ms-tau-sdk 2.0.27.dev80__tar.gz → 2.0.28.dev82__tar.gz

{ms_tau_sdk-2.0.27.dev80 → ms_tau_sdk-2.0.28.dev82}/CHANGELOG.md

@@ -2,6 +2,28 @@
 
 ## Unreleased
 
+- **Models that read images receive them on OpenAI-compatible, Google and Mistral providers
+  ([#121](https://github.com/mainsequence-sdk/ms-tau-sdk/issues/121)).** The runtime built these
+  providers without telling Tau which models read images, and `tau-ai==0.4.2` then assumes that
+  none does: every image in a user message or a tool result, such as the image content of an MCP
+  tool result or Tau's `read` of an image file, was replaced by a note telling the model that the
+  image was unavailable. A model whose platform projection lists `image` in its input now receives
+  the image: an `image_url` part on Chat Completions (`openai-completions`), an `input_image` part
+  on Responses (`openai-responses`), and the image parts of the Google and Mistral APIs.
+  Organization custom providers and providers built from the platform's catalog definition, such
+  as Meta, follow the same rule. A text-only model still receives the note in place of the image.
+  The `tau-ai` pin and the ADR 0016 provider patch are unchanged.
+
+- **Models that read images receive them on Anthropic Messages and OpenAI Codex providers
+  ([#123](https://github.com/mainsequence-sdk/ms-tau-sdk/issues/123)).** These two transports had
+  the same gap: every image reached the model as the note saying it was unavailable. A model whose
+  platform projection lists `image` in its input now receives an `image` block on Anthropic
+  Messages, which serves Anthropic's Claude models and the other catalog models Tau routes to it,
+  such as GitHub Copilot's Claude models, and an `input_image` part on OpenAI Codex. A text-only
+  model still receives the note.
+
+## 2.0.27 — 2026-10-28
+
 - **Platform providers that this Tau release does not know run from the platform's catalog.**
   Beside each built-in provider credential, the platform now sends `catalog_definition`: the
   address and Tau compatibility settings its provider catalog publishes for the selected model.

{ms_tau_sdk-2.0.27.dev80 → ms_tau_sdk-2.0.28.dev82}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: ms-tau-sdk
-Version: 2.0.27.dev80
+Version: 2.0.28.dev82
 Summary: Workspace-bound Tau application primitives for Main Sequence projects
 Project-URL: Changelog, https://github.com/mainsequence-sdk/ms-tau-sdk/blob/development/CHANGELOG.md
 Project-URL: Documentation, https://github.com/mainsequence-sdk/ms-tau-sdk/tree/development/docs

{ms_tau_sdk-2.0.27.dev80 → ms_tau_sdk-2.0.28.dev82}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "ms-tau-sdk"
-version = "2.0.27.dev80"
+version = "2.0.28.dev82"
 description = "Workspace-bound Tau application primitives for Main Sequence projects"
 readme = "README.md"
 requires-python = ">=3.13"

{ms_tau_sdk-2.0.27.dev80 → ms_tau_sdk-2.0.28.dev82}/src/ms_tau_sdk/providers/factory.py

@@ -443,6 +443,8 @@ class ProviderFactory:
             ):
                 mapped_thinking = model_metadata.thinking_level_map[normalized_thinking]
         api = provider_control.model.api
+        # Tau puts a placeholder in place of every image unless the model reads images.
+        supports_images = "image" in provider_control.model.input
         if outside_tau_catalog:
             if api not in CUSTOM_API_TRANSPORTS:
                 raise ConfigurationError(
@@ -520,6 +522,7 @@ class ProviderFactory:
                     headers=headers,
                     timeout_seconds=timeout_seconds,
                     reasoning_effort=mapped_thinking,
+                    supports_images=supports_images,
                     provider_name=credential.provider,
                 )
             )
@@ -533,6 +536,7 @@ class ProviderFactory:
                     provider_name=credential.provider,
                     timeout_seconds=timeout_seconds,
                     max_tokens=max_tokens,
+                    supports_images=supports_images,
                     thinking_budget_tokens=(
                         anthropic_thinking_budget_for_level(normalized_thinking)
                         if normalized_thinking is not None and mapped_thinking is not None
@@ -549,6 +553,7 @@ class ProviderFactory:
             provider_name=credential.provider,
             omit_authorization_header=is_custom_provider and not bool(secret),
             max_tokens=max_tokens,
+            supports_images=supports_images,
             timeout_seconds=timeout_seconds,
             reasoning_effort=mapped_thinking,
             reasoning_effort_parameter=(