xpander-sdk 2.0.366__tar.gz → 2.0.368__tar.gz

{xpander_sdk-2.0.366/src/xpander_sdk.egg-info → xpander_sdk-2.0.368}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xpander-sdk
-Version: 2.0.366
+Version: 2.0.368
 Summary: xpander.ai Backend-as-a-service for AI Agents - SDK
 Home-page: https://www.xpander.ai
 Author: xpanderAI

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/setup.py

@@ -5,7 +5,7 @@ with open("README.md", "r", encoding="utf-8") as fh:
 
 setup(
     name="xpander-sdk",
-    version="2.0.366",
+    version="2.0.368",
     author="xpanderAI",
     author_email="dev@xpander.ai",
     description="xpander.ai Backend-as-a-service for AI Agents - SDK",

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/src/xpander_sdk/__init__.py

@@ -15,7 +15,7 @@ Main Components:
 For more information, visit: https://xpander.ai
 """
 
-__version__ = "2.0.366"
+__version__ = "2.0.368"
 
 # Backend-related imports
 from .modules.backend.backend_module import Backend

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/src/xpander_sdk/core/context_optimizer/constants.py

@@ -119,6 +119,10 @@ SESSION_COMPACT_TIMEOUT = 120
 # wedging the agent loop forever while still tolerating long
 # provider stalls.
 PRE_RETRY_COMPACT_MAX_ATTEMPTS = 10
+# Non-timeout failures (deterministic provider errors) won't self-heal across
+# retries, so they get a much smaller budget than transient timeouts — the full
+# 10x60s backoff on a deterministic error is what wedged the agent for hours.
+PRE_RETRY_COMPACT_MAX_NONTIMEOUT_ATTEMPTS = 2
 PRE_RETRY_COMPACT_RETRY_BASE_DELAY = 2.0  # seconds
 PRE_RETRY_COMPACT_RETRY_MAX_DELAY = 60.0  # seconds — cap on backoff
 PRE_RETRY_COMPACT_RETRY_JITTER = 0.2  # ± fraction of computed delay

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/src/xpander_sdk/core/context_optimizer/context_optimizer.py

@@ -70,6 +70,7 @@ from xpander_sdk.core.context_optimizer.constants import (
     MAX_STAGNANT_COMPACTIONS,
     MIN_TOKENS_FOR_PRE_RETRY_COMPACT,
     PRE_RETRY_COMPACT_MAX_ATTEMPTS,
+    PRE_RETRY_COMPACT_MAX_NONTIMEOUT_ATTEMPTS,
     PRE_RETRY_COMPACT_RETRY_BASE_DELAY,
     PRE_RETRY_COMPACT_RETRY_JITTER,
     PRE_RETRY_COMPACT_RETRY_MAX_DELAY,
@@ -2133,7 +2134,23 @@ class XPanderContextOptimizer(MapReduceMixin, CompressionManager):
                     raise
                 except (asyncio.TimeoutError, Exception) as exc:
                     is_timeout = isinstance(exc, asyncio.TimeoutError)
+                    # Context-overflow can't be cleared by re-sending the same
+                    # oversized payload — layer_2 already exhausted its chunked
+                    # fallback before raising. Bail to plan-retry at once.
+                    if not is_timeout and _is_context_overflow_error(exc):
+                        logger.error(
+                            f"[context-optimizer] pre-retry: compaction hit a "
+                            f"context-overflow ({exc}) — not retryable, giving up"
+                        )
+                        return result
                     last_attempt = attempt == PRE_RETRY_COMPACT_MAX_ATTEMPTS - 1
+                    # Deterministic provider errors won't self-heal, so cap them
+                    # well below the timeout budget instead of burning all 10x.
+                    if (
+                        not is_timeout
+                        and attempt >= PRE_RETRY_COMPACT_MAX_NONTIMEOUT_ATTEMPTS - 1
+                    ):
+                        last_attempt = True
                     reason = (
                         f"timed out after {SESSION_COMPACT_TIMEOUT}s"
                         if is_timeout

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/src/xpander_sdk/models/events.py

@@ -111,3 +111,8 @@ class TaskUpdateEventType(str, Enum):
     InviteMember = "invite_member"
     # a member invite was sent, rendered as a confirmation chip
     MemberInvited = "member_invited"
+    # omni gateway asks the user to provide one or more secrets a downstream
+    # skill / custom code needs, rendered as an interactive card
+    AskForSecret = "ask_for_secret"
+    # secrets were saved, rendered as a confirmation chip (names only, no values)
+    SecretSaved = "secret_saved"

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/src/xpander_sdk/modules/backend/frameworks/agno.py

@@ -3454,12 +3454,18 @@ _MODEL_CONTEXT_WINDOWS_SUBSTRING: List[Tuple[str, int]] = [
 _LOGGED_UNKNOWN_MODEL_IDS: set = set()
 _EMPTY_MODEL_ID_LOGGED = False
 
+# Conservative window for unmapped models: over-budgeting a small-window model
+# pushes the real request past its hard limit (fatal overflow + retry spin),
+# whereas under-budgeting only triggers extra compaction. 128K is a safe floor
+# for modern models; all frontier IDs are in the tables above and bypass this.
+_DEFAULT_UNKNOWN_CONTEXT_WINDOW = 128_000
+
 
 def _detect_context_window(model: Model) -> int:
     """Return the model's max input context in tokens.
 
     Resolution order: exact match → case-insensitive exact → ordered
-    substring fallback → 200K default with a once-per-process warning
+    substring fallback → conservative default with a once-per-process warning
     log so the miss is triagable without flooding output.
     """
     global _EMPTY_MODEL_ID_LOGGED
@@ -3467,10 +3473,11 @@ def _detect_context_window(model: Model) -> int:
     if not raw_id:
         if not _EMPTY_MODEL_ID_LOGGED:
             logger.warning(
-                "[context-optimizer] model.id is empty; defaulting to 200K window"
+                f"[context-optimizer] model.id is empty; defaulting to "
+                f"{_DEFAULT_UNKNOWN_CONTEXT_WINDOW:,} window"
             )
             _EMPTY_MODEL_ID_LOGGED = True
-        return 200_000
+        return _DEFAULT_UNKNOWN_CONTEXT_WINDOW
     if raw_id in _MODEL_CONTEXT_WINDOWS_EXACT:
         return _MODEL_CONTEXT_WINDOWS_EXACT[raw_id]
     lower_id = raw_id.lower()
@@ -3485,11 +3492,12 @@ def _detect_context_window(model: Model) -> int:
             return window
     if raw_id not in _LOGGED_UNKNOWN_MODEL_IDS:
         logger.warning(
-            f"[context-optimizer] unknown model.id='{raw_id}'; "
-            f"defaulting to 200K window. Add to _MODEL_CONTEXT_WINDOWS_EXACT."
+            f"[context-optimizer] unknown model.id='{raw_id}'; defaulting to "
+            f"{_DEFAULT_UNKNOWN_CONTEXT_WINDOW:,} window. "
+            f"Add to _MODEL_CONTEXT_WINDOWS_EXACT."
         )
         _LOGGED_UNKNOWN_MODEL_IDS.add(raw_id)
-    return 200_000
+    return _DEFAULT_UNKNOWN_CONTEXT_WINDOW
 
 
 def _configure_context_optimizer(

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368/src/xpander_sdk.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xpander-sdk
-Version: 2.0.366
+Version: 2.0.368
 Summary: xpander.ai Backend-as-a-service for AI Agents - SDK
 Home-page: https://www.xpander.ai
 Author: xpanderAI

{xpander_sdk-2.0.366 → xpander_sdk-2.0.368}/tests/test_context_window_detection.py

@@ -11,6 +11,7 @@ from __future__ import annotations
 import pytest
 
 from xpander_sdk.modules.backend.frameworks.agno import (
+    _DEFAULT_UNKNOWN_CONTEXT_WINDOW,
     _MODEL_CONTEXT_WINDOWS_EXACT,
     _detect_context_window,
 )
@@ -96,14 +97,19 @@ def test_known_model_resolves_to_documented_window(
     assert _detect_context_window(_M(model_id)) == expected
 
 
-def test_unknown_model_defaults_to_200k() -> None:
-    # An entirely unknown model id falls back to 200K (the safe Claude
-    # default) and the warning log is the only side-effect.
-    assert _detect_context_window(_M("totally-made-up-2099-model")) == 200_000
+def test_unknown_model_defaults_to_conservative_window() -> None:
+    # An unmapped model id falls back to a conservative window so it can't
+    # be told it has more context than it does (over-budgeting an unmapped
+    # small-window model is what caused the non-frontier overflow + spin).
+    assert (
+        _detect_context_window(_M("totally-made-up-2099-model"))
+        == _DEFAULT_UNKNOWN_CONTEXT_WINDOW
+    )
+    assert _DEFAULT_UNKNOWN_CONTEXT_WINDOW < 200_000
 
 
-def test_empty_model_id_defaults_to_200k() -> None:
-    assert _detect_context_window(_M("")) == 200_000
+def test_empty_model_id_defaults_to_conservative_window() -> None:
+    assert _detect_context_window(_M("")) == _DEFAULT_UNKNOWN_CONTEXT_WINDOW
 
 
 def test_substring_fallback_for_unmapped_1m_variant() -> None: