TalkPipe 1.0.2b2__tar.gz → 1.0.2b3__tar.gz

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/CHANGELOG.md

@@ -2,6 +2,23 @@
 
 ## Unreleased
 
+- **`max_tokens` caps the length of an LLM response.** `llmPrompt` and the
+  guided-generation segments built on it (`llmScore`, `llmExtractTerms`,
+  `llmBinaryAnswer`) accept `max_tokens`, as do the Ollama/OpenAI/Anthropic
+  prompt adapters: the most tokens the model may generate per response
+  (output only; prompt tokens do not count). The Ollama adapter's chat path
+  sent only `temperature`, so a model that never emitted a stop token
+  generated until its context window filled; `timeout` fails the pipeline
+  in that case, while `max_tokens` bounds the item. It maps to `num_predict`
+  on Ollama, `max_output_tokens` on the OpenAI Responses API, and
+  `max_tokens` on Anthropic, where it replaces the adapter's 4096 (still the
+  value when unset). The default is `None`, which leaves every request
+  exactly as it was; there is no global config key, the option is
+  per-segment. On reasoning/thinking models the thinking tokens count
+  against the cap, and a truncated response from a guided-generation segment
+  fails schema validation rather than returning a partial result. Prompt
+  adapters that take no `max_tokens` (such as `eliza`, or third-party
+  adapters that predate it) keep working and ignore it.
 - **App Center: a launch opens the page of a released application that does
   not open its own.** A second newcomer run against the PyPI releases. The
   catalog's `opens_browser` describes the newest version of an application,

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: talkpipe
-Version: 1.0.2b2
+Version: 1.0.2b3
 Summary: Python internal and external DSL for writing generative AI analytics
 Author-email: Travis Bauer <tlbauer@sandia.gov>
 Maintainer-email: Travis Bauer <tlbauer@sandia.gov>
@@ -113,12 +113,7 @@ Entry points for different usage scenarios, from interactive development to prod
 - **[chatterlang_workbench](docs/api-reference/chatterlang-workbench.md)**
   A browser-based IDE for writing, testing, and running ChatterLang scripts: editor with autocomplete and live error checking, real-time execution, pipeline save/load, next-component suggestions, logging, and documentation lookup.
 
-<video width="640" height="360" controls>
-  <source src="docs/workbench_demo.mp4" type="video/mp4">
-  Your browser does not support the video tag.
-</video>
-
-  ([workbench demo video](docs/workbench_demo.mp4) — direct link, for renderers that strip the embed above.)
+<center><img src="docs/chatterlang_workbench_screenshot.png" width=700 alt="ChatterLang Workbench: pipeline editor with autocomplete, next-component suggestions, and pipeline output"></center>
 
 - **[chatterlang_script](docs/api-reference/chatterlang-script.md)**
   Runs ChatterLang scripts from files or directly from the command line, for batch processing and automation.
@@ -148,9 +143,11 @@ Entry points for different usage scenarios, from interactive development to prod
   Each release attaches the version-stamped file, so the URL above needs a full
   release that post-dates the App Center — against an older one, or where the
   only such release so far is a pre-release, it returns "Not Found" and uv
-  reports that as a `SyntaxError`. Pre-releases *of the App Center itself* are
+  reports that as a `SyntaxError`. **That is the case today: the App Center is
+  still in beta**, and pre-releases *of the App Center itself* are
   served by the experimental channel instead
-  (`releases/download/experimental/talkpipe_appcenter.py`). Either copy
+  (`releases/download/experimental/talkpipe_appcenter.py`; see
+  [Testing the App Center while it is in beta](appcenter/README.md#testing-the-app-center-while-it-is-in-beta)). Either copy
   installs the release or the pre-release of each application — press `e` on
   its row to choose the channel and `i` to install, or pass
   `--experimental` — so you never need a particular copy to

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/README.md

@@ -44,12 +44,7 @@ Entry points for different usage scenarios, from interactive development to prod
 - **[chatterlang_workbench](docs/api-reference/chatterlang-workbench.md)**
   A browser-based IDE for writing, testing, and running ChatterLang scripts: editor with autocomplete and live error checking, real-time execution, pipeline save/load, next-component suggestions, logging, and documentation lookup.
 
-<video width="640" height="360" controls>
-  <source src="docs/workbench_demo.mp4" type="video/mp4">
-  Your browser does not support the video tag.
-</video>
-
-  ([workbench demo video](docs/workbench_demo.mp4) — direct link, for renderers that strip the embed above.)
+<center><img src="docs/chatterlang_workbench_screenshot.png" width=700 alt="ChatterLang Workbench: pipeline editor with autocomplete, next-component suggestions, and pipeline output"></center>
 
 - **[chatterlang_script](docs/api-reference/chatterlang-script.md)**
   Runs ChatterLang scripts from files or directly from the command line, for batch processing and automation.
@@ -79,9 +74,11 @@ Entry points for different usage scenarios, from interactive development to prod
   Each release attaches the version-stamped file, so the URL above needs a full
   release that post-dates the App Center — against an older one, or where the
   only such release so far is a pre-release, it returns "Not Found" and uv
-  reports that as a `SyntaxError`. Pre-releases *of the App Center itself* are
+  reports that as a `SyntaxError`. **That is the case today: the App Center is
+  still in beta**, and pre-releases *of the App Center itself* are
   served by the experimental channel instead
-  (`releases/download/experimental/talkpipe_appcenter.py`). Either copy
+  (`releases/download/experimental/talkpipe_appcenter.py`; see
+  [Testing the App Center while it is in beta](appcenter/README.md#testing-the-app-center-while-it-is-in-beta)). Either copy
   installs the release or the pre-release of each application — press `e` on
   its row to choose the channel and `i` to install, or pass
   `--experimental` — so you never need a particular copy to

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/appcenter/README.md

@@ -3,6 +3,68 @@
 > Install, launch, and manage applications from one screen, with nothing to
 > set up first but [uv](https://docs.astral.sh/uv/).
 
+## Testing the App Center while it is in beta
+
+> [!IMPORTANT]
+> **These directions are temporary.** The App Center has only shipped in
+> pre-releases so far, so the release commands in the rest of this page do
+> not work yet: `releases/latest` is the newest release GitHub does *not*
+> consider a pre-release, and against it those URLs return "Not Found" (uv
+> reports that as a `SyntaxError`, since it tries to run the 404 body as
+> Python). Once the first full release carries the App Center, this section
+> goes away and the rest of the page applies as written.
+>
+> To try it now, fetch it from the `experimental` release instead, which CI
+> keeps pointed at the assets of the newest release of either kind:
+>
+> ```bash
+> uv run https://github.com/sandialabs/talkpipe/releases/download/experimental/talkpipe_appcenter.py
+> ```
+>
+> The bootstrap scripts, for a computer without uv, take the channel by name.
+> On Linux and macOS:
+>
+> ```bash
+> curl -fsSL https://github.com/sandialabs/talkpipe/releases/download/experimental/install.sh | sh -s -- --experimental
+> ```
+>
+> On Windows, set the channel in the environment first, because a piped
+> `irm | iex` cannot be given arguments:
+>
+> ```powershell
+> $env:TALKPIPE_APPCENTER_CHANNEL="experimental"
+> powershell -ExecutionPolicy Bypass -c "irm https://github.com/sandialabs/talkpipe/releases/download/experimental/install.ps1 | iex"
+> ```
+>
+> Or run the file straight from a checkout of the repository, which needs no
+> release at all and is how to test a change before it is tagged:
+>
+> ```bash
+> uv run appcenter/talkpipe_appcenter.py
+> ```
+>
+> Three things to know while testing:
+>
+> - **The copy you run and the versions it installs are separate choices.**
+>   The beta of the App Center installs the newest *release* of each
+>   application, exactly as the released copy will; it does not put betas of
+>   the vault or the writing assistant on your computer. To test an
+>   application's pre-release too, press `e` on its row and then `i`, or pass
+>   `--experimental` (see [Channels](#channels-stable-and-experimental)).
+> - **`--version` says which beta you have.** `uv run <url> --version` prints
+>   the talkpipe tag the file was attached to (`talkpipe-appcenter 1.0.2b2`,
+>   say); a copy run from a checkout prints `0.0.0+unknown`. Include that line
+>   when reporting a problem. uv fetches the file anew on every run (only the
+>   environment is cached), so a newly published beta is picked up the next
+>   time you run the command.
+> - **The `experimental` URL rolls.** It serves whatever the newest release
+>   is, including the first full one when it comes. For a copy that stays put,
+>   use the beta's own tag:
+>   `https://github.com/sandialabs/talkpipe/releases/download/v1.0.2b2/talkpipe_appcenter.py`.
+>
+> The assets are published from github.com only; the Gitea mirror attaches
+> none.
+
 The App Center is the part of TalkPipe that installs applications. It is
 designed to make the TalkPipe-based applications (the vault, the writing
 assistant, the workbench) easy to install, upgrade, and launch, and it comes
@@ -34,13 +96,10 @@ powershell -ExecutionPolicy Bypass -c "irm https://github.com/sandialabs/talkpip
 ```
 
 All three URLs are release assets, attached by CI to each release, so they need
-a release that post-dates the App Center; against an older `latest` they return
-"Not Found" (uv reports that as a `SyntaxError`, since it tries to run the 404
-body as Python). From a checkout, run the file directly instead:
-
-```bash
-uv run appcenter/talkpipe_appcenter.py
-```
+a full release that post-dates the App Center. **There is no such release yet**;
+until there is, use the commands under
+[Testing the App Center while it is in beta](#testing-the-app-center-while-it-is-in-beta)
+at the top of this page.
 
 ## Channels: stable and experimental
 
@@ -109,7 +168,9 @@ installs from it.
 
 ### Which copy of the App Center you run
 
-This choice matters only while no full release carries the App Center.
+This choice matters only while no full release carries the App Center, which
+is the case today; the commands to use meanwhile are collected under
+[Testing the App Center while it is in beta](#testing-the-app-center-while-it-is-in-beta).
 `releases/latest` is the newest release GitHub does *not* consider a
 pre-release, so those URLs serve the last full release even while a beta is
 out — and if the only release carrying the App Center is a beta, they return

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/docs/architecture/chatterlang.md

@@ -118,6 +118,28 @@ Example:
 Adapter implementations get deterministic and truncate summarization from the shared prompt-adapter memory mixin.
 Adapters only need to implement `complete_text_without_context(...)` when they support `memory_mode="summary_llm"`.
 
+#### Capping the response length with `max_tokens`
+
+`llmPrompt` and the guided-generation segments built on it (`llmScore`,
+`llmExtractTerms`, `llmBinaryAnswer`) accept `max_tokens`: the most tokens the
+model may generate per response. It is an **output** cap — prompt tokens do not
+count — and it defaults to `None`, which leaves the backend default. Without it,
+a model that fails to emit a stop token generates until its context window
+fills; `timeout` only fails the pipeline, while `max_tokens` bounds the item.
+
+```chatterlang
+| llmScore[system_prompt=p, field="toanalyze", set_as="ai_eval", max_tokens=1024]
+```
+
+- It maps to `num_predict` on Ollama, `max_output_tokens` on OpenAI, and
+  `max_tokens` on Anthropic (where it replaces the adapter's 4096 default).
+  Adapters that generate nothing to cap (`eliza`) ignore it.
+- On reasoning/thinking models (Anthropic extended thinking, OpenAI reasoning
+  models, Ollama thinking models) the thinking tokens count against the cap, so
+  leave room for them.
+- A truncated response from a guided-generation segment such as `llmScore`
+  fails schema validation rather than returning a partial score.
+
 **Variables**: Store intermediate results
 ```chatterlang
 | @my_variable    # Store results in variable

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe/llm/chat.py

@@ -65,6 +65,19 @@ class LLMPrompt(AbstractSegment[Any, Any]):
     - unsummarized_message_count controls how many recent messages are kept verbatim during compaction.
       This is a message count, not a user+assistant turn-pair count.
     - memory_size controls the target max tokens for summaries created during compaction.
+
+    Output cap:
+    - max_tokens caps the tokens the model may generate per response. It is an
+      output cap only: prompt tokens do not count. None (the default) leaves the
+      backend default, so a model that never emits a stop token generates until
+      its context window fills.
+    - On reasoning/thinking models (Anthropic extended thinking, OpenAI reasoning
+      models, Ollama thinking models) the thinking tokens count against it, so
+      leave room for them.
+    - A response cut off by the cap is returned truncated. In a guided-generation
+      segment such as llmScore the truncated response fails schema validation
+      rather than returning a partial result.
+    - Adapters that generate nothing (eliza) or that predate the option ignore it.
     """
 
     def __init__(
@@ -115,6 +128,10 @@ class LLMPrompt(AbstractSegment[Any, Any]):
             float | None,
             "Request timeout in seconds for the model server; defaults to the llm_timeout config value (120 if unset)",
         ] = None,
+        max_tokens: Annotated[
+            int | None,
+            "Maximum tokens the model may generate per response (output only; prompt tokens do not count). None leaves the backend default.",
+        ] = None,
     ):
         super().__init__()
         logger.debug(f"Initializing LLMPrompt with name={model}, source={source}")
@@ -152,6 +169,7 @@ class LLMPrompt(AbstractSegment[Any, Any]):
             "memory_size": memory_size,
             "debug_messages": debug_messages,
             "timeout": timeout,
+            "max_tokens": max_tokens,
         }
         self.chat = self._create_prompt_adapter(
             getPromptAdapter(source), source, adapter_kwargs
@@ -178,6 +196,12 @@ class LLMPrompt(AbstractSegment[Any, Any]):
                     "remove the timeout option or update the adapter."
                 )
             adapter_kwargs = {k: v for k, v in adapter_kwargs.items() if k != "timeout"}
+        # ``max_tokens`` is dropped for such adapters even when set: eliza
+        # generates nothing to cap, so the option is harmless there.
+        if "max_tokens" in adapter_kwargs and "max_tokens" not in accepted_kwargs:
+            adapter_kwargs = {
+                k: v for k, v in adapter_kwargs.items() if k != "max_tokens"
+            }
         unsupported_compat_kwargs = (
             set(PROMPT_ADAPTER_COMPAT_KWARG_DEFAULTS) - accepted_kwargs
         )
@@ -250,6 +274,12 @@ class AbstractLLMGuidedGeneration(LLMPrompt):
     This class is used to create segments that generate output based on LLM responses.
     It is an abstract class and should not be used directly.
     Subclasses must implement the get_output_format method to specify the output format.
+
+    max_tokens caps the model's output per response (prompt tokens do not count;
+    on reasoning/thinking models the thinking tokens do). A response truncated by
+    the cap is not valid for the output format, so it fails schema validation
+    rather than returning a partial result -- size the cap for the whole
+    structured answer, explanation included.
     """
 
     @staticmethod
@@ -296,6 +326,10 @@ class AbstractLLMGuidedGeneration(LLMPrompt):
         debug_messages: Annotated[
             bool, "Whether to log outbound LLM request messages"
         ] = False,
+        max_tokens: Annotated[
+            int | None,
+            "Maximum tokens the model may generate per response (output only; prompt tokens do not count). None leaves the backend default.",
+        ] = None,
     ):
 
         super().__init__(
@@ -314,6 +348,7 @@ class AbstractLLMGuidedGeneration(LLMPrompt):
             context_token_trigger=context_token_trigger,
             memory_size=memory_size,
             debug_messages=debug_messages,
+            max_tokens=max_tokens,
         )
 
 
@@ -377,6 +412,11 @@ class LlmScore(AbstractLLMGuidedGeneration):
     - set_as (str, optional):
          If specified, the output will be appended as this field in the
          final response.
+    - max_tokens (int, optional):
+         Cap on the tokens generated per response (output only; on
+         reasoning/thinking models the thinking tokens count against it).
+         A response truncated by the cap fails schema validation rather than
+         returning a partial score, so leave room for the explanation.
 
     Data Model
     ----------

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe/llm/prompt_adapter_base.py

@@ -64,13 +64,26 @@ class AbstractLLMPromptAdapter(PromptAdapterMemoryMixin, ABC):
         debug_messages: Annotated[
             bool, "Whether to log outbound LLM request messages"
         ] = False,
+        max_tokens: Annotated[
+            int | None,
+            "Maximum tokens the model may generate per response (output only; prompt tokens do not count). None leaves the backend default.",
+        ] = None,
     ):
-        """Initialize the chat model."""
+        """Initialize the chat model.
+
+        ``max_tokens`` caps the model's output per response; prompt tokens do
+        not count. On reasoning/thinking models (Anthropic extended thinking,
+        OpenAI reasoning models, Ollama thinking models) the thinking tokens
+        count against it. A response truncated by the cap is returned as is,
+        so with an ``output_format`` it fails schema validation rather than
+        yielding a partial result.
+        """
         self._model_name = model
         self._source = source
         self._multi_turn = multi_turn
         self._temperature = temperature
         self._temperature_explicit = temperature is not None
+        self._max_tokens = max_tokens
         self._output_format = output_format
         self._messages = []
         self._summary_message = None

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe/llm/prompt_adapters_anthropic.py

@@ -28,6 +28,7 @@ class AnthropicPromptAdapter(AbstractLLMPromptAdapter):
         memory_size: int = 512,
         debug_messages: bool = False,
         timeout: float | None = None,
+        max_tokens: int | None = None,
     ):
         anthropic = self._require_dependency("anthropic", "Anthropic", "anthropic")
 
@@ -54,6 +55,7 @@ class AnthropicPromptAdapter(AbstractLLMPromptAdapter):
             context_token_trigger,
             memory_size,
             debug_messages,
+            max_tokens=max_tokens,
         )
         self._timeout = resolve_timeout(timeout, LLM_TIMEOUT, DEFAULT_LLM_TIMEOUT)
         self.client = self._build_client(
@@ -61,7 +63,9 @@ class AnthropicPromptAdapter(AbstractLLMPromptAdapter):
             "Anthropic",
             "ANTHROPIC_API_KEY",
         )
-        self._max_tokens = 4096  # Default max tokens for response
+        # The Messages API requires max_tokens, so an unset cap falls back to 4096.
+        if self._max_tokens is None:
+            self._max_tokens = 4096
 
     def execute(self, prompt: str) -> str | BaseModel:
         """Execute the chat model.

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe/llm/prompt_adapters_ollama.py

@@ -35,6 +35,7 @@ class OllamaPromptAdapter(AbstractLLMPromptAdapter):
         memory_size: int = 512,
         debug_messages: bool = False,
         timeout: float | None = None,
+        max_tokens: int | None = None,
     ):
         super().__init__(
             model,
@@ -49,6 +50,7 @@ class OllamaPromptAdapter(AbstractLLMPromptAdapter):
             context_token_trigger,
             memory_size,
             debug_messages,
+            max_tokens=max_tokens,
         )
         # Ollama uses 0.5 as default when temperature is not specified
         if self._temperature is None:
@@ -76,13 +78,16 @@ class OllamaPromptAdapter(AbstractLLMPromptAdapter):
         # assembled history and parse the reply the same way.
         logger.debug(f"Sending chat request to Ollama model {self._model_name}")
         self._log_message_payload("messages", self._request_messages())
+        options: dict[str, Any] = {"temperature": self._temperature}
+        if self._max_tokens is not None:
+            options["num_predict"] = self._max_tokens
         response = self._chat_completion(
             model=self._model_name,
             messages=self._request_messages(),
             format_schema=self._output_format.model_json_schema()
             if self._output_format
             else None,
-            options={"temperature": self._temperature},
+            options=options,
         )
 
         self._record_assistant_response(str(response.message.content))

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe/llm/prompt_adapters_openai.py

@@ -27,6 +27,7 @@ class OpenAIPromptAdapter(AbstractLLMPromptAdapter):
         memory_size: int = 512,
         debug_messages: bool = False,
         timeout: float | None = None,
+        max_tokens: int | None = None,
     ):
         openai = self._require_dependency("openai", "OpenAI", "openai")
         super().__init__(
@@ -42,6 +43,7 @@ class OpenAIPromptAdapter(AbstractLLMPromptAdapter):
             context_token_trigger,
             memory_size,
             debug_messages,
+            max_tokens=max_tokens,
         )
         self._timeout = resolve_timeout(timeout, LLM_TIMEOUT, DEFAULT_LLM_TIMEOUT)
         self.client = self._build_client(
@@ -80,6 +82,8 @@ class OpenAIPromptAdapter(AbstractLLMPromptAdapter):
         }
 
         self._apply_temperature_if_explicit(request_params)
+        if self._max_tokens is not None:
+            request_params["max_output_tokens"] = self._max_tokens
 
         self._log_message_payload("input", request_params["input"])
         response = self._responses_request(parse=True, **request_params)

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: talkpipe
-Version: 1.0.2b2
+Version: 1.0.2b3
 Summary: Python internal and external DSL for writing generative AI analytics
 Author-email: Travis Bauer <tlbauer@sandia.gov>
 Maintainer-email: Travis Bauer <tlbauer@sandia.gov>
@@ -113,12 +113,7 @@ Entry points for different usage scenarios, from interactive development to prod
 - **[chatterlang_workbench](docs/api-reference/chatterlang-workbench.md)**
   A browser-based IDE for writing, testing, and running ChatterLang scripts: editor with autocomplete and live error checking, real-time execution, pipeline save/load, next-component suggestions, logging, and documentation lookup.
 
-<video width="640" height="360" controls>
-  <source src="docs/workbench_demo.mp4" type="video/mp4">
-  Your browser does not support the video tag.
-</video>
-
-  ([workbench demo video](docs/workbench_demo.mp4) — direct link, for renderers that strip the embed above.)
+<center><img src="docs/chatterlang_workbench_screenshot.png" width=700 alt="ChatterLang Workbench: pipeline editor with autocomplete, next-component suggestions, and pipeline output"></center>
 
 - **[chatterlang_script](docs/api-reference/chatterlang-script.md)**
   Runs ChatterLang scripts from files or directly from the command line, for batch processing and automation.
@@ -148,9 +143,11 @@ Entry points for different usage scenarios, from interactive development to prod
   Each release attaches the version-stamped file, so the URL above needs a full
   release that post-dates the App Center — against an older one, or where the
   only such release so far is a pre-release, it returns "Not Found" and uv
-  reports that as a `SyntaxError`. Pre-releases *of the App Center itself* are
+  reports that as a `SyntaxError`. **That is the case today: the App Center is
+  still in beta**, and pre-releases *of the App Center itself* are
   served by the experimental channel instead
-  (`releases/download/experimental/talkpipe_appcenter.py`). Either copy
+  (`releases/download/experimental/talkpipe_appcenter.py`; see
+  [Testing the App Center while it is in beta](appcenter/README.md#testing-the-app-center-while-it-is-in-beta)). Either copy
   installs the release or the pre-release of each application — press `e` on
   its row to choose the channel and `i` to install, or pass
   `--experimental` — so you never need a particular copy to

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe.egg-info/SOURCES.txt

@@ -57,6 +57,7 @@ docs/LLMEvaluateWebPage.png
 docs/README.md
 docs/TalkPipe.png
 docs/TalkPipe.xcf
+docs/chatterlang_workbench_screenshot.png
 docs/make_structure_diagram.sh
 docs/quickstart.md
 docs/structure.tex

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/src/talkpipe.egg-info/scm_file_list.json

@@ -73,6 +73,7 @@
     "docs/architecture/protocol.md",
     "docs/architecture/security.md",
     "docs/architecture/streaming.md",
+    "docs/chatterlang_workbench_screenshot.png",
     "docs/contributing/developer-handbook.md",
     "docs/guides/container-images.md",
     "docs/guides/makevectordatabase-and-serverag.md",

talkpipe-1.0.2b3/src/talkpipe.egg-info/scm_version.json

@@ -0,0 +1,8 @@
+{
+  "tag": "1.0.2b3",
+  "distance": 0,
+  "node": "gc0c71951294cdc7328b6b67881f73f24ebea2100",
+  "dirty": false,
+  "branch": "HEAD",
+  "node_date": "2026-09-17"
+}

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/tests/talkpipe/llm/test_chat.py

@@ -145,6 +145,74 @@ def test_guided_generation_passes_context_params(monkeypatch):
     assert captured["memory_size"] == 128
 
 
+def _install_capture_adapter(monkeypatch):
+    captured = {}
+
+    class CaptureAdapter:
+        def __init__(self, **kwargs):
+            captured.update(kwargs)
+
+        def execute(self, prompt):
+            return prompt
+
+    monkeypatch.setattr("talkpipe.llm.chat.getPromptSources", lambda: ["capture"])
+    monkeypatch.setattr(
+        "talkpipe.llm.chat.getPromptAdapter", lambda _source: CaptureAdapter
+    )
+    return captured
+
+
+def test_llmprompt_max_tokens_defaults_to_none(monkeypatch):
+    captured = _install_capture_adapter(monkeypatch)
+    LLMPrompt(model="model-a", source="capture")
+    assert captured["max_tokens"] is None
+
+
+def test_llmscore_passes_max_tokens(monkeypatch):
+    captured = _install_capture_adapter(monkeypatch)
+    LlmScore(system_prompt="score it", model="model-a", source="capture", max_tokens=64)
+    assert captured["max_tokens"] == 64
+
+
+def test_chatterlang_script_passes_max_tokens(monkeypatch):
+    captured = _install_capture_adapter(monkeypatch)
+    compiler.compile(
+        'CONST p = "score it"; '
+        '| llmScore[system_prompt=p, model="model-a", source="capture", '
+        'field="toanalyze", set_as="ai_eval", max_tokens=1024]'
+    )
+    assert captured["max_tokens"] == 1024
+
+
+def test_llmprompt_drops_max_tokens_for_adapter_that_predates_it(monkeypatch):
+    captured = {}
+
+    class OldAdapter:
+        def __init__(
+            self,
+            model,
+            system_prompt=None,
+            multi_turn=True,
+            temperature=None,
+            output_format=None,
+            role_map=None,
+            memory_mode="full",
+            unsummarized_message_count=6,
+            context_token_trigger=None,
+            memory_size=512,
+            debug_messages=False,
+        ):
+            captured["model"] = model
+
+    monkeypatch.setattr("talkpipe.llm.chat.getPromptSources", lambda: ["old"])
+    monkeypatch.setattr(
+        "talkpipe.llm.chat.getPromptAdapter", lambda _source: OldAdapter
+    )
+
+    LLMPrompt(model="model-a", source="old", max_tokens=64)
+    assert captured["model"] == "model-a"
+
+
 def test_llmprompt_supports_old_adapter_defaults(monkeypatch, caplog):
     captured = {}
 

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/tests/talkpipe/llm/test_prompt_adapters_anthropic.py

@@ -148,3 +148,22 @@ def test_anthropic_no_temperature_warning_when_none_requested(monkeypatch, caplo
 
     assert "temperature" not in captured
     assert "temperature" not in caplog.text
+
+
+def test_anthropic_execute_without_max_tokens_keeps_4096(monkeypatch):
+    _patch_anthropic_constructor(monkeypatch)
+    adapter = AnthropicPromptAdapter("claude-3-5-haiku-latest")
+    captured = _capture_messages_create(monkeypatch, adapter, "ok")
+
+    assert adapter.execute("prompt") == "ok"
+    assert captured["max_tokens"] == 4096
+    assert set(captured) == {"model", "messages", "max_tokens", "system"}
+
+
+def test_anthropic_execute_max_tokens_replaces_4096(monkeypatch):
+    _patch_anthropic_constructor(monkeypatch)
+    adapter = AnthropicPromptAdapter("claude-3-5-haiku-latest", max_tokens=1024)
+    captured = _capture_messages_create(monkeypatch, adapter, "ok")
+
+    assert adapter.execute("prompt") == "ok"
+    assert captured["max_tokens"] == 1024

{talkpipe-1.0.2b2 → talkpipe-1.0.2b3}/tests/talkpipe/llm/test_prompt_adapters_eliza.py

@@ -1,3 +1,5 @@
+import logging
+
 from pydantic import BaseModel
 
 from prompt_adapter_contract_suite import (
@@ -5,6 +7,7 @@ from prompt_adapter_contract_suite import (
     run_shared_live_contract_checks,
     run_shared_offline_contract_checks,
 )
+from talkpipe.llm.chat import LLMPrompt, LlmScore
 from talkpipe.llm.config import getPromptAdapter, getPromptSources
 from talkpipe.llm.prompt_adapters import ElizaPromptAdapter
 
@@ -114,3 +117,18 @@ def test_eliza_reintroduces_i_am_on_second_name_query():
     assert "i am dr. eliza" not in name_first.lower()
     assert "my name is dr. eliza" in name_first.lower()
     assert "i am dr. eliza" in name_second.lower()
+
+
+def test_eliza_segment_ignores_max_tokens_without_warning(caplog):
+    # eliza generates nothing to cap, so max_tokens is dropped silently
+    # (unlike timeout, which eliza refuses).
+    with caplog.at_level(logging.WARNING):
+        prompt = LLMPrompt(model="eliza", source="eliza", max_tokens=16)
+        score = LlmScore(
+            system_prompt="Score 0-10.", model="eliza", source="eliza", max_tokens=16
+        )
+
+    assert isinstance(prompt.chat, ElizaPromptAdapter)
+    assert list(prompt(["hello"]))
+    assert isinstance(next(iter(score(["hello"]))), LlmScore.Score)
+    assert not [r for r in caplog.records if r.levelno >= logging.WARNING]