Fix Qwen session tool calls and expose thinking compatibility

2026-09-21 16:23:51 +02:00
parent efd7d788d9
commit 84084bba37
10 changed files with 220 additions and 4 deletions
@@ -128,6 +128,51 @@ endpoint expects a model name different from the catalog key.
 Provider integrations remain declarative. Do not register providers from
 `harness/.pi/extensions/`; those extensions implement the workflow and human gates only.

+### Qwen 3.6 sessions and thinking controls
+
+For Qwen served through a vLLM-compatible chat template, declare the following inside the
+model's `session` block, alongside its context and output limits:
+
+```yaml
+reasoning: true
+compatibility:
+  supportsDeveloperRole: false
+  supportsReasoningEffort: false
+  supportsStore: false
+  maxTokensField: max_tokens
+  thinkingFormat: qwen-chat-template
+```
+
+This makes Pi send `chat_template_kwargs.enable_thinking` from the selected thinking level,
+with `preserve_thinking: true`. Choose **off** to explicitly disable thinking. The alternative
+`thinkingFormat: qwen` is for endpoints expecting top-level `enable_thinking`. Both formats
+require `reasoning: true`; declaring `reasoning: false` does not tell the server to disable
+thinking. Omit `thinkingFormat` to preserve Pi's default behavior for other providers.
+
+Regenerate projections with the updated host CLI and recreate the local core container after
+rebuilding it. Do not add these fields directly to generated Pi files. These controls do not
+force tool calls or certify the workflow; perform the operator verification above.
+
+```sh
+tht --installation /absolute/path/thothii-installation.yaml installation generate
+```
+
+Use the model identifier exposed by your endpoint, such as `qwen3.6-35b-a3b`, and limits
+supported by that deployment. The thinking format configures the Pi session adapter;
+metadata generation continues to use its separate LiteLLM settings.
+
+If a session displays text such as `{"type":"bash","command":"tht session show … --json"}`
+and never opens a review widget, that text is not an executed tool call. A verified cause
+was the Evidence JSON extension being loaded into interactive sessions and forcing
+`response_format: {type: "json_object"}`. Upgrade to the core image containing the fix:
+the extension belongs in `.pi/evidence-extensions/` and is loaded explicitly only by
+Evidence authoring. It must not also remain in the automatically loaded `.pi/extensions/`
+directory. Regenerating model configuration alone does not remove an extension from an old image.
+
+After upgrading, reload the browser and resume the session. Verify that Pi executes
+`tht session show` and opens a review widget. This fix does not require changing the Qwen
+server, forcing every turn to call a tool, or teaching the model to print tool-call JSON.
+
 ## Authentication

 Every provider chooses one explicit mode: