Skip to content

DeepTutor: Lifelong Personalized Tutoring. https://deeptutor.info/.

fix(cli): respect configured token budget in doctor probe

OrdoAbChao757dc901c8ae2
diff --git a/deeptutor/services/doctor.py b/deeptutor/services/doctor.py
index 335b50033..133d84e8b 100644
--- a/deeptutor/services/doctor.py
+++ b/deeptutor/services/doctor.py
@@ -336,8 +336,21 @@ def _redact_error(exc: Exception, config: Any) -> str:
async def _probe_provider(config: Any) -> None:
+ from deeptutor.services.config.loader import get_agent_params
from deeptutor.services.llm import complete
+ # Share the Settings probe's configurable budget. A 64-token cap can be
+ # exhausted by hidden reasoning before a healthy model emits even "OK".
+ # Doctor must also work before agents.yaml has been created.
+ try:
+ probe_params = get_agent_params("llm_probe")
+ except FileNotFoundError:
+ probe_params = {}
+ try:
+ max_tokens = max(1, int(probe_params.get("max_tokens", 4096)))
+ except (TypeError, ValueError):
+ max_tokens = 4096
+
response = await complete(
model=str(config.model),
prompt="Reply with OK.",
@@ -351,7 +364,7 @@ async def _probe_provider(config: Any) -> None:
reasoning_effort=config.reasoning_effort,
max_retries=0,
allow_image_fallback=False,
- max_tokens=64,
+ max_tokens=max_tokens,
)
if not (response or "").strip():
raise RuntimeError("The model returned an empty response.")
diff --git a/tests/cli/test_doctor_cli.py b/tests/cli/test_doctor_cli.py
index 030d366f3..dabea65b9 100644
--- a/tests/cli/test_doctor_cli.py
+++ b/tests/cli/test_doctor_cli.py
@@ -203,7 +203,13 @@ async def test_online_probe_success_is_reported(tmp_path) -> None:
@pytest.mark.asyncio
-async def test_online_probe_passes_one_bounded_token_parameter(monkeypatch, tmp_path) -> None:
+ ("configured_budget", "expected"),
+ [(None, 4096), (512, 512), (8192, 8192), ("2048", 2048), ("invalid", 4096), (0, 1)],
+)
+async def test_online_probe_uses_configured_budget(
+ monkeypatch, tmp_path, configured_budget, expected
+) -> None:
import deeptutor.services.llm as llm
captured = {}
@@ -213,6 +219,13 @@ async def test_online_probe_passes_one_bounded_token_parameter(monkeypatch, tmp_
return "OK"
monkeypatch.setattr(llm, "complete", fake_complete)
+ (tmp_path / "agents.yaml").write_text(
+ json.dumps({"diagnostics": {"llm_probe": {"max_tokens": configured_budget}}}),
+ encoding="utf-8",
+ )
+ monkeypatch.setattr(
+ "deeptutor.services.config.loader.get_runtime_settings_dir", lambda _root: tmp_path
+ )
report = await run_diagnostics(
online=True,
@@ -227,10 +240,36 @@ async def test_online_probe_passes_one_bounded_token_parameter(monkeypatch, tmp_
)
assert report.ok is True
- assert captured["max_tokens"] == 64
+ assert captured["max_tokens"] == expected
assert "max_completion_tokens" not in captured
[email protected](("response", "expected_status"), [("OK", "pass"), (" ", "fail")])
+async def test_online_probe_allows_reasoning_before_visible_output(
+ monkeypatch, tmp_path, response, expected_status
+) -> None:
+ """A healthy reasoning endpoint must not fail because the probe truncates it."""
+
+ async def fake_complete(**kwargs):
+ return response if kwargs["max_tokens"] >= 4096 else ""
+
+ def missing_settings(module):
+ raise FileNotFoundError("agents.yaml")
+
+ monkeypatch.setattr("deeptutor.services.llm.complete", fake_complete)
+ monkeypatch.setattr("deeptutor.services.config.loader.get_agent_params", missing_settings)
+ report = await run_diagnostics(
+ online=True,
+ resolve_llm=lambda: _llm_config(model="gpt-5-mini"),
+ data_root=tmp_path,
+ load_rag_config=lambda: {"defaults": {}, "knowledge_bases": {}},
+ )
+
+ assert next(check for check in report.checks if check.key == "online").status == expected_status
+ assert report.ok is (expected_status == "pass")
+
+
@pytest.mark.asyncio
async def test_online_probe_preserves_empty_custom_endpoint_api_key(monkeypatch, tmp_path) -> None:
import deeptutor.services.llm as llm