Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

Studio: keep checkpoint compaction under --disable-tools (#12119)

Daniel Han9290d18ad3da
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index 02e4f9a800e..815cf0bb1e9 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -203,13 +203,11 @@ def _can_reset_epoch(
# carried-forward header tells the model to search. Same refusal, one scope down.
return False
try:
- from state.tool_policy import get_tool_policy
+ from state.tool_policy import conversation_recall_allowed
- # `supports_tools` is the TEMPLATE's capability, not "will this request be given
- # the tool". `--disable-tools` refuses every tool for the life of the process, so
- # resetting would strand the epoch behind a tool that never arrives. Rolling keeps
- # the newest turns in view and re-injects the inline recall, needing no tool.
- if get_tool_policy() is False:
+ # `supports_tools` is the TEMPLATE's capability; where recall is refused a reset would
+ # strand the epoch behind a tool that never arrives, and rolling needs no tool.
+ if not conversation_recall_allowed():
return False
except Exception: # noqa: BLE001 -- an unreadable policy is "no", never an error
return False
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index a9de136bca0..0514be2f75e 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -5916,8 +5916,11 @@ async def _select_request_tools(
if getattr(payload, "deep_research_armed", None):
# Appended past every filter above, including the tools_on gate: arming research in the
# composer is what offers this tool, and it is the only way the model can start a run.
+ # Not under `--disable-tools`, where the loop can open only for search_conversation.
from core.inference.tools import DEEP_RESEARCH_TOOL
- tools = tools + [DEEP_RESEARCH_TOOL]
+ from state.tool_policy import get_tool_policy
+ if get_tool_policy() is not False:
+ tools = tools + [DEEP_RESEARCH_TOOL]
return tools
@@ -5984,10 +5987,10 @@ def _checkpoint_needs_search(payload = None) -> bool:
def _checkpoint_recall_may_enable_tools(payload) -> bool:
- """Whether checkpoint recovery can force the internal recall tool loop."""
- from state.tool_policy import get_tool_policy
+ """Whether checkpoint recovery can force the internal recall tool loop (also under --disable-tools)."""
+ from state.tool_policy import conversation_recall_allowed
return bool(
- get_tool_policy() is not False
+ conversation_recall_allowed()
and _rolling_context_policy(payload) is not None
and _checkpoint_needs_search(payload)
and _thread_has_conversation_archive(getattr(payload, "thread_id", None))
diff --git a/studio/backend/routes/preview.py b/studio/backend/routes/preview.py
index 719756e696e..7153ba2e438 100644
--- a/studio/backend/routes/preview.py
+++ b/studio/backend/routes/preview.py
@@ -128,6 +128,8 @@ def _sanitize_preview_payload(
"bypass_permissions": False,
"confirm_tool_calls": False,
"session_id": None,
+ # A streamed body outlives tools_force_disabled(), so no thread archive may be reachable.
+ "thread_id": None,
"rag_scope": None,
"openai_code_exec_container_id": None,
"anthropic_code_exec_container_id": None,
diff --git a/studio/backend/state/tool_policy.py b/studio/backend/state/tool_policy.py
index 0938a11f71e..18f475b9949 100644
--- a/studio/backend/state/tool_policy.py
+++ b/studio/backend/state/tool_policy.py
@@ -100,6 +100,11 @@ def get_tool_policy_default() -> Optional[bool]:
return _tool_policy_default
+def conversation_recall_allowed() -> bool:
+ """Recall reads only this thread's archive: `--disable-tools` keeps it, `tools_force_disabled` does not."""
+ return not _force_disabled.get()
+
+
@contextmanager
def tools_force_disabled() -> Iterator[None]:
"""Hard-disable server-side tools for the current async context."""
diff --git a/studio/backend/tests/data/refactor_guard/patch_targets.json b/studio/backend/tests/data/refactor_guard/patch_targets.json
index 954bec4d13e..93a3eb5d847 100644
--- a/studio/backend/tests/data/refactor_guard/patch_targets.json
+++ b/studio/backend/tests/data/refactor_guard/patch_targets.json
@@ -938,8 +938,6 @@
"tests/test_media_auto_switch.py"
],
"state.tool_policy.get_tool_policy": [
- "tests/test_checkpoint_compaction.py",
- "tests/test_checkpoint_compaction.py",
"tests/test_checkpoint_compaction.py"
],
"storage.studio_db.add_scan_folder_with_status": [
diff --git a/studio/backend/tests/test_checkpoint_compaction.py b/studio/backend/tests/test_checkpoint_compaction.py
index d70bf141773..04528b70d28 100644
--- a/studio/backend/tests/test_checkpoint_compaction.py
+++ b/studio/backend/tests/test_checkpoint_compaction.py
@@ -537,23 +537,70 @@ def test_a_restated_instruction_does_not_crowd_out_every_other_rule():
assert any(item.startswith("Second standing rule") for item in items)
-def test_a_process_with_tools_disabled_never_resets(monkeypatch):
- """`supports_tools` is the TEMPLATE's capability, not "this request gets the tool".
-
- `--disable-tools` sets the process policy to False, so every tool is refused and the
- checkpoint override is blocked. Resetting anyway would leave the epoch behind a tool
- that never arrives while the header tells the model to search for what was dropped.
- """
+def test_a_process_with_tools_disabled_still_resets(monkeypatch):
+ """`--disable-tools` keeps checkpoint resets; a per-context hard-off still refuses them."""
from core.inference import llama_cpp
+ from state.tool_policy import tools_force_disabled
monkeypatch.setattr("core.rag.conversation_archive.enabled", lambda: True)
monkeypatch.setattr("core.rag.conversation_archive.can_archive", lambda thread_id: True)
- monkeypatch.setattr("state.tool_policy.get_tool_policy", lambda: None)
+ monkeypatch.setattr("state.tool_policy._tool_policy", None)
assert llama_cpp._can_reset_epoch("thread-1", True) is True
- monkeypatch.setattr("state.tool_policy.get_tool_policy", lambda: False)
- assert llama_cpp._can_reset_epoch("thread-1", True) is False
+ monkeypatch.setattr("state.tool_policy._tool_policy", False)
+ assert llama_cpp._can_reset_epoch("thread-1", True) is True
+
+ with tools_force_disabled():
+ assert llama_cpp._can_reset_epoch("thread-1", True) is False
+
+
+def test_disable_tools_reopens_the_loop_for_recall_only(monkeypatch):
+ """The recall loop under `--disable-tools` offers search_conversation alone."""
+ import asyncio
+ import types
+
+ import routes.inference as routes_mod
+ from state.tool_policy import tools_force_disabled
+
+ monkeypatch.setattr("state.tool_policy._tool_policy", False)
+ monkeypatch.setattr(routes_mod, "_thread_has_conversation_archive", lambda _tid: True)
+ monkeypatch.setattr(routes_mod, "_thread_has_checkpoint", lambda *_a: True)
+ monkeypatch.setattr(routes_mod, "_enabled_agent_skills", lambda: [{"name": "hf-cli"}])
+ monkeypatch.setattr("core.inference.checkpoint.enabled", lambda: True)
+ monkeypatch.delenv("UNSLOTH_CONTEXT_OVERFLOW", raising = False)
+
+ payload = types.SimpleNamespace(
+ enabled_tools = ["web_search", "read_skill"],
+ rag_scope = None,
+ thread_id = "t1",
+ messages = [],
+ bypass_permissions = False,
+ context_overflow = "truncate_oldest",
+ context_policy = None,
+ deep_research_armed = True,
+ )
+ assert routes_mod._checkpoint_recall_may_enable_tools(payload) is True
+ with tools_force_disabled():
+ assert routes_mod._checkpoint_recall_may_enable_tools(payload) is False
+
+ tools = asyncio.run(
+ routes_mod._select_request_tools(
+ payload, tools_on = False, mcp_allowed = False, checkpoint_fitted = True
+ )
+ )
+ assert [tool["function"]["name"] for tool in tools] == ["search_conversation"]
+
+ monkeypatch.setattr("state.tool_policy._tool_policy", None)
+ tools = asyncio.run(
+ routes_mod._select_request_tools(
+ payload, tools_on = False, mcp_allowed = False, checkpoint_fitted = True
+ )
+ )
+ assert [tool["function"]["name"] for tool in tools] == [
+ "search_conversation",
+ "deep_research",
+ ]
def _memory_tool_branch():
diff --git a/studio/backend/tests/test_preview_routes.py b/studio/backend/tests/test_preview_routes.py
index 77774ab91a0..44c9114ff8c 100644
--- a/studio/backend/tests/test_preview_routes.py
+++ b/studio/backend/tests/test_preview_routes.py
@@ -290,6 +290,7 @@ def test_chat_payload_sanitized(client, captured):
"use_adapter": False,
"confirm_tool_calls": True,
"session_id": "abc",
+ "thread_id": "someone-elses-thread",
"rag_scope": {"project_id": "x"},
"enable_thinking": True,
"reasoning_effort": "high",
@@ -308,6 +309,7 @@ def test_chat_payload_sanitized(client, captured):
# Tool-loop levers neutralized regardless of the tool gate.
assert p.confirm_tool_calls is False
assert p.session_id is None
+ assert p.thread_id is None
assert p.rag_scope is None
# Provider routing stripped so /p can't proxy an arbitrary endpoint.
assert p.provider_id is None