pythinker-code 0.54.0__tar.gz → 0.55.0__tar.gz

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: pythinker-code
-Version: 0.54.0
+Version: 0.55.0
 Summary: Pythinker — an agentic CLI developed by Pythoughts-labs.
 Keywords: cli,agent,ai,coding-assistant,llm,claude,openai,terminal
 Author: Pythoughts-labs
@@ -115,11 +115,11 @@ It speaks the [**Agent Client Protocol (ACP)**](https://github.com/agentclientpr
 
 ---
 
-## 🆕 What's New in 0.54.0
+## 🆕 What's New in 0.55.0
 
-- **New `Workflow` tool for deterministic multi-agent orchestration.** The default agent can author and run a script that fans work out across subagents with `agent()`, `parallel()`, and `pipeline()`, with live progress, structured-schema output, an optional token budget, and a single up-front approval.
+- **Local-model reliability fixes for compaction, Qwen3 reasoning, and stuck loops.** Capped context-compaction summary output so a slow/degenerate local model completion can't hang the soul loop, switched self-hosted `openai_legacy` endpoints (llama.cpp/vLLM/LM Studio) to Qwen3.x's binary `enable_thinking` toggle instead of a tiered `reasoning_effort` value, and added a second stuck-loop backstop that catches consecutive identical tool calls even when each reports success.
 
-Upgrade with `pythinker update`, `pip install --upgrade pythinker-code==0.54.0`, or use the native installer for your platform from the [Releases page](https://github.com/Pythoughts-labs/pythinker-code/releases/latest).
+Upgrade with `pythinker update`, `pip install --upgrade pythinker-code==0.55.0`, or use the native installer for your platform from the [Releases page](https://github.com/Pythoughts-labs/pythinker-code/releases/latest).
 
 
 ---
@@ -209,7 +209,7 @@ matches your OS — no Python, Node, or `uv` prerequisite.
 
 | Platform | Recommended install | Artifact source |
 |---|---|---|
-| **🪟 Windows** | `irm https://pythinker.com/install.ps1 \| iex` | `PythinkerSetup-0.54.0.exe` from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
+| **🪟 Windows** | `irm https://pythinker.com/install.ps1 \| iex` | `PythinkerSetup-0.55.0.exe` from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
 | **<img src="https://img.shields.io/badge/-macOS-000000?style=flat-square&logo=apple&logoColor=white" alt="macOS"> / <img src="https://img.shields.io/badge/-Linux-FCC624?style=flat-square&logo=linux&logoColor=black" alt="Linux">** | `curl -fsSL https://pythinker.com/install.sh \| bash` | native tarball from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
 | **<img src="https://img.shields.io/badge/-macOS-000000?style=flat-square&logo=apple&logoColor=white" alt="macOS"> — Homebrew** | `brew install Pythoughts-labs/pythinker/pythinker-code` | auto-published Homebrew tap |
 | **🐳 Docker** | `docker run --rm -it ghcr.io/pythoughts-labs/pythinker-code` | GHCR multi-arch image |
@@ -237,7 +237,7 @@ pythinker                      # start the interactive TUI
 
 ### 🪟 Windows — native installer
 
-`PythinkerSetup-0.54.0.exe` is a signed* Inno Setup wizard. Installs per-user
+`PythinkerSetup-0.55.0.exe` is a signed* Inno Setup wizard. Installs per-user
 into `%LOCALAPPDATA%\Programs\Pythinker`, registers `pythinker` on your user
 PATH (`HKCU\Environment`), broadcasts `WM_SETTINGCHANGE` so new shells see
 the change. **No UAC prompt.**
@@ -248,13 +248,13 @@ irm https://pythinker.com/install.ps1 | iex
 
 # Or manually download the installer + checksum from the Releases page,
 # verify with Get-FileHash, then run:
-.\PythinkerSetup-0.54.0.exe
+.\PythinkerSetup-0.55.0.exe
 
 # Open a fresh PowerShell
 pythinker --version
 ```
 
-**Per-machine install** (IT-managed boxes): `.\PythinkerSetup-0.54.0.exe /ALLUSERS`
+**Per-machine install** (IT-managed boxes): `.\PythinkerSetup-0.55.0.exe /ALLUSERS`
 installs to `%ProgramFiles%\Pythinker` and writes PATH to HKLM (requires admin).
 
 **Upgrade:** `pythinker update` from inside the running app — it downloads
@@ -312,26 +312,26 @@ attached to every GitHub Release.
 
 ```sh
 # Debian / Ubuntu (x86_64)
-sudo dpkg -i pythinker-code_0.54.0_amd64.deb
+sudo dpkg -i pythinker-code_0.55.0_amd64.deb
 sudo apt-get install -f       # only if dpkg reports missing deps
 
 # Debian / Ubuntu (ARM64)
-sudo dpkg -i pythinker-code_0.54.0_arm64.deb
+sudo dpkg -i pythinker-code_0.55.0_arm64.deb
 
 # Fedora / RHEL / openSUSE (x86_64)
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.x86_64.rpm
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.x86_64.rpm.sha256
-sha256sum -c pythinker-code-0.54.0.x86_64.rpm.sha256
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.x86_64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.x86_64.rpm.sha256
+sha256sum -c pythinker-code-0.55.0.x86_64.rpm.sha256
 # Fedora / RHEL:
-sudo dnf install ./pythinker-code-0.54.0.x86_64.rpm
+sudo dnf install ./pythinker-code-0.55.0.x86_64.rpm
 # openSUSE:
-sudo zypper install ./pythinker-code-0.54.0.x86_64.rpm
+sudo zypper install ./pythinker-code-0.55.0.x86_64.rpm
 
 # Fedora / RHEL (aarch64)
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.aarch64.rpm
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.aarch64.rpm.sha256
-sha256sum -c pythinker-code-0.54.0.aarch64.rpm.sha256
-sudo dnf install ./pythinker-code-0.54.0.aarch64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.aarch64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.aarch64.rpm.sha256
+sha256sum -c pythinker-code-0.55.0.aarch64.rpm.sha256
+sudo dnf install ./pythinker-code-0.55.0.aarch64.rpm
 ```
 
 Both packages drop a small `/usr/bin/pythinker` launcher that execs the real
@@ -340,8 +340,8 @@ binary under `/usr/lib/pythinker/`, so your `$PATH` stays tidy.
 **Verify before install:**
 
 ```sh
-sha256sum -c pythinker-code_0.54.0_amd64.deb.sha256        # Debian/Ubuntu
-sha256sum -c pythinker-code-0.54.0.x86_64.rpm.sha256       # Fedora/RHEL
+sha256sum -c pythinker-code_0.55.0_amd64.deb.sha256        # Debian/Ubuntu
+sha256sum -c pythinker-code-0.55.0.x86_64.rpm.sha256       # Fedora/RHEL
 ```
 
 **Upgrade:** download the new `.deb`/`.rpm` from Releases and `dpkg -i` /

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/README.md

@@ -50,11 +50,11 @@ It speaks the [**Agent Client Protocol (ACP)**](https://github.com/agentclientpr
 
 ---
 
-## 🆕 What's New in 0.54.0
+## 🆕 What's New in 0.55.0
 
-- **New `Workflow` tool for deterministic multi-agent orchestration.** The default agent can author and run a script that fans work out across subagents with `agent()`, `parallel()`, and `pipeline()`, with live progress, structured-schema output, an optional token budget, and a single up-front approval.
+- **Local-model reliability fixes for compaction, Qwen3 reasoning, and stuck loops.** Capped context-compaction summary output so a slow/degenerate local model completion can't hang the soul loop, switched self-hosted `openai_legacy` endpoints (llama.cpp/vLLM/LM Studio) to Qwen3.x's binary `enable_thinking` toggle instead of a tiered `reasoning_effort` value, and added a second stuck-loop backstop that catches consecutive identical tool calls even when each reports success.
 
-Upgrade with `pythinker update`, `pip install --upgrade pythinker-code==0.54.0`, or use the native installer for your platform from the [Releases page](https://github.com/Pythoughts-labs/pythinker-code/releases/latest).
+Upgrade with `pythinker update`, `pip install --upgrade pythinker-code==0.55.0`, or use the native installer for your platform from the [Releases page](https://github.com/Pythoughts-labs/pythinker-code/releases/latest).
 
 
 ---
@@ -144,7 +144,7 @@ matches your OS — no Python, Node, or `uv` prerequisite.
 
 | Platform | Recommended install | Artifact source |
 |---|---|---|
-| **🪟 Windows** | `irm https://pythinker.com/install.ps1 \| iex` | `PythinkerSetup-0.54.0.exe` from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
+| **🪟 Windows** | `irm https://pythinker.com/install.ps1 \| iex` | `PythinkerSetup-0.55.0.exe` from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
 | **<img src="https://img.shields.io/badge/-macOS-000000?style=flat-square&logo=apple&logoColor=white" alt="macOS"> / <img src="https://img.shields.io/badge/-Linux-FCC624?style=flat-square&logo=linux&logoColor=black" alt="Linux">** | `curl -fsSL https://pythinker.com/install.sh \| bash` | native tarball from [Releases](https://github.com/Pythoughts-labs/pythinker-code/releases/latest) |
 | **<img src="https://img.shields.io/badge/-macOS-000000?style=flat-square&logo=apple&logoColor=white" alt="macOS"> — Homebrew** | `brew install Pythoughts-labs/pythinker/pythinker-code` | auto-published Homebrew tap |
 | **🐳 Docker** | `docker run --rm -it ghcr.io/pythoughts-labs/pythinker-code` | GHCR multi-arch image |
@@ -172,7 +172,7 @@ pythinker                      # start the interactive TUI
 
 ### 🪟 Windows — native installer
 
-`PythinkerSetup-0.54.0.exe` is a signed* Inno Setup wizard. Installs per-user
+`PythinkerSetup-0.55.0.exe` is a signed* Inno Setup wizard. Installs per-user
 into `%LOCALAPPDATA%\Programs\Pythinker`, registers `pythinker` on your user
 PATH (`HKCU\Environment`), broadcasts `WM_SETTINGCHANGE` so new shells see
 the change. **No UAC prompt.**
@@ -183,13 +183,13 @@ irm https://pythinker.com/install.ps1 | iex
 
 # Or manually download the installer + checksum from the Releases page,
 # verify with Get-FileHash, then run:
-.\PythinkerSetup-0.54.0.exe
+.\PythinkerSetup-0.55.0.exe
 
 # Open a fresh PowerShell
 pythinker --version
 ```
 
-**Per-machine install** (IT-managed boxes): `.\PythinkerSetup-0.54.0.exe /ALLUSERS`
+**Per-machine install** (IT-managed boxes): `.\PythinkerSetup-0.55.0.exe /ALLUSERS`
 installs to `%ProgramFiles%\Pythinker` and writes PATH to HKLM (requires admin).
 
 **Upgrade:** `pythinker update` from inside the running app — it downloads
@@ -247,26 +247,26 @@ attached to every GitHub Release.
 
 ```sh
 # Debian / Ubuntu (x86_64)
-sudo dpkg -i pythinker-code_0.54.0_amd64.deb
+sudo dpkg -i pythinker-code_0.55.0_amd64.deb
 sudo apt-get install -f       # only if dpkg reports missing deps
 
 # Debian / Ubuntu (ARM64)
-sudo dpkg -i pythinker-code_0.54.0_arm64.deb
+sudo dpkg -i pythinker-code_0.55.0_arm64.deb
 
 # Fedora / RHEL / openSUSE (x86_64)
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.x86_64.rpm
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.x86_64.rpm.sha256
-sha256sum -c pythinker-code-0.54.0.x86_64.rpm.sha256
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.x86_64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.x86_64.rpm.sha256
+sha256sum -c pythinker-code-0.55.0.x86_64.rpm.sha256
 # Fedora / RHEL:
-sudo dnf install ./pythinker-code-0.54.0.x86_64.rpm
+sudo dnf install ./pythinker-code-0.55.0.x86_64.rpm
 # openSUSE:
-sudo zypper install ./pythinker-code-0.54.0.x86_64.rpm
+sudo zypper install ./pythinker-code-0.55.0.x86_64.rpm
 
 # Fedora / RHEL (aarch64)
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.aarch64.rpm
-curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.54.0/pythinker-code-0.54.0.aarch64.rpm.sha256
-sha256sum -c pythinker-code-0.54.0.aarch64.rpm.sha256
-sudo dnf install ./pythinker-code-0.54.0.aarch64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.aarch64.rpm
+curl -LO https://github.com/Pythoughts-labs/pythinker-code/releases/download/v0.55.0/pythinker-code-0.55.0.aarch64.rpm.sha256
+sha256sum -c pythinker-code-0.55.0.aarch64.rpm.sha256
+sudo dnf install ./pythinker-code-0.55.0.aarch64.rpm
 ```
 
 Both packages drop a small `/usr/bin/pythinker` launcher that execs the real
@@ -275,8 +275,8 @@ binary under `/usr/lib/pythinker/`, so your `$PATH` stays tidy.
 **Verify before install:**
 
 ```sh
-sha256sum -c pythinker-code_0.54.0_amd64.deb.sha256        # Debian/Ubuntu
-sha256sum -c pythinker-code-0.54.0.x86_64.rpm.sha256       # Fedora/RHEL
+sha256sum -c pythinker-code_0.55.0_amd64.deb.sha256        # Debian/Ubuntu
+sha256sum -c pythinker-code-0.55.0.x86_64.rpm.sha256       # Fedora/RHEL
 ```
 
 **Upgrade:** download the new `.deb`/`.rpm` from Releases and `dpkg -i` /

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "pythinker-code"
-version = "0.54.0"
+version = "0.55.0"
 description = "Pythinker — an agentic CLI developed by Pythoughts-labs."
 readme = "README.md"
 requires-python = ">=3.12"

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/src/pythinker_code/CHANGELOG.md

@@ -15,6 +15,22 @@ GitHub Releases page; `0.8.0` is the new starting line.
 
 ## Unreleased
 
+## 0.55.0 (2026-06-30)
+
+- **Local-model reliability fixes for compaction, Qwen3 reasoning, and stuck loops.**
+  Context-compaction summary output is now capped so a slow/degenerate local model
+  completion can no longer hang the soul loop indefinitely on local OpenAI-compatible
+  backends. Self-hosted `openai_legacy` endpoints (llama.cpp/vLLM/LM Studio) now send
+  Qwen3.x's binary `enable_thinking` chat-template toggle instead of a tiered
+  `reasoning_effort` value, since these models only support on/off and were silently
+  being promoted to full reasoning at every configured effort level. A second,
+  independent stuck-loop backstop (`max_consecutive_identical_calls`, default 10) now
+  stops a turn after enough consecutive tool calls with identical arguments,
+  regardless of whether each call reports success — catching loops where a tool
+  falsely reports success on a call that never made progress.
+
+Upgrade with `pythinker update`, `pip install --upgrade pythinker-code==0.55.0`, or use the native installer for your platform from the [Releases page](https://github.com/Pythoughts-labs/pythinker-code/releases/latest).
+
 ## 0.54.0 (2026-06-30)
 
 - **New `Workflow` tool for deterministic multi-agent orchestration.** The default

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/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

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/src/pythinker_code/llm.py

@@ -61,6 +61,31 @@ class LLM:
         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

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/src/pythinker_code/soul/compaction.py

@@ -9,7 +9,7 @@ from pythinker_core.message import Message
 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 @@ class SimpleCompaction:
         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],

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/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 @@ class PythinkerSoul:
                     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

{pythinker_code-0.54.0 → pythinker_code-0.55.0}/src/pythinker_code/soul/toolset.py

@@ -818,6 +818,16 @@ class PythinkerToolset:
         """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:

pythinker_code-0.55.0/src/pythinker_code/web/static/assets/architecture-7EHR7CIX-Ji0TtuXL.js

@@ -0,0 +1 @@
+import"./chunk-NNHCCRGN-DmZbXBJw.js";import{x as e}from"./mermaid-parser.core-DsHOq9qM.js";export{e as createArchitectureServices};