add deepseek compatibility

.gitignore

@@ -18,3 +18,7 @@ gaia_eval_dashboard.html
 .claude/
 .codegraph/
 .cursor/
+.pycache/
+.__pycache__/
+.ruff_cache/
+.venv/

CLAUDE.md

@@ -49,7 +49,7 @@ Untrusted-input boundary: the user's first `HumanMessage` is wrapped in `<gaia_q
 
 ### Config (`src/lilith_agent/config.py`)
 
-`Config.from_env()` reads `GAIA_*` env vars (note: `GAIA_` prefix, not `LILITH_`). Three model tiers: `cheap` / `strong` / `extra_strong`, each with independent provider+model. The agent's main model is **`extra_strong`**; `cheap` powers the tool-result summarizer. `extra_strong_*` defaults to `strong_*` if unset. The `vision_*` pair is separate, used by `inspect_visual_content`. Behavior flags: `caveman`, `caveman_mode`, `recursion_limit` (50), `budget_hard_cap` (25), `budget_warn_at` (15), `semantic_dedup_threshold` (0.5), `compact_summarize`, `llm_formatter_enabled`.
+`Config.from_env()` reads `GAIA_*` env vars (note: `GAIA_` prefix, not `LILITH_`). Three model tiers: `cheap` / `strong` / `extra_strong`, each with independent provider+model. The agent's main model defaults to **`extra_strong`**, can select `cheap` or `strong` via `GAIA_AGENT_MODEL_TIER`, and can be directly overridden with `GAIA_AGENT_PROVIDER` / `GAIA_AGENT_MODEL`; `cheap` powers the tool-result summarizer. `extra_strong_*` defaults to `strong_*` if unset. DeepSeek is available as provider `deepseek` using `DEEPSEEK_API_KEY` or `GAIA_DEEPSEEK_API_KEY`. The `vision_*` pair is separate, used by `inspect_visual_content`. Behavior flags: `caveman`, `caveman_mode`, `recursion_limit` (50), `budget_hard_cap` (25), `budget_warn_at` (15), `semantic_dedup_threshold` (0.5), `compact_summarize`, `llm_formatter_enabled`.
 
 ### Tools (`src/lilith_agent/tools/`)
 

README.md

@@ -65,6 +65,19 @@ GAIA_STRONG_PROVIDER=anthropic
 GAIA_STRONG_MODEL=claude-sonnet-4-6
 GAIA_EXTRA_STRONG_PROVIDER=anthropic
 GAIA_EXTRA_STRONG_MODEL=claude-sonnet-4-6
+GAIA_AGENT_MODEL_TIER=extra_strong  # cheap | strong | extra_strong
+# Optional one-off override for the main agent model.
+# GAIA_AGENT_PROVIDER=deepseek
+# GAIA_AGENT_MODEL=deepseek-v4-pro
+
+# DeepSeek uses an OpenAI-compatible API.
+# DEEPSEEK_API_KEY=...
+# GAIA_DEEPSEEK_BASE_URL=https://api.deepseek.com
+# GAIA_CHEAP_PROVIDER=deepseek
+# GAIA_CHEAP_MODEL=deepseek-v4-flash
+# GAIA_STRONG_PROVIDER=deepseek
+# GAIA_STRONG_MODEL=deepseek-v4-pro
+
 GAIA_VISION_PROVIDER=fal
 GAIA_VISION_MODEL=gemini-3-flash-preview
 GAIA_CAVEMAN=true

src/lilith_agent/app.py

@@ -922,7 +922,10 @@ def build_react_agent(cfg: Config):
         guidance = str(state.get("supervisor_guidance", "") or "").strip()
         original_question = _initial_question_from_state(state)
         compacted = _compact_old_tool_messages(state["messages"], summarize_fn=summarize_fn)
-        response = base_model.invoke([
+        # Thinking mode (DeepSeek v4) emits tool-call markup as raw text here, where no
+        # tools are bound to parse it; disable it so the finalizer returns plain prose.
+        finalizer_model = get_extra_strong_model(cfg, thinking=False)
+        response = finalizer_model.invoke([
             SystemMessage(content=(
                 "SUPERVISOR FINALIZER: Stop tool use. Answer the original question, not an intermediate hop. "
                 "Produce a bare final answer in 'Final Answer: ...' form using the existing evidence. "
@@ -1012,7 +1015,9 @@ def build_react_agent(cfg: Config):
             log.debug("[memory] skipping extraction: only %d messages", len(messages))
             return state
         try:
-            cheap_model = get_cheap_model(cfg)
+            # langmem/trustcall forces tool_choice for structured extraction, which
+            # DeepSeek thinking mode rejects with a 400 it then retries forever.
+            cheap_model = get_cheap_model(cfg, thinking=False)
             extract_and_compress_facts(messages, cheap_model)
         except Exception as e:
             log.warning("[memory] failed to run extraction: %s", e)

src/lilith_agent/config.py

@@ -41,6 +41,11 @@ class Config:
     tavily_api_key: str
     lmstudio_base_url: str
     max_tokens: int
+    deepseek_api_key: str = ""
+    deepseek_base_url: str = "https://api.deepseek.com"
+    agent_model_tier: str = "extra_strong"
+    agent_provider: str = ""
+    agent_model: str = ""
     caveman: bool = False
     caveman_mode: str = "full"
     recursion_limit: int = 100
@@ -54,6 +59,12 @@ class Config:
 
     @classmethod
     def from_env(cls) -> "Config":
+        agent_model_tier = os.getenv("GAIA_AGENT_MODEL_TIER", "extra_strong").strip().lower()
+        if agent_model_tier not in {"cheap", "strong", "extra_strong"}:
+            raise ValueError(
+                "GAIA_AGENT_MODEL_TIER must be one of: cheap, strong, extra_strong"
+            )
+
         return cls(
             cheap_provider=os.getenv("GAIA_CHEAP_PROVIDER", "google"),
             cheap_model=os.getenv("GAIA_CHEAP_MODEL", "gemini-3-flash-preview"),
@@ -72,6 +83,11 @@ class Config:
             huggingface_api_key=os.getenv("GAIA_HUGGINGFACE_API_KEY", ""),
             tavily_api_key=os.getenv("GAIA_TAVILY_API_KEY", ""),
             lmstudio_base_url=os.getenv("GAIA_LMSTUDIO_BASE_URL", ""),
+            deepseek_api_key=os.getenv("GAIA_DEEPSEEK_API_KEY", os.getenv("DEEPSEEK_API_KEY", "")),
+            deepseek_base_url=os.getenv("GAIA_DEEPSEEK_BASE_URL", "https://api.deepseek.com"),
+            agent_model_tier=agent_model_tier,
+            agent_provider=os.getenv("GAIA_AGENT_PROVIDER", "").strip(),
+            agent_model=os.getenv("GAIA_AGENT_MODEL", "").strip(),
             max_tokens=_get_int_env("GAIA_MAX_TOKENS", "65536"),
             caveman=os.getenv("GAIA_CAVEMAN", "true").lower() == "true",
             caveman_mode=os.getenv("GAIA_CAVEMAN_MODE", "full"),

src/lilith_agent/models.py

@@ -109,6 +109,7 @@ if SQLiteCache:
 log = logging.getLogger(__name__)
 
 LMSTUDIO_DEFAULT_BASE_URL = "http://localhost:1234/v1"
+DEEPSEEK_DEFAULT_BASE_URL = "https://api.deepseek.com"
 _NO_THINK = "/no_think"
 _GEMINI_COOLDOWN_MODELS = {"gemini-3-flash-preview", "gemini-3.1-pro"}
 _COOLDOWN_LADDER_SECONDS = (60, 120, 300)
@@ -460,7 +461,27 @@ class _BoundRetryWrapper(Runnable):
         return getattr(self._bound, name)
 
 
-def _build(provider: str, model: str, cfg: Config) -> BaseChatModel:
+def _resolve_agent_model_choice(cfg: Config) -> tuple[str, str]:
+    tier = (cfg.agent_model_tier or "extra_strong").strip().lower()
+    tiers = {
+        "cheap": (cfg.cheap_provider, cfg.cheap_model),
+        "strong": (cfg.strong_provider, cfg.strong_model),
+        "extra_strong": (cfg.extra_strong_provider, cfg.extra_strong_model),
+    }
+    if tier not in tiers:
+        raise ValueError(
+            "GAIA_AGENT_MODEL_TIER must be one of: cheap, strong, extra_strong"
+        )
+
+    provider, model = tiers[tier]
+    return (
+        (cfg.agent_provider or provider).strip(),
+        (cfg.agent_model or model).strip(),
+    )
+
+
+def _build(provider: str, model: str, cfg: Config, thinking: bool = True) -> BaseChatModel:
+    provider = provider.strip().lower()
     log.info("Building model for provider=%r, model=%r", provider, model)
     max_tokens = cfg.max_tokens
     
@@ -493,6 +514,19 @@ def _build(provider: str, model: str, cfg: Config) -> BaseChatModel:
             repo_id=model, huggingfacehub_api_token=cfg.huggingface_api_key, max_new_tokens=max_tokens
         )
         return _wrap(ChatHuggingFace(llm=endpoint))
+    if provider == "deepseek":
+        ds_kwargs = dict(
+            model=model,
+            base_url=cfg.deepseek_base_url or DEEPSEEK_DEFAULT_BASE_URL,
+            api_key=cfg.deepseek_api_key,
+            max_tokens=max_tokens,
+        )
+        # DeepSeek v4 defaults to thinking mode, which rejects forced tool_choice
+        # (langmem extraction) and leaks raw tool-call markup in free-text calls
+        # (finalizer). Opt out explicitly when the caller needs a plain response.
+        if not thinking:
+            ds_kwargs["extra_body"] = {"thinking": {"type": "disabled"}}
+        return _wrap(ChatOpenAI(**ds_kwargs))
     if provider == "lmstudio":
         # LM Studio exposes an OpenAI-compatible API.
         inner = ChatOpenAI(
@@ -509,13 +543,14 @@ def _build(provider: str, model: str, cfg: Config) -> BaseChatModel:
     raise ValueError(f"Unknown provider: {provider}")
 
 
-def get_cheap_model(cfg: Config) -> BaseChatModel:
-    return _build(cfg.cheap_provider, cfg.cheap_model, cfg)
+def get_cheap_model(cfg: Config, thinking: bool = True) -> BaseChatModel:
+    return _build(cfg.cheap_provider, cfg.cheap_model, cfg, thinking=thinking)
 
 
 def get_strong_model(cfg: Config) -> BaseChatModel:
     return _build(cfg.strong_provider, cfg.strong_model, cfg)
 
 
-def get_extra_strong_model(cfg: Config) -> BaseChatModel:
-    return _build(cfg.extra_strong_provider, cfg.extra_strong_model, cfg)
+def get_extra_strong_model(cfg: Config, thinking: bool = True) -> BaseChatModel:
+    provider, model = _resolve_agent_model_choice(cfg)
+    return _build(provider, model, cfg, thinking=thinking)

tests/test_graph.py

@@ -256,7 +256,7 @@ def test_supervisor_nudges_agent_to_answer_when_evidence_is_enough(monkeypatch,
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -313,7 +313,7 @@ def test_supervisor_uses_extra_strong_model_not_cheap_model(monkeypatch, tmp_pat
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", cheap_should_not_be_used)
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -362,7 +362,7 @@ def test_supervisor_finalizer_prompt_reinforces_original_question_contract(monke
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -413,7 +413,7 @@ def test_supervisor_finalizer_rejects_unknown_best_answer_and_forces_best_guess(
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -471,7 +471,7 @@ def test_supervisor_finalizes_even_with_placeholder_if_requested(monkeypatch, tm
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -528,7 +528,7 @@ def test_supervisor_forces_finalize_after_max_nudges(monkeypatch, tmp_path):
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MAX_NUDGES", 5, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -580,7 +580,7 @@ def test_final_answer_gets_supervisor_review_and_can_be_returned_for_revision(mo
     cfg.budget_warn_at = 99
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -638,7 +638,7 @@ def test_fail_safe_falls_back_to_supervisor_best_answer_when_empty(monkeypatch,
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -690,7 +690,7 @@ def test_fail_safe_never_returns_empty_answer_without_best_answer(monkeypatch, t
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 99, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -749,7 +749,7 @@ def test_supervisor_review_auto_approves_after_fail_safe(monkeypatch, tmp_path):
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 99, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -821,7 +821,7 @@ def test_supervisor_finalizes_when_agent_ignores_prior_nudge(monkeypatch, tmp_pa
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: object())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)
@@ -874,7 +874,7 @@ def test_supervisor_overhead_leaves_room_for_hard_cap_fail_safe(monkeypatch, tmp
     cfg.compact_summarize = False
     monkeypatch.setenv("LILITH_HOME", str(tmp_path / ".lilith"))
     monkeypatch.setattr("lilith_agent.app._SUPERVISOR_MIN_TOOL_CALLS", 1, raising=False)
-    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg: strong)
+    monkeypatch.setattr("lilith_agent.app.get_extra_strong_model", lambda cfg, thinking=True: strong)
     monkeypatch.setattr("lilith_agent.app.get_cheap_model", lambda cfg: FakeSupervisorModel())
     monkeypatch.setattr("lilith_agent.tools.build_tools", lambda cfg: [echo_tool])
     monkeypatch.setattr("lilith_agent.memory.extract_and_compress_facts", lambda messages, model: None)

tests/test_models.py

@@ -423,3 +423,132 @@ async def test_async_bound_retry_wrapper_raises_cooldown_for_gemini_lane(monkeyp
         await bound.ainvoke([("user", "hi")])
 
     assert raised.value.model == "gemini-3.1-pro"
+
+
+def test_deepseek_config_defaults_and_env(monkeypatch):
+    from lilith_agent.config import Config
+
+    monkeypatch.delenv("GAIA_DEEPSEEK_API_KEY", raising=False)
+    monkeypatch.delenv("GAIA_DEEPSEEK_BASE_URL", raising=False)
+    monkeypatch.setenv("DEEPSEEK_API_KEY", "ds-env-key")
+
+    cfg = Config.from_env()
+
+    assert cfg.deepseek_api_key == "ds-env-key"
+    assert cfg.deepseek_base_url == "https://api.deepseek.com"
+
+    monkeypatch.setenv("GAIA_DEEPSEEK_API_KEY", "gaia-ds-key")
+    monkeypatch.setenv("GAIA_DEEPSEEK_BASE_URL", "https://deepseek.internal")
+
+    cfg = Config.from_env()
+
+    assert cfg.deepseek_api_key == "gaia-ds-key"
+    assert cfg.deepseek_base_url == "https://deepseek.internal"
+
+
+def test_agent_model_tier_selects_configured_tier(monkeypatch):
+    from lilith_agent.config import Config
+    from lilith_agent.models import _resolve_agent_model_choice
+
+    monkeypatch.setenv("GAIA_CHEAP_PROVIDER", "deepseek")
+    monkeypatch.setenv("GAIA_CHEAP_MODEL", "deepseek-v4-flash")
+    monkeypatch.setenv("GAIA_STRONG_PROVIDER", "deepseek")
+    monkeypatch.setenv("GAIA_STRONG_MODEL", "deepseek-v4-pro")
+    monkeypatch.setenv("GAIA_AGENT_MODEL_TIER", "strong")
+
+    assert _resolve_agent_model_choice(Config.from_env()) == ("deepseek", "deepseek-v4-pro")
+
+
+def test_agent_model_direct_override_wins_over_tier(monkeypatch):
+    from lilith_agent.config import Config
+    from lilith_agent.models import _resolve_agent_model_choice
+
+    monkeypatch.setenv("GAIA_AGENT_MODEL_TIER", "cheap")
+    monkeypatch.setenv("GAIA_AGENT_PROVIDER", "deepseek")
+    monkeypatch.setenv("GAIA_AGENT_MODEL", "deepseek-v4-pro")
+
+    assert _resolve_agent_model_choice(Config.from_env()) == ("deepseek", "deepseek-v4-pro")
+
+
+def test_invalid_agent_model_tier_is_rejected(monkeypatch):
+    from lilith_agent.config import Config
+
+    monkeypatch.setenv("GAIA_AGENT_MODEL_TIER", "medium")
+
+    with pytest.raises(ValueError, match="GAIA_AGENT_MODEL_TIER"):
+        Config.from_env()
+
+
+def test_build_deepseek_uses_openai_compatible_endpoint(monkeypatch):
+    from dataclasses import replace
+
+    from lilith_agent.config import Config
+    from lilith_agent.models import _build
+
+    class FakeChatOpenAI:
+        def __init__(self, **kwargs):
+            self.kwargs = kwargs
+
+    monkeypatch.setattr("lilith_agent.models.ChatOpenAI", FakeChatOpenAI)
+    monkeypatch.setattr(
+        "lilith_agent.models._RetryWrapper",
+        lambda inner, provider, model_name: inner,
+    )
+
+    cfg = replace(
+        Config.from_env(),
+        deepseek_api_key="ds-key",
+        deepseek_base_url="https://deepseek.internal",
+        max_tokens=123,
+    )
+
+    model = _build("deepseek", "deepseek-v4-flash", cfg)
+
+    assert model.kwargs == {
+        "model": "deepseek-v4-flash",
+        "api_key": "ds-key",
+        "base_url": "https://deepseek.internal",
+        "max_tokens": 123,
+    }
+
+
+def _fake_deepseek_cfg(monkeypatch):
+    from dataclasses import replace
+
+    from lilith_agent.config import Config
+
+    class FakeChatOpenAI:
+        def __init__(self, **kwargs):
+            self.kwargs = kwargs
+
+    monkeypatch.setattr("lilith_agent.models.ChatOpenAI", FakeChatOpenAI)
+    monkeypatch.setattr(
+        "lilith_agent.models._RetryWrapper",
+        lambda inner, provider, model_name: inner,
+    )
+    return replace(
+        Config.from_env(),
+        deepseek_api_key="ds-key",
+        deepseek_base_url="https://deepseek.internal",
+        max_tokens=123,
+    )
+
+
+def test_build_deepseek_thinking_disabled_sets_extra_body(monkeypatch):
+    from lilith_agent.models import _build
+
+    cfg = _fake_deepseek_cfg(monkeypatch)
+
+    model = _build("deepseek", "deepseek-v4-flash", cfg, thinking=False)
+
+    assert model.kwargs["extra_body"] == {"thinking": {"type": "disabled"}}
+
+
+def test_build_deepseek_thinking_enabled_omits_extra_body(monkeypatch):
+    from lilith_agent.models import _build
+
+    cfg = _fake_deepseek_cfg(monkeypatch)
+
+    model = _build("deepseek", "deepseek-v4-flash", cfg)
+
+    assert "extra_body" not in model.kwargs