fix: local-model reliability — compaction hang, Qwen3 reasoning toggle, stuck-loop backstop

CHANGELOG.md

@@ -15,6 +15,19 @@ GitHub Releases page; `0.8.0` is the new starting line.
 
 ## Unreleased
 
+- Cap context-compaction summary output length so a slow/degenerate local
+  model completion (observed hanging the soul loop indefinitely on local
+  OpenAI-compatible backends) can no longer run unbounded.
+- Send Qwen3.x's binary `enable_thinking` chat-template toggle instead of a
+  tiered `reasoning_effort` value on self-hosted openai_legacy endpoints
+  (llama.cpp/vLLM/LM Studio), where the model only supports on/off and was
+  silently promoting every configured effort level to full reasoning.
+- Add a second, independent stuck-loop backstop (`max_consecutive_identical_calls`,
+  default 10) that stops a turn after enough consecutive tool calls with identical
+  arguments, regardless of whether each call reports success — the existing
+  all-error backstop can't catch a loop where a tool falsely reports success on a
+  call that never made progress.
+
 ## 0.54.0 (2026-06-30)
 
 - **New `Workflow` tool for deterministic multi-agent orchestration.** The default

packages/pythinker-core/src/pythinker_core/chat_provider/echo/scripted_echo.py

@@ -67,6 +67,13 @@ def with_thinking(self, effort: ThinkingEffort) -> Self:
         copied._scripts = deque(self._scripts)
         return copied
 
+    def with_generation_kwargs(self, **kwargs: object) -> Self:
+        # Scripted replay carries no real request body, so generation
+        # kwargs (e.g. an output-length cap) have nothing to attach to.
+        copied = copy.copy(self)
+        copied._scripts = deque(self._scripts)
+        return copied
+
 
 class ScriptedEchoStreamedMessage(StreamedMessage):
     """Streamed message for ScriptedEchoChatProvider."""

packages/pythinker-core/tests/test_scripted_echo_chat_provider.py

@@ -135,3 +135,15 @@ async def test_scripted_echo_chat_provider_requires_dsl_content():
 
     with pytest.raises(ChatProviderError):
         await provider.generate(system_prompt="", tools=[], history=[])
+
+
+async def test_scripted_echo_chat_provider_with_generation_kwargs_preserves_scripts():
+    provider = ScriptedEchoChatProvider(["text: first", "text: second"])
+
+    capped = provider.with_generation_kwargs(max_tokens=4000)
+
+    assert capped is not provider
+    first_stream = await capped.generate(system_prompt="", tools=[], history=[])
+    assert [part async for part in first_stream] == [TextPart(text="first")]
+    second_stream = await capped.generate(system_prompt="", tools=[], history=[])
+    assert [part async for part in second_stream] == [TextPart(text="second")]

src/pythinker_code/config.py

@@ -586,6 +586,11 @@ class LoopControl(BaseModel):
     call failed (a degenerate stuck loop), instead of continuing to
     ``max_steps_per_turn``. The turn ends with a ``stuck`` outcome and a handoff
     summary of what was tried. ``0`` disables the backstop. Default: 8."""
+    max_consecutive_identical_calls: int = Field(default=10, ge=0)
+    """Yield to the user after this many consecutive tool calls with identical
+    arguments, even if each call reports success. Tracked independently of
+    ``max_consecutive_failures`` so a tool that falsely reports success on a call
+    that made no progress can't defeat the backstop. ``0`` disables it. Default: 10."""
     max_truncation_recoveries: int = Field(default=3, ge=0)
     """When a model response is cut off by the output-token limit and makes no tool call,
     nudge the model to continue at most this many times per turn before surfacing the

src/pythinker_code/llm.py

@@ -61,6 +61,31 @@ def model_name(self) -> str:
         return self.chat_provider.model_name
 
 
+# Providers whose generation kwargs use `max_output_tokens` instead of the
+# more common `max_tokens` (OpenAI Chat-Completions-shaped and Anthropic
+# providers, plus the first-party `pythinker` provider, all use `max_tokens`).
+_MAX_OUTPUT_TOKENS_KWARG_OVERRIDES: dict[str, str] = {
+    "openai_responses": "max_output_tokens",
+    "openai_codex": "max_output_tokens",
+    "google_genai": "max_output_tokens",
+    "gemini": "max_output_tokens",
+    "vertexai": "max_output_tokens",
+}
+
+
+def capped_chat_provider(llm: LLM, max_output_tokens: int) -> ChatProvider:
+    """Return a copy of ``llm.chat_provider`` with its output length capped.
+
+    Callers needing a small, predictable cap regardless of the model's
+    usual output budget (e.g. a context-compaction summary) can use this
+    instead of hand-picking a provider-specific kwarg name.
+    """
+    provider_config = getattr(llm, "provider_config", None)
+    provider_type = getattr(provider_config, "type", None)
+    kwarg = _MAX_OUTPUT_TOKENS_KWARG_OVERRIDES.get(provider_type or "", "max_tokens")
+    return cast(Any, llm.chat_provider).with_generation_kwargs(**{kwarg: max_output_tokens})
+
+
 # Hosts that serve the genuine Anthropic API and therefore accept the
 # `tool_reference` / `defer_loading` beta content blocks that deferred tool
 # search depends on. The Claude API-key path and Anthropic OAuth both route
@@ -490,12 +515,22 @@ def create_llm(
         and not is_dashscope_legacy
     )
     is_glm_openai_legacy = provider.type == "openai_legacy" and _is_glm_model(model.model)
+    # Qwen3.x exposes a binary `enable_thinking` template toggle, not tiered
+    # reasoning effort; DashScope's own hosted endpoint already sends
+    # `enable_thinking` below, so only apply this for other openai_legacy
+    # routes (e.g. local llama.cpp/vLLM/LM Studio servers).
+    is_qwen3_openai_legacy = (
+        provider.type == "openai_legacy"
+        and _is_qwen3_model(model.model)
+        and not is_dashscope_legacy
+    )
     if (
         effective_effort is not None
         and supports_thinking
         and not is_kimi_openai_legacy
         and not is_glm_openai_legacy
         and not is_dashscope_legacy
+        and not is_qwen3_openai_legacy
     ):
         # Only explicitly send thinking controls for models that advertise
         # reasoning. Some OpenAI-compatible non-reasoning models reject even a
@@ -526,6 +561,13 @@ def create_llm(
             extra_body={"enable_thinking": thinking_on}
         )
 
+    # Self-hosted Qwen3.x servers (llama.cpp/vLLM/LM Studio) take the same
+    # `enable_thinking` toggle via the template-kwargs extra_body shape.
+    if is_qwen3_openai_legacy and effective_effort is not None:
+        chat_provider = cast(Any, chat_provider).with_generation_kwargs(
+            extra_body={"chat_template_kwargs": {"enable_thinking": thinking_on}}
+        )
+
     # Apply Pythinker AI-specific ``thinking.keep`` (preserved thinking) only when
     # the model is actually in thinking mode; otherwise the API would see a
     # ``thinking.keep`` without an accompanying ``thinking.type`` it honors.
@@ -663,6 +705,14 @@ def _is_glm_model(model_name: str) -> bool:
     return model_name.lower().replace("_", "-").startswith("glm-")
 
 
+def _is_qwen3_model(model_name: str) -> bool:
+    """Qwen3.x (dense and MoE) models only support the chat template's
+    binary `enable_thinking` toggle, not tiered reasoning effort levels.
+    """
+    normalized = model_name.lower().replace("_", "-")
+    return "qwen3" in normalized or "qwen-3" in normalized
+
+
 def _is_dashscope_endpoint(base_url: str) -> bool:
     """True for any Alibaba DashScope endpoint (standard, intl, workspace)."""
     from urllib.parse import urlparse

src/pythinker_code/soul/compaction.py

@@ -9,7 +9,7 @@
 from pythinker_core.tooling.empty import EmptyToolset
 
 import pythinker_code.prompts as prompts
-from pythinker_code.llm import LLM
+from pythinker_code.llm import LLM, capped_chat_provider
 from pythinker_code.soul.api_errors import is_context_overflow_error
 from pythinker_code.soul.message import system
 from pythinker_code.utils.logging import logger
@@ -87,6 +87,11 @@ def should_prune(token_count: int, max_context_size: int, *, ratio: float) -> bo
 PRUNE_PLACEHOLDER = "[tool output elided to save context: {n} chars]"
 CAP_PLACEHOLDER = "\n[tool output capped: removed {n} chars]"
 
+# A compaction summary is a short digest, not a full response — cap it well
+# below the model's normal output budget so a degenerate/repetitive
+# completion can't run unbounded and hang the soul loop.
+_COMPACTION_MAX_OUTPUT_TOKENS = 4000
+
 
 def prune_stale_tool_outputs(
     messages: Sequence[Message], *, protect_last: int, min_chars: int
@@ -270,16 +275,15 @@ async def _summarize_to_message(
         half is dropped and the request retried; ``(None, None)`` means even
         a single message did not fit.
 
-        NOTE: the summary length is bounded by the chat provider's
-        construction-time max output tokens (LLM default_max_tokens, or
-        PYTHINKER_MODEL_MAX_TOKENS). A tighter per-call cap would require a
-        max-tokens parameter on ``ChatProvider.generate`` (and
-        ``pythinker_core.step``), which neither exposes today.
+        The summary's output length is capped at ``_COMPACTION_MAX_OUTPUT_TOKENS``
+        via a per-call generation kwarg (see ``capped_chat_provider``), so a
+        slow/degenerate completion can't run unbounded.
         """
+        capped_provider = capped_chat_provider(llm, _COMPACTION_MAX_OUTPUT_TOKENS)
         while True:
             try:
                 result = await pythinker_core.step(
-                    chat_provider=llm.chat_provider,
+                    chat_provider=capped_provider,
                     system_prompt="You are a helpful assistant that compacts conversation context.",
                     toolset=EmptyToolset(),
                     history=[compact_message],

src/pythinker_code/soul/pythinkersoul.py

@@ -388,12 +388,17 @@ def _should_nudge_truncation(
 
 
 def _stuck_summary_message(
-    failures: int, tool_calls: Sequence[ToolCall], tool_results: Sequence[ToolResult]
+    count: int,
+    tool_calls: Sequence[ToolCall],
+    tool_results: Sequence[ToolResult],
+    *,
+    reason: str = "steps each had every tool call fail",
 ) -> Message:
     """Build a concise handoff message when the loop yields on a degenerate stuck loop.
 
-    Surfaces a count of consecutive all-error steps and a brief of what the last
-    step tried, so the human can take over without reconstructing state.
+    Surfaces a count (consecutive all-error steps, or consecutive identical calls,
+    per ``reason``) and a brief of what the last step tried, so the human can take
+    over without reconstructing state.
     """
     calls_by_id = {call.id: call for call in tool_calls}
     tried: list[str] = []
@@ -409,8 +414,8 @@ def _stuck_summary_message(
             brief = brief[:200] + "…"
         tried.append(f"- {name}: {brief}")
     text = (
-        f"I appear to be stuck — the last {failures} steps each had every tool call "
-        "fail, so I'm stopping and handing control back to you rather than continuing.\n\n"
+        f"I appear to be stuck — the last {count} {reason}, so I'm stopping and "
+        "handing control back to you rather than continuing.\n\n"
         "What I last tried:\n" + "\n".join(tried) + "\n\n"
         "You can adjust the request, fix the underlying issue, or tell me how to proceed."
     )
@@ -2197,6 +2202,32 @@ async def _pythinker_core_step_with_retry() -> StepResult:
                     return StepOutcome(stop_reason="stuck", assistant_message=summary)
             else:
                 self._consecutive_failures = 0
+
+            # Second, independent backstop: the same tool call repeated with
+            # identical arguments enough times in a row, regardless of whether
+            # each call reports success — catches a tool that falsely reports
+            # success on a call that never made progress, which the all-error
+            # check above can't see.
+            repeat_threshold = self._loop_control.max_consecutive_identical_calls
+            if repeat_threshold and isinstance(self._agent.toolset, PythinkerToolset):
+                repeat_count = self._agent.toolset.consecutive_repeat_count
+                if repeat_count >= repeat_threshold:
+                    from pythinker_code.telemetry import track
+
+                    summary = _stuck_summary_message(
+                        repeat_count,
+                        result.tool_calls,
+                        results,
+                        reason="tool calls were identical",
+                    )
+                    await self._context.append_message(summary)
+                    wire_send(TextPart(text=summary.extract_text(" ")))
+                    track(
+                        "agent_stuck_repeat",
+                        consecutive_repeat_calls=repeat_count,
+                        model=self._runtime.llm.model_name,
+                    )
+                    return StepOutcome(stop_reason="stuck", assistant_message=summary)
             return None
 
         # A tool-call-free message normally ends the turn. If it is only a

src/pythinker_code/soul/toolset.py

@@ -818,6 +818,16 @@ def dedup_triggered(self) -> bool:
         """Whether a cross-step duplicate was blocked in the current step."""
         return self._dedup_triggered
 
+    @property
+    def consecutive_repeat_count(self) -> int:
+        """Length of the current streak of identical-argument tool calls.
+
+        Tracked independently of each call's reported success/failure, so it
+        still catches a degenerate loop even if a tool falsely reports success
+        on a call that made no progress.
+        """
+        return self._consecutive_count
+
     def handle(self, tool_call: ToolCall) -> HandleResult:
         token = current_tool_call.set(tool_call)
         try:

tests/core/test_compaction_overflow.py

@@ -18,7 +18,7 @@
 from pythinker_core.chat_provider import APIStatusError
 from pythinker_core.message import Message
 
-from pythinker_code.llm import LLM
+from pythinker_code.llm import LLM, capped_chat_provider
 from pythinker_code.soul.compaction import SimpleCompaction
 from pythinker_code.wire.types import TextPart
 
@@ -31,8 +31,59 @@ def _history(n_pairs: int = 4) -> list[Message]:
     return messages
 
 
+class _FakeChatProvider:
+    """Minimal provider double recording `with_generation_kwargs` calls.
+
+    Returns a fresh instance rather than mutating in place, matching the real
+    providers' copy-on-write contract — so tests must assert on the returned
+    provider, not the original, to actually exercise that contract.
+    """
+
+    def __init__(self, generation_kwargs: dict[str, object] | None = None) -> None:
+        self.generation_kwargs: dict[str, object] = generation_kwargs or {}
+
+    def with_generation_kwargs(self, **kwargs: object) -> _FakeChatProvider:
+        return _FakeChatProvider(kwargs)
+
+
 def _fake_llm() -> LLM:
-    return cast(LLM, SimpleNamespace(chat_provider=None))
+    return cast(LLM, SimpleNamespace(chat_provider=_FakeChatProvider(), provider_config=None))
+
+
+def _fake_llm_with_provider_type(provider_type: str) -> LLM:
+    return cast(
+        LLM,
+        SimpleNamespace(
+            chat_provider=_FakeChatProvider(),
+            provider_config=SimpleNamespace(type=provider_type),
+        ),
+    )
+
+
+@pytest.mark.parametrize(
+    ("provider_type", "expected_kwarg"),
+    [
+        ("openai_legacy", "max_tokens"),
+        ("anthropic", "max_tokens"),
+        ("pythinker", "max_tokens"),
+        ("openai_responses", "max_output_tokens"),
+        # ChatGPT/Codex sessions build the same OpenAIResponses provider as
+        # "openai_responses" (see create_llm's "openai_codex" case), so they
+        # take the same max_output_tokens kwarg, not the max_tokens default.
+        ("openai_codex", "max_output_tokens"),
+        ("google_genai", "max_output_tokens"),
+        ("gemini", "max_output_tokens"),
+        ("vertexai", "max_output_tokens"),
+    ],
+)
+def test_capped_chat_provider_picks_kwarg_by_provider_type(
+    provider_type: str, expected_kwarg: str
+) -> None:
+    llm = _fake_llm_with_provider_type(provider_type)
+
+    capped_provider = cast(_FakeChatProvider, capped_chat_provider(llm, 4000))
+
+    assert capped_provider.generation_kwargs == {expected_kwarg: 4000}
 
 
 def _overflow_error() -> APIStatusError:
@@ -50,9 +101,11 @@ class _FakeStep:
     def __init__(self, failures_before_success: int) -> None:
         self.failures_before_success = failures_before_success
         self.histories: list[list[Message]] = []
+        self.chat_providers: list[object] = []
 
     async def __call__(self, *, chat_provider, system_prompt, toolset, history):
         self.histories.append(list(history))
+        self.chat_providers.append(chat_provider)
         if len(self.histories) <= self.failures_before_success:
             raise _overflow_error()
         return _summary_result()
@@ -89,6 +142,18 @@ async def test_exhausted_retries_fall_back_to_tail_with_note(monkeypatch) -> Non
     assert "reply 3" in joined
 
 
+@pytest.mark.asyncio
+async def test_compaction_caps_output_tokens(monkeypatch) -> None:
+    fake_step = _FakeStep(failures_before_success=0)
+    monkeypatch.setattr(pythinker_core, "step", fake_step)
+
+    llm = _fake_llm()
+    await SimpleCompaction(max_preserved_messages=2).compact(_history(), llm=llm)
+
+    used_provider = cast(_FakeChatProvider, fake_step.chat_providers[0])
+    assert used_provider.generation_kwargs == {"max_tokens": 4000}
+
+
 @pytest.mark.asyncio
 async def test_non_overflow_error_propagates(monkeypatch) -> None:
     async def _step_raises(**kwargs):

tests/core/test_config.py

@@ -50,6 +50,7 @@ def test_default_config_dump():
             "loop_control": {
                 "max_steps_per_turn": 1000,
                 "max_consecutive_failures": 8,
+                "max_consecutive_identical_calls": 10,
                 "max_truncation_recoveries": 3,
                 "max_compaction_failures": 1,
                 "max_session_cost_usd": None,

tests/core/test_model_switch_carryover.py

@@ -23,8 +23,15 @@
 from pythinker_code.wire.types import TextPart
 
 
+class _FakeChatProvider:
+    """Minimal provider double supporting `with_generation_kwargs`."""
+
+    def with_generation_kwargs(self, **kwargs: object) -> _FakeChatProvider:
+        return self
+
+
 def _fake_llm() -> LLM:
-    return cast(LLM, SimpleNamespace(chat_provider=None))
+    return cast(LLM, SimpleNamespace(chat_provider=_FakeChatProvider(), provider_config=None))
 
 
 def _history(n_pairs: int = 3) -> list[Message]: