steerable-agent-runtime 0.6.0__tar.gz → 0.6.2__tar.gz

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: steerable-agent-runtime
-Version: 0.6.0
+Version: 0.6.2
 Summary: Steerable agent runtime: LLM, tool, storage, and transport adapters.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "steerable-agent-runtime"
-version = "0.6.0"
+version = "0.6.2"
 description = "Steerable agent runtime: LLM, tool, storage, and transport adapters."
 readme = "README.md"
 requires-python = ">=3.10"

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/src/steerable_agent_runtime/llm/compat.py

@@ -57,6 +57,18 @@ _FLAG_WIRE_SPEC: tuple[tuple[str, str, str, str], ...] = (
         "string-list",
         "Delta keys read as reasoning text, in preference order",
     ),
+    (
+        "reasoningEchoField",
+        "reasoning_echo_field",
+        "string",
+        "Message key for echoing plaintext reasoning back to the vendor",
+    ),
+    (
+        "supportsForcedToolChoice",
+        "supports_forced_tool_choice",
+        "bool",
+        "Accept tool_choice=required; disable to downgrade it to auto",
+    ),
     (
         "cachedTokensFields",
         "cached_tokens_fields",
@@ -109,11 +121,23 @@ class OpenAICompatFlags:
     #: Whether ``temperature`` may be sent. Reasoning-only model families
     #: (o-series class) reject it.
     supports_temperature: bool = True
+    #: Whether ``tool_choice: "required"`` is accepted. Thinking-mode
+    #: gateways (DeepSeek live-verified 2026-09-08: "Thinking mode does not
+    #: support this tool_choice") 400 the forced value while plain ``auto``
+    #: works — flagging off downgrades ``required`` → ``auto`` on the wire.
+    supports_forced_tool_choice: bool = True
 
     # ── response shape ───────────────────────────────────────────────
     #: Delta keys read as reasoning text, in preference order. DeepSeek uses
     #: ``reasoning_content``; OpenRouter's GLM path uses ``reasoning``.
     reasoning_delta_fields: tuple[str, ...] = ("reasoning_content", "reasoning")
+    #: Message key for echoing plaintext reasoning back on the next request.
+    #: OpenRouter normalizes to ``reasoning`` (the reference default);
+    #: DeepSeek thinking mode *requires* the round-trip under its own
+    #: ``reasoning_content`` key — echoing as ``reasoning`` is an HTTP 400
+    #: ("The `reasoning_content` in the thinking mode must be passed back to
+    #: the API.", live-verified 2026-09-08 with deepseek-v4-flash).
+    reasoning_echo_field: str = "reasoning"
     #: Usage locations read for cached prompt tokens, in preference order.
     #: Dotted paths resolve nested objects. OpenAI nests under
     #: ``prompt_tokens_details.cached_tokens``; DeepSeek reports top-level
@@ -142,6 +166,8 @@ class OpenAICompatFlags:
             value = data[wire]
             if field in ("reasoning_delta_fields", "cached_tokens_fields"):
                 kwargs[field] = tuple(str(v) for v in value)
+            elif field == "reasoning_echo_field":
+                kwargs[field] = str(value)
             elif field == "max_tokens_field":
                 if value not in _MAX_TOKENS_FIELDS:
                     raise ValueError(
@@ -162,13 +188,22 @@ class OpenAICompatFlags:
 #: as data — adding a vendor must not touch request-building or parsing.
 PROVIDER_COMPAT_HOSTS: list[tuple[str, OpenAICompatFlags]] = [
     # DeepSeek: reasoning arrives as ``reasoning_content`` and cache hits as
-    # top-level ``prompt_cache_hit_tokens``. The tolerant defaults already
-    # cover both — this entry pins that coverage as data so a future default
-    # change cannot silently regress it.
+    # top-level ``prompt_cache_hit_tokens``. Thinking mode additionally
+    # REQUIRES the reasoning round-trip under ``reasoning_content`` — the
+    # reference echo key ``reasoning`` is ignored and the follow-up request
+    # 400s (live-verified 2026-09-08, deepseek-v4-flash tool-turn: "The
+    # `reasoning_content` in the thinking mode must be passed back to the
+    # API."). The tolerant delta-field defaults already cover parsing; the
+    # echo key is the divergence pinned here.
     (
         "api.deepseek.com",
         OpenAICompatFlags(
             reasoning_delta_fields=("reasoning_content", "reasoning"),
+            reasoning_echo_field="reasoning_content",
+            # Thinking mode 400s ``tool_choice="required"`` ("Thinking mode
+            # does not support this tool_choice", live-verified 2026-09-08
+            # with deepseek-v4-flash); ``auto`` and omission both work.
+            supports_forced_tool_choice=False,
             cached_tokens_fields=(
                 "prompt_cache_hit_tokens",
                 "prompt_tokens_details.cached_tokens",

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/src/steerable_agent_runtime/llm/openai_compat.py

@@ -312,7 +312,7 @@ class OpenAICompatProvider:
         compat = self.compat or OpenAICompatFlags()
         body: dict[str, Any] = {
             "model": self.model,
-            "messages": [_encode_message(m) for m in messages],
+            "messages": [_encode_message(m, compat=compat) for m in messages],
             "stream": stream,
         }
         if stream and compat.supports_usage_in_streaming:
@@ -331,11 +331,16 @@ class OpenAICompatProvider:
             if tools_list:
                 body["tools"] = tools_list
         body.update(extra)
-        # Z.AI (direct or OpenRouter pin) 400s ``tool_choice=required``.
-        # Harbor still logs the hook; the wire must send auto or the trial
-        # dies on round 0 (failed-prev 33335200327).
-        if body.get("tool_choice") == "required" and _z_ai_tool_choice_auto_only(
-            self.model, self.base_url
+        # ``tool_choice=required`` downgrade paths:
+        #  - compat flag: vendors whose thinking mode 400s the forced value
+        #    (DeepSeek: "Thinking mode does not support this tool_choice");
+        #  - Z.AI (direct or OpenRouter pin) 400s it outright — model-name
+        #    based, which host flags can't cover. Harbor still logs the hook;
+        #    the wire must send auto or the trial dies on round 0
+        #    (failed-prev 33335200327).
+        if body.get("tool_choice") == "required" and (
+            not compat.supports_forced_tool_choice
+            or _z_ai_tool_choice_auto_only(self.model, self.base_url)
         ):
             body["tool_choice"] = "auto"
         # W6-8: clamp the env-requested reasoning effort to a level the model
@@ -476,7 +481,10 @@ def _encode_content(message: LLMMessage) -> str | list[dict[str, Any]]:
     return out
 
 
-def _encode_message(message: LLMMessage) -> dict[str, Any]:
+def _encode_message(
+    message: LLMMessage, *, compat: OpenAICompatFlags | None = None
+) -> dict[str, Any]:
+    flags = compat or OpenAICompatFlags()
     out: dict[str, Any] = {"role": message.role, "content": _encode_content(message)}
     if message.name is not None:
         out["name"] = message.name
@@ -495,11 +503,13 @@ def _encode_message(message: LLMMessage) -> dict[str, Any]:
             for tc in message.tool_calls
         ]
     # OpenRouter: echo reasoning_details unmodified so GLM continues after
-    # tools. Prefer the structured block; plaintext is the fallback.
+    # tools. Prefer the structured block; plaintext is the fallback. The
+    # plaintext echo key is compat data: DeepSeek thinking mode 400s unless
+    # the round-trip uses its own ``reasoning_content`` field.
     if message.reasoning_details:
         out["reasoning_details"] = message.reasoning_details
     elif message.reasoning:
-        out["reasoning"] = message.reasoning
+        out[flags.reasoning_echo_field] = message.reasoning
     return out
 
 

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/src/steerable_agent_runtime.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: steerable-agent-runtime
-Version: 0.6.0
+Version: 0.6.2
 Summary: Steerable agent runtime: LLM, tool, storage, and transport adapters.
 Requires-Python: >=3.10
 Description-Content-Type: text/markdown

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/tests/test_llm_wire_helpers.py

@@ -154,6 +154,34 @@ def test_openai_parse_and_encode_reasoning_details() -> None:
     assert "reasoning_details" not in plaintext
 
 
+def test_openai_encode_reasoning_echo_field_from_compat() -> None:
+    """DeepSeek thinking mode 400s unless reasoning round-trips under its own
+    ``reasoning_content`` key (live-verified 2026-09-08, deepseek-v4-flash);
+    the echo field name is compat data, not a hardcoded constant."""
+    from steerable_agent_runtime.llm import OpenAICompatFlags, compat_for_base_url
+
+    msg = LLMMessage.text_of("assistant", "", reasoning="think step")
+
+    deepseek = compat_for_base_url("https://api.deepseek.com")
+    assert deepseek is not None
+    encoded = _encode_message(msg, compat=deepseek)
+    assert encoded["reasoning_content"] == "think step"
+    assert "reasoning" not in encoded
+
+    # Reference default (and OpenRouter) keep the ``reasoning`` echo key.
+    default = _encode_message(msg)
+    assert default["reasoning"] == "think step"
+    assert "reasoning_content" not in default
+
+    openrouter = compat_for_base_url("https://openrouter.ai/api/v1")
+    assert openrouter is not None
+    assert _encode_message(msg, compat=openrouter)["reasoning"] == "think step"
+
+    # Host-passed camelCase payload builds the same flags.
+    from_dict = OpenAICompatFlags.from_dict({"reasoningEchoField": "reasoning_content"})
+    assert _encode_message(msg, compat=from_dict)["reasoning_content"] == "think step"
+
+
 def test_openai_parse_stream_chunk_usage_only() -> None:
     chunk = {
         "choices": [],

{steerable_agent_runtime-0.6.0 → steerable_agent_runtime-0.6.2}/tests/test_provider_compat.py

@@ -96,6 +96,45 @@ def test_temperature_unsupported_suppresses_field() -> None:
     assert "temperature" not in body
 
 
+def test_forced_tool_choice_downgraded_when_flagged_off() -> None:
+    """DeepSeek thinking mode 400s ``tool_choice="required"`` ("Thinking mode
+    does not support this tool_choice", live-verified 2026-09-08 with
+    deepseek-v4-flash); the flag downgrades it to ``auto`` on the wire."""
+    from steerable_agent_runtime.llm import LLMMessage
+
+    provider = _provider(compat=OpenAICompatFlags(supports_forced_tool_choice=False))
+    body = provider._build_body(
+        messages=[LLMMessage.text_of("user", "hi")],
+        tools=None,
+        temperature=None,
+        max_tokens=None,
+        stream=True,
+        extra={"tool_choice": "required"},
+    )
+    assert body["tool_choice"] == "auto"
+
+
+def test_forced_tool_choice_kept_by_default() -> None:
+    from steerable_agent_runtime.llm import LLMMessage
+
+    provider = _provider()
+    body = provider._build_body(
+        messages=[LLMMessage.text_of("user", "hi")],
+        tools=None,
+        temperature=None,
+        max_tokens=None,
+        stream=True,
+        extra={"tool_choice": "required"},
+    )
+    assert body["tool_choice"] == "required"
+
+
+def test_registry_deepseek_disables_forced_tool_choice() -> None:
+    entry = compat_for_base_url("https://api.deepseek.com")
+    assert entry is not None
+    assert entry.supports_forced_tool_choice is False
+
+
 # ---------------------------------------------------------------------------
 # Response-shape flags
 # ---------------------------------------------------------------------------