release: promote develop → master (v0.2.0)

.github/workflows/branch-policy.yml

@@ -2,9 +2,7 @@ name: Enforce branch flow
 on:
   pull_request:
     branches:
-      - stable
-      - main
-      - 'main/**'
+      - master
 
 jobs:
   check-branch-policy:
@@ -17,19 +15,9 @@ jobs:
 
           echo "PR: $SOURCE → $TARGET"
 
-          # stable ← main, main/*, hotfix/*
-          # main   ← develop, hotfix/*
-          # main/* ← develop, hotfix/*
-          if [[ "$TARGET" == "stable" ]]; then
-            PATTERN="^(main(/.*)?|hotfix/.+)$"
-            ALLOWED="main, main/*, hotfix/*"
-          elif [[ "$TARGET" == "main" || "$TARGET" == main/* ]]; then
-            PATTERN="^(develop|hotfix/.+)$"
-            ALLOWED="develop, hotfix/*"
-          else
-            echo "✅ No branch policy for target '$TARGET'"
-            exit 0
-          fi
+          # master ← develop, hotfix/*
+          PATTERN="^(develop|hotfix/.+)$"
+          ALLOWED="develop, hotfix/*"
 
           if [[ "$SOURCE" =~ $PATTERN ]]; then
             echo "✅ '$SOURCE' → '$TARGET' is allowed"

.github/workflows/bump-version-on-pr-to-stable.yml → .github/workflows/bump-version-on-pr-to-master.yml

@@ -1,14 +1,15 @@
-name: Bump version on PR to stable
+name: Bump version on PR to master
 
 on:
   pull_request:
     types: [opened, synchronize]
-    branches: [stable]
+    branches: [master]
 
 jobs:
   bump-version:
-    # Only run for main/* branches
-    if: startsWith(github.head_ref, 'main/')
+    # Only run for the develop → master release train.
+    # Hotfix branches bump their version manually before opening the PR.
+    if: github.head_ref == 'develop'
     runs-on: ubuntu-latest
     permissions:
       contents: write
@@ -23,20 +24,21 @@ jobs:
       - name: Read versions and compare
         id: compare
         run: |
-          # Read version from source branch (main/*)
+          # Read version from source branch (develop)
           SOURCE_VERSION=$(grep -m1 '^version' pyproject.toml | sed 's/.*"\(.*\)"/\1/')
           echo "source=$SOURCE_VERSION" >> "$GITHUB_OUTPUT"
 
-          # Read version from stable
-          git fetch origin stable
-          STABLE_VERSION=$(git show origin/stable:pyproject.toml | grep -m1 '^version' | sed 's/.*"\(.*\)"/\1/')
-          echo "stable=$STABLE_VERSION" >> "$GITHUB_OUTPUT"
+          # Read version from master
+          git fetch origin master
+          MASTER_VERSION=$(git show origin/master:pyproject.toml | grep -m1 '^version' | sed 's/.*"\(.*\)"/\1/')
+          echo "master=$MASTER_VERSION" >> "$GITHUB_OUTPUT"
 
-          echo "Source: $SOURCE_VERSION | Stable: $STABLE_VERSION"
+          echo "Source: $SOURCE_VERSION | Master: $MASTER_VERSION"
 
-          # Compare using sort -V (version sort)
-          HIGHER=$(printf '%s\n%s' "$SOURCE_VERSION" "$STABLE_VERSION" | sort -V | tail -1)
-          if [ "$SOURCE_VERSION" != "$STABLE_VERSION" ] && [ "$SOURCE_VERSION" = "$HIGHER" ]; then
+          # Compare using sort -V (version sort). If develop is already ahead
+          # (e.g. a manual minor/major bump), no patch bump is needed.
+          HIGHER=$(printf '%s\n%s' "$SOURCE_VERSION" "$MASTER_VERSION" | sort -V | tail -1)
+          if [ "$SOURCE_VERSION" != "$MASTER_VERSION" ] && [ "$SOURCE_VERSION" = "$HIGHER" ]; then
             echo "needs_bump=false" >> "$GITHUB_OUTPUT"
             echo "Source is already ahead, no bump needed"
           else
@@ -48,7 +50,7 @@ jobs:
         if: steps.compare.outputs.needs_bump == 'true'
         id: next
         run: |
-          IFS='.' read -r MAJOR MINOR PATCH <<< "${{ steps.compare.outputs.stable }}"
+          IFS='.' read -r MAJOR MINOR PATCH <<< "${{ steps.compare.outputs.master }}"
           PATCH=$((PATCH + 1))
           echo "version=$MAJOR.$MINOR.$PATCH" >> "$GITHUB_OUTPUT"
           echo "Next version: $MAJOR.$MINOR.$PATCH"

.github/workflows/pr-lint-and-test.yml

@@ -5,7 +5,7 @@ name: PR Lint & Test
 
 on:
   pull_request:
-    branches: [develop, stable, main, 'main/**']
+    branches: [develop, master]
   workflow_dispatch:
 
 jobs:

.github/workflows/release-on-merge-to-stable.yml → .github/workflows/release-on-merge-to-master.yml

@@ -1,9 +1,9 @@
-name: Release on merge to stable
+name: Release on merge to master
 
 on:
   pull_request:
     types: [closed]
-    branches: [stable]
+    branches: [master]
 
 jobs:
   release:
@@ -13,10 +13,10 @@ jobs:
       contents: write
 
     steps:
-      - name: Checkout stable
+      - name: Checkout master
         uses: actions/checkout@v4
         with:
-          ref: stable
+          ref: master
           fetch-depth: 0
 
       - name: Read version from pyproject.toml
@@ -37,4 +37,4 @@ jobs:
           fi
           git tag "$TAG"
           git push origin "$TAG"
-          gh release create "$TAG" --title "$TAG" --generate-notes --target stable
+          gh release create "$TAG" --title "$TAG" --generate-notes --target master

dana/common/llm/providers/anthropic.py

@@ -350,6 +350,22 @@ async def chat(self, messages: list[LLMMessage], tools: list[dict] | None = None
 
         except anthropic.APITimeoutError as e:
             raise LLMTimeoutError(f"Anthropic API timeout: {e}") from e
+        except anthropic.BadRequestError as e:
+            # Map Anthropic prompt-too-long to typed PromptTooLongError so the
+            # caller-layer (llm_caller.py) can trigger reactive_compact + retry.
+            from dana.common.llm.types import PromptTooLongError
+
+            err_body: dict = {}
+            try:
+                err_body = (e.response.json() or {}).get("error", {}) if hasattr(e, "response") else {}
+            except Exception:
+                err_body = {}
+            err_type = err_body.get("type")
+            err_msg = err_body.get("message", "") or str(e)
+            if err_type == "invalid_request_error" and "prompt is too long" in err_msg.lower():
+                raise PromptTooLongError(f"Anthropic: {err_msg}") from e
+            logger.error("Anthropic API error", error=str(e))
+            raise
         except Exception as e:
             logger.error("Anthropic API error", error=str(e))
             raise

dana/common/llm/providers/azure.py

@@ -4,7 +4,7 @@
 import structlog
 
 from ...config import config_manager
-from .openai_compatible_base import OpenAICompatibleProvider
+from .openai_compatible_base import OpenAICompatibleProvider, make_logging_http_client
 
 
 logger = structlog.get_logger()
@@ -13,6 +13,34 @@
 class AzureProvider(OpenAICompatibleProvider):
     """Azure OpenAI provider."""
 
+    # Azure exposes the Responses API only on api-version >= this date.
+    # Older versions return HTTP 400 BadRequest for /openai/responses.
+    _RESPONSES_API_MIN_DATE = "2025-03-01"
+
+    # Env var operators use to dial reasoning effort for Azure deployments.
+    # Valid values: "minimal" | "low" | "medium" | "high".
+    _REASONING_EFFORT_ENV_VAR = "AZURE_THINKING_EFFORT"
+
+    # Env var to force/disable the Responses API for this provider.
+    # Truthy: 1/true/on/yes — falsy: 0/false/off/no. Overrides the config flag
+    # and the api-version gate (forcing on an old api-version will likely 400).
+    _RESPONSES_API_ENV_VAR = "AZURE_USE_RESPONSES_API"
+
+    @property
+    def name(self) -> str:
+        return "azure"
+
+    def _endpoint_url(self) -> str:
+        return getattr(self, "azure_endpoint", "") or ""
+
+    def _responses_api_supported(self) -> bool:
+        # api-version format is "YYYY-MM-DD" or "YYYY-MM-DD-preview"; first 10 chars
+        # are the ISO date which sorts correctly lexicographically.
+        version = getattr(self, "api_version", None)
+        if not version or len(version) < 10:
+            return False
+        return version[:10] >= self._RESPONSES_API_MIN_DATE
+
     def __init__(
         self, api_key: str | None = None, model: str = "gpt-35-turbo", base_url: str | None = None, api_version: str | None = None
     ):
@@ -38,6 +66,7 @@ def __init__(
             raise ValueError("Azure OpenAI endpoint URL not found. Set AZURE_OPENAI_API_URL environment variable.")
 
         azure_endpoint = azure_endpoint.rstrip("/")
+        self.azure_endpoint = azure_endpoint
 
         if api_version:
             self.api_version = api_version
@@ -48,6 +77,7 @@ def __init__(
             api_key=self.api_key,
             azure_endpoint=azure_endpoint,
             api_version=self.api_version,
+            http_client=make_logging_http_client(self.DEFAULT_TIMEOUT_SECONDS),
         )
 
         # Check for use_responses_api config flag

dana/common/llm/providers/gemini.py

@@ -251,6 +251,17 @@ async def chat(self, messages: list[LLMMessage], tools: list | None = None, **kw
                 if fr:
                     finish_reason = str(fr)
 
+            # Gemini silently truncates over-budget prompts via MAX_TOKENS finish
+            # reason — there is no SDK PTL error. Best-effort WARNING log only;
+            # reactive_compact cannot run here (indistinguishable from output-limit
+            # truncation). Ops must tune DANA_COMPACT_TRIGGER_TOKENS conservatively.
+            if finish_reason and "MAX_TOKENS" in finish_reason:
+                logger.warning(
+                    "gemini_max_tokens_finish",
+                    model=self.model,
+                    note="may indicate context-window overflow; tune DANA_COMPACT_TRIGGER_TOKENS",
+                )
+
             return LLMResponse(
                 content=content,
                 model=self.model,

dana/common/llm/providers/moonshot.py

@@ -9,7 +9,7 @@
 
 from ...config import config_manager
 from ..types import read_media_as_base64, unsupported_placeholder
-from .openai_compatible_base import OpenAICompatibleProvider
+from .openai_compatible_base import OpenAICompatibleProvider, make_logging_http_client
 
 
 logger = structlog.get_logger()
@@ -36,11 +36,13 @@ def __init__(self, api_key: str | None = None, model: str = "kimi-k2.5", base_ur
         if base_url:
             self.base_url = base_url
         else:
-            self.base_url = (
-                config_manager.get_provider_base_url(MOONSHOT_PROVIDER_NAME) or "https://api.moonshot.ai/v1"
-            )
+            self.base_url = config_manager.get_provider_base_url(MOONSHOT_PROVIDER_NAME) or "https://api.moonshot.ai/v1"
 
-        self.client = AsyncOpenAI(api_key=self.api_key, base_url=self.base_url)
+        self.client = AsyncOpenAI(
+            api_key=self.api_key,
+            base_url=self.base_url,
+            http_client=make_logging_http_client(self.DEFAULT_TIMEOUT_SECONDS),
+        )
 
         # Moonshot never uses Responses API
         self._use_responses_api = False

dana/common/llm/providers/openai.py

@@ -4,7 +4,7 @@
 import structlog
 
 from ...config import config_manager
-from .openai_compatible_base import OpenAICompatibleProvider
+from .openai_compatible_base import OpenAICompatibleProvider, make_logging_http_client
 
 
 logger = structlog.get_logger()
@@ -13,6 +13,21 @@
 class OpenAIProvider(OpenAICompatibleProvider):
     """OpenAI API provider."""
 
+    # Env var operators use to dial reasoning effort for OpenAI direct API.
+    # Valid values: "minimal" | "low" | "medium" | "high".
+    _REASONING_EFFORT_ENV_VAR = "OPENAI_THINKING_EFFORT"
+
+    # Env var to force/disable the Responses API for this provider.
+    # Truthy: 1/true/on/yes — falsy: 0/false/off/no. Overrides the config flag.
+    _RESPONSES_API_ENV_VAR = "OPENAI_USE_RESPONSES_API"
+
+    @property
+    def name(self) -> str:
+        return "openai"
+
+    def _endpoint_url(self) -> str:
+        return getattr(self, "base_url", None) or "https://api.openai.com/v1"
+
     def __init__(self, api_key: str | None = None, model: str = "gpt-3.5-turbo", base_url: str | None = None):
         self.model = model
 
@@ -31,7 +46,11 @@ def __init__(self, api_key: str | None = None, model: str = "gpt-3.5-turbo", bas
         else:
             self.base_url = config_manager.get_provider_base_url("openai")
 
-        self.client = AsyncOpenAI(api_key=self.api_key, base_url=self.base_url)
+        self.client = AsyncOpenAI(
+            api_key=self.api_key,
+            base_url=self.base_url,
+            http_client=make_logging_http_client(self.DEFAULT_TIMEOUT_SECONDS),
+        )
 
         # Check for use_responses_api config flag
         provider_config = config_manager.get_provider_config("openai")