == before["tools"]
+ assert after.get("instructions") == before.get("instructions")
+ if trigger != "pressure":
+ await loop.process_direct("Check again", session_key=session.key)
current = provider.chat_stream_with_retry.await_args.kwargs["messages"][-1]["content"]
assert current.count("Current project (JSON path):") == 1
assert "File access: not restricted to this project." in current
+async def test_workspace_context_transient_retry_keeps_exact_wire_request(
+ tmp_path: Path, monkeypatch: pytest.MonkeyPatch, wire_request,
+) -> None:
+ loop = _make_loop(tmp_path)
+ provider = OpenAICompatProvider(api_key="test-key")
+ requests = []
+
+ async def respond(**kwargs):
+ requests.append(wire_request(kwargs))
+ if len(requests) == 1:
+ return LLMResponse(
+ content="Temporarily unavailable", finish_reason="error",
+ error_status_code=503, error_should_retry=True,
+ )
+ return LLMResponse(content="done")
+
+ # Keep the production retry policy; replace only transport and backoff waiting.
+ monkeypatch.setattr(provider, "chat_stream_with_context", respond)
+ monkeypatch.setattr(provider, "_sleep_with_heartbeat", AsyncMock())
+ loop.provider.chat_stream_with_retry = provider.chat_stream_with_retry
+ loop.provider.generation = GenerationSettings()
+ result = await loop.process_direct("Check access", session_key="cli:retry")
+
+ assert result.content == "done"
+ assert len(requests) == 2
+ assert json.dumps(requests[0]) == json.dumps(requests[1])
+ assert json.dumps(requests[0]).count("Current project (JSON path):") == 1
+ provider._sleep_with_heartbeat.assert_awaited_once()
+
+
@pytest.fixture
def runtime() -> LLMRuntime:
return LLMRuntime(MagicMock(), "test", GenerationSettings(), 128_000)