From 3aa144a377a37269e476f1123f1c7887668201c1 Mon Sep 17 00:00:00 2001
From: yangfucheng123 <140314128+yangfucheng123@users.noreply.github.com>
Date: Mon, 19 Oct 2026 15:45:56 +0800
Subject: [PATCH] fix(managed-models): map DeepSeek "minimal" effort to "low"
 (#6717)

The official DeepSeek Chat Completions endpoint rejects the generic
"minimal" effort token (issue #4514). The declared effort contract for
DeepSeek models now advertises the provider's real vocabulary
(low/medium/high/max/xhigh), defaults to "low", and aliases the generic
"minimal" preset onto "low" so the "flash" dialogue mode serializes a
value the endpoint accepts.

Co-authored-by: Willem Jiang <willem.jiang@gmail.com>
---
 .../config/managed_model_providers.py         |  2 +-
 .../harness/deerflow/models/AGENTS.md         |  2 +-
 backend/tests/test_managed_deepseek.py        | 30 +++++++++++++++++++
 3 files changed, 32 insertions(+), 2 deletions(-)

diff --git a/backend/packages/harness/deerflow/config/managed_model_providers.py b/backend/packages/harness/deerflow/config/managed_model_providers.py
index a69e143ccb..14ffa42a8b 100644
--- a/backend/packages/harness/deerflow/config/managed_model_providers.py
+++ b/backend/packages/harness/deerflow/config/managed_model_providers.py
@@ -95,7 +95,7 @@ def _capabilities(provider: str, model: str) -> dict[str, Any]:
             when_thinking_enabled={"extra_body": {"thinking": {"type": "enabled"}}},
             when_thinking_disabled={"extra_body": {"thinking": {"type": "disabled"}}},
         )
-        result["reasoning"]["effort"] = {"values": ["high", "max"], "default": "high", "aliases": {"minimal": "high", "low": "high", "medium": "high"}}
+        result["reasoning"]["effort"] = {"values": ["low", "medium", "high", "max", "xhigh"], "default": "low", "aliases": {"minimal": "low"}}
     elif provider in {"zai", "zhipu-cn"} and model in {"glm-4.5", "glm-4.6", "glm-4.7", "glm-5.3-flash"}:
         result.update(
             reasoning={"thinking": "optional", "dialect": "openai_extra_body", "history": "preserve"},
diff --git a/backend/packages/harness/deerflow/models/AGENTS.md b/backend/packages/harness/deerflow/models/AGENTS.md
index 731a6e53f5..f9a861d1c0 100644
--- a/backend/packages/harness/deerflow/models/AGENTS.md
+++ b/backend/packages/harness/deerflow/models/AGENTS.md
@@ -218,7 +218,7 @@ Direct `AppConfig.from_file()` continues to read only operator configuration.
 `config/managed_model_providers.py` is the sole registry for provider IDs, labels,
 endpoints, protocols, authentication parameter mappings, adapters and request
 defaults. Capability rules match exact verified model IDs; unknown preset IDs
-have an unsupported-thinking contract, no vision and no inferred context window.
+have an unsupported-thinking contract, no vision and no inferred context window. Verified DeepSeek presets declare an effort vocabulary of `low`/`medium`/`high`/`max`/`xhigh` (default `low`) and alias the generic `minimal` preset onto `low`, so a saved per-agent `minimal` default serializes as a value the official endpoint accepts (issue #4514).
 Preset drafts cannot supply endpoints or adapter paths. Legacy custom
 `openai-compatible` records retain endpoint-based official DeepSeek detection;
 never infer that adapter from a model-name substring. Reads do not migrate the
diff --git a/backend/tests/test_managed_deepseek.py b/backend/tests/test_managed_deepseek.py
index 0c5129db9a..936669b5de 100644
--- a/backend/tests/test_managed_deepseek.py
+++ b/backend/tests/test_managed_deepseek.py
@@ -212,6 +212,36 @@ async def test_saved_model_streaming_replays_reasoning_and_honors_token_limit(st
     assert model_config.supports_reasoning_effort is True
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "effort,expected",
+    [("minimal", "low"), ("low", "low"), ("medium", "medium"), ("high", "high"), ("max", "max"), ("xhigh", "xhigh")],
+)
+async def test_deepseek_effort_contract_maps_generic_effort_values(store, provider, effort, expected):
+    saved = store.save(ManagedModel(provider="deepseek", model="deepseek-flash", api_key=_KEY), expected_revision=None)
+    config = merge_managed_models(_base_config())
+    model_config = config.get_model_config(saved.name)
+
+    # The official endpoint rejects the generic "minimal" token (issue #4514),
+    # so the declared contract advertises the provider's real vocabulary and
+    # aliases "minimal" onto "low". Parametrized over every accepted value so
+    # a future rename/drop fails here, not at request time.
+    assert model_config.reasoning is not None
+    assert model_config.reasoning.effort is not None
+    assert model_config.reasoning.effort.values == ["low", "medium", "high", "max", "xhigh"]
+    assert model_config.reasoning.effort.default == "low"
+    assert model_config.reasoning.effort.aliases == {"minimal": "low"}
+    assert model_config.supports_reasoning_effort is True
+
+    model = create_chat_model(saved.name, thinking_enabled=False, reasoning_effort=effort, app_config=config, attach_tracing=False)
+    await _collect(model, [HumanMessage(content="Hello")])
+
+    assert len(provider.requests) == 1
+    body = provider.requests[0][1]
+    assert body["reasoning_effort"] == expected
+    assert body["thinking"] == {"type": "disabled"}
+
+
 @pytest.mark.parametrize(
     "base_url",
     [