[AAASM-4434] ⬆️ deps: Upgrade Python dependencies to latest stable, incl. LangGraph

agent_assembly/adapters/langchain/callback_handler.py

@@ -4,7 +4,8 @@
 
 import importlib
 import inspect
-from typing import Any, Literal, Mapping, cast
+from collections.abc import Mapping
+from typing import Any, Literal, cast
 from uuid import UUID
 
 from agent_assembly.exceptions import ToolExecutionBlockedError

agent_assembly/adapters/langgraph/patch.py

@@ -216,6 +216,24 @@ def _is_compiled_subgraph(node_executor: Any) -> bool:
     return hasattr(node_executor, "nodes") and hasattr(node_executor, "invoke")
 
 
+def _is_pregel_node_wrapper(node_executor: Any) -> bool:
+    """Return True when node_executor looks like a langgraph PregelNode.
+
+    AAASM-4434: since langgraph's Pregel graph-compilation rewrite (the node-map
+    entries in a CompiledStateGraph.nodes dict stopped being directly-callable/
+    directly-invokable node executors and became ``PregelNode`` wrapper objects),
+    ``PregelNode.invoke``/``.ainvoke`` are present but are *not* what the Pregel
+    runtime calls when executing a node — it dispatches through the wrapped
+    Runnable at ``PregelNode.bound`` instead. A node_executor with a `.bound`
+    attribute that itself exposes invoke/ainvoke is that shape; wrap `.bound`,
+    not the PregelNode itself, or the governance hooks silently never fire.
+    """
+    bound = getattr(node_executor, "bound", None)
+    if bound is None:
+        return False
+    return callable(getattr(bound, "invoke", None)) or callable(getattr(bound, "ainvoke", None))
+
+
 def _current_spawn_depth() -> int:
     """Return depth for a new spawn context: current depth + 1, or 1 if root."""
     current = _SPAWN_CTX.get()
@@ -373,6 +391,12 @@ def _wrap_node_entry(
     if _is_compiled_subgraph(node_executor):
         return _wrap_subgraph_spawn_node(node_map, node_name, node_executor, process_agent_id)
 
+    # AAASM-4434: current langgraph node-map entries are PregelNode wrappers whose
+    # own .invoke/.ainvoke are never called by the Pregel runtime — it dispatches
+    # through PregelNode.bound. Wrap that inner Runnable instead, or hooks never fire.
+    if _is_pregel_node_wrapper(node_executor):
+        return _wrap_node_invoke_methods(node_name, node_executor.bound, callback_handler)
+
     return _wrap_node_invoke_methods(node_name, node_executor, callback_handler)
 
 

agent_assembly/adapters/mcp/patch.py

@@ -5,8 +5,9 @@
 import importlib as importlib
 import importlib.util
 import inspect
+from collections.abc import Mapping
 from dataclasses import dataclass
-from typing import TYPE_CHECKING, Any, Literal, Mapping
+from typing import TYPE_CHECKING, Any, Literal
 
 if TYPE_CHECKING:
     from agent_assembly.exceptions import MCPToolBlockedError
@@ -301,7 +302,7 @@ async def patched_call_tool(self: Any, *args: Any, **kwargs: Any) -> Any:
         return result
 
     setattr(client_session_cls, _ORIGINAL_CALL_TOOL, original_call_tool)
-    setattr(client_session_cls, "call_tool", patched_call_tool)
+    client_session_cls.call_tool = patched_call_tool
     setattr(client_session_cls, _PATCHED_FLAG, True)
 
 
@@ -311,7 +312,7 @@ def _revert_client_session_patch(client_session_cls: type[Any]) -> None:
 
     original_call_tool = getattr(client_session_cls, _ORIGINAL_CALL_TOOL, None)
     if callable(original_call_tool):
-        setattr(client_session_cls, "call_tool", original_call_tool)
+        client_session_cls.call_tool = original_call_tool
 
     if hasattr(client_session_cls, _ORIGINAL_CALL_TOOL):
         delattr(client_session_cls, _ORIGINAL_CALL_TOOL)

agent_assembly/adapters/registry.py

@@ -1,9 +1,10 @@
 from __future__ import annotations
 
+from collections.abc import Callable
 from dataclasses import dataclass
 from importlib import metadata
 from threading import Lock
-from typing import Callable, Literal
+from typing import Literal
 
 from agent_assembly.adapters.agno.adapter import AgnoAdapter
 from agent_assembly.adapters.base import FrameworkAdapter

agent_assembly/cli/adapter_validator.py

@@ -116,9 +116,9 @@ def _check_supported_versions(instance: FrameworkAdapter) -> AdapterValidationRe
     )
 
 
-def _check_register_hooks_signature(cls: type) -> AdapterValidationResult:
+def _check_register_hooks_signature(cls: type[FrameworkAdapter]) -> AdapterValidationResult:
     """Check that register_hooks accepts a GovernanceInterceptor argument."""
-    register_hooks = getattr(cls, "register_hooks")
+    register_hooks = cls.register_hooks
     sig = inspect.signature(register_hooks)
     params = [p for name, p in sig.parameters.items() if name != "self"]
     if not params:

agent_assembly/core/assembly.py

@@ -5,9 +5,10 @@
 import os
 import re
 import sys
+from collections.abc import Callable
 from dataclasses import dataclass, field
 from threading import Lock
-from typing import Any, Callable, Literal, Protocol
+from typing import Any, Literal, Protocol
 
 from agent_assembly.adapters.base import FrameworkAdapter
 from agent_assembly.adapters.langchain.adapter import LangChainAdapter
@@ -58,9 +59,7 @@ def _validate_agent_id(agent_id: str) -> str:
     a malformed socket path.
     """
     if not _AGENT_ID_RE.fullmatch(agent_id):
-        raise ValueError(
-            f"invalid agent_id: {agent_id!r}; must match {_AGENT_ID_RE.pattern}"
-        )
+        raise ValueError(f"invalid agent_id: {agent_id!r}; must match {_AGENT_ID_RE.pattern}")
     return agent_id
 
 

agent_assembly/models/agent.py

@@ -3,7 +3,7 @@
 from __future__ import annotations
 
 from datetime import datetime
-from typing import Any, Optional
+from typing import Any
 
 from pydantic import BaseModel, Field
 
@@ -13,7 +13,7 @@ class AgentConfig(BaseModel):
 
     agent_id: str = Field(..., description="Unique identifier for the agent")
     name: str = Field(..., description="Human-readable name for the agent")
-    description: Optional[str] = Field(None, description="Description of the agent")
+    description: str | None = Field(None, description="Description of the agent")
     version: str = Field(default="0.1.0", description="Agent version")
     created_at: datetime = Field(default_factory=datetime.utcnow, description="Creation timestamp")
     updated_at: datetime = Field(default_factory=datetime.utcnow, description="Last update timestamp")
@@ -24,7 +24,7 @@ class AgentState(BaseModel):
 
     agent_id: str = Field(..., description="Unique identifier for the agent")
     status: str = Field(default="idle", description="Current status of the agent")
-    last_activity: Optional[datetime] = Field(None, description="Last activity timestamp")
+    last_activity: datetime | None = Field(None, description="Last activity timestamp")
     metadata: dict[str, Any] = Field(default_factory=dict, description="Additional state metadata")
 
 
@@ -33,5 +33,5 @@ class PolicyEvaluation(BaseModel):
 
     action: str = Field(..., description="Action that was evaluated")
     allowed: bool = Field(..., description="Whether the action is allowed")
-    reason: Optional[str] = Field(None, description="Reason for the decision")
-    policy_id: Optional[str] = Field(None, description="ID of the policy that made the decision")
+    reason: str | None = Field(None, description="Reason for the decision")
+    policy_id: str | None = Field(None, description="ID of the policy that made the decision")

pyproject.toml

@@ -24,13 +24,28 @@ classifiers = [
     "Programming Language :: Python :: 3.14",
 ]
 dependencies = [
-    "pydantic>=2.0.0,<3.0.0",
+    "pydantic>=2.13.4,<3.0.0",
     "httpx>=0.27.0,<1.0.0",
-    "typing-extensions>=4.0.0",
+    "typing-extensions>=4.16.0",
     # AAASM-1654 (PR-E): OpControlSubscriber consumes PolicyService.OpControlStream
     # via gRPC. Held to grpcio's modern stable line.
-    "grpcio>=1.66,<2",
-    "protobuf>=5,<8",
+    "grpcio>=1.82.1,<2",
+    # AAASM-4434 (dependency completeness re-audit): resolves to 6.33.6, not raw
+    # PyPI latest (7.35.1), even though >=5,<8 permitted 7.x. Confirmed this is a
+    # REAL transitive ceiling, not a stale "permissive floor" resolution: the
+    # `dev`/test group's pydantic-ai>=2.9.0 pulls in
+    # `pydantic-ai-slim[...,logfire,...]`, and logfire (<=4.37.0) pins
+    # `opentelemetry-exporter-otlp-proto-http<1.43.0`; every otlp-proto-http
+    # release in that range depends on an `opentelemetry-proto` that caps
+    # `protobuf<7.0`. Verified by temporarily setting `protobuf>=7.35.1,<8` and
+    # re-running `uv lock`, which fails with a resolver conflict rooted at
+    # pydantic-ai's logfire extra; a bare `uv lock --upgrade` (no floor change)
+    # also leaves protobuf pinned at 6.33.6. Re-audit once logfire raises its
+    # opentelemetry-exporter-otlp-proto-http ceiling above 1.43.0.
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open "5" to the exact resolved 6.33.6 so the manifest states what is
+    # actually locked, not just the outer bound of what's permitted.
+    "protobuf>=6.33.6,<8",
 ]
 
 [project.optional-dependencies]
@@ -50,51 +65,75 @@ Repository = "https://github.com/AI-agent-assembly/python-sdk"
 
 [dependency-groups]
 dev = [
-    "coverage~=7.10",
+    "coverage~=7.15",
     "python-dotenv>=1.0.1,<2",
     # AAASM-1654 (PR-E): grpcio-tools provides protoc + Python plugin used by
     # scripts/gen_proto.py to regenerate agent_assembly/proto/*_pb2*.py from
     # the sibling agent-assembly/proto/ checkout.
-    "grpcio-tools>=1.66,<2",
+    # AAASM-4434 (dependency completeness re-audit): resolves to 1.81.1, not raw
+    # PyPI latest (1.82.1). Not a stale-resolution bug — grpcio-tools 1.82.1
+    # requires protobuf>=7.35.1, which conflicts with the protobuf<7.0 ceiling
+    # forced transitively by pydantic-ai's logfire extra (see the protobuf
+    # comment under [project].dependencies for the full evidence chain).
+    # Re-audit alongside protobuf once that transitive ceiling lifts.
+    "grpcio-tools>=1.81.1,<2",
     { include-group = "lint" },
     { include-group = "test" },
 ]
 lint = [
-    "ruff>=0.1.0",
-    "mypy>=1.11,<3",
+    "ruff>=0.15.21",
+    "mypy>=2.2.0,<3",
 ]
 test = [
-    "pytest>=8.1.1,<10",
-    "pytest-cov>=5.0.0,<8",
-    "pytest-rerunfailures>=14.0,<17",
-    "pytest-asyncio>=0.23.0,<2",
-    "pytest-benchmark>=4.0.0,<6",
+    "pytest>=9.1.1,<10",
+    "pytest-cov>=7.1.0,<8",
+    "pytest-rerunfailures>=16.4,<17",
+    "pytest-asyncio>=1.4.0,<2",
+    "pytest-benchmark>=5.2.3,<6",
     # AAASM-2943: dev/test-only (NOT a runtime dependency). Installing the
     # framework lets the `importorskip`-guarded Pydantic AI integration tests
     # run in CI (the `dev` group includes `test`, and the integration-test job
     # installs `dev`), so the function-tool governance regression is actually
     # exercised instead of skipped.
-    "pydantic-ai>=0.3.0",
+    # AAASM-4434: floor raised 0.3.0 -> 2.0.0 — with an unbounded floor, `uv lock
+    # --upgrade` resolved pydantic-ai down to 0.7.2 rather than latest, because an
+    # unconstrained solve found a cheaper-to-satisfy graph at that older version.
+    # Raising the floor forces resolution to latest stable, which is verified
+    # compatible (protobuf still resolves within our >=5,<8 range, just at
+    # 6.33.6 instead of 7.x).
+    # AAASM-4434 (manifest transparency follow-up): floor tightened again to the
+    # exact resolved version (2.9.0) so the manifest doesn't understate what's
+    # actually locked.
+    "pydantic-ai>=2.9.0",
     # AAASM-3528: dev/test-only (NOT a runtime dependency). The shipped OpenAI
     # Agents framework is the top-level `agents` package (NOT `openai.agents`),
     # and a tool runs via its per-instance `on_invoke_tool` coroutine (NOT a
     # `FunctionTool.__call__`). Installing it lets the `importorskip`-guarded
     # integration test drive a real tool call, so a regression to the old
     # fail-open no-op patch is actually caught instead of silently skipped.
-    "openai-agents>=0.1.0",
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open 0.1.0 to the exact resolved version (0.18.2) so the manifest
+    # doesn't understate what's actually locked.
+    "openai-agents>=0.18.2",
     # AAASM-3539: dev/test-only (NOT a runtime dependency). smolagents routes
     # every tool execution through `smolagents.tools.Tool.__call__` (which calls
     # `self.forward`); installing it lets the `importorskip`-guarded adapter
     # tests drive a real `Tool` subclass, so a regression to a fail-open no-op
     # patch is caught (deny must block `forward`) instead of silently skipped.
-    "smolagents>=1.0.0,<2.0.0",
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open 1.0.0 to the exact resolved version (1.26.0) so the manifest
+    # doesn't understate what's actually locked. Upper bound <2.0.0 unchanged.
+    "smolagents>=1.26.0,<2.0.0",
     # AAASM-3537: dev/test-only (NOT a runtime dependency). Agno (formerly
     # Phidata) runs every function-tool body through
     # `agno.tools.function.FunctionCall.execute` / `aexecute`. Installing it lets
     # the `importorskip`-guarded integration test drive a real `FunctionCall`, so
     # the negative-control deny test fails if the patch ever regresses to a no-op
     # instead of being silently skipped.
-    "agno>=2.0.0",
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open 2.0.0 to the exact resolved version (2.7.2) so the manifest
+    # doesn't understate what's actually locked.
+    "agno>=2.7.2",
     # AAASM-3536: dev/test-only (NOT a runtime dependency). LlamaIndex routes
     # tool execution through the concrete `FunctionTool.call` / `acall` methods;
     # installing it lets the `importorskip`-guarded governance tests drive a
@@ -108,11 +147,25 @@ test = [
     # by the SDK or its tests. No upstream fix exists (3.9.4 is the latest release;
     # `first_patched_version: none`), so no lock upgrade removes it. Tracked in
     # AAASM-4169; Dependabot #27 recommended for dismissal as dev/test-only.
-    "llama-index-core>=0.10.0",
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open 0.10.0 to the exact resolved version (0.14.23) so the manifest
+    # doesn't understate what's actually locked.
+    "llama-index-core>=0.14.23",
+    # AAASM-4434: dev/test-only (NOT a runtime dependency). The LangGraph adapter
+    # (agent_assembly/adapters/langgraph/) never hard-imports langgraph — it
+    # reflectively loads `langgraph.graph.state.StateGraph` via importlib and
+    # structurally duck-types compiled-graph node maps and ToolNode so it degrades
+    # gracefully when the framework isn't installed. That resilience previously
+    # meant the patch was never exercised against a *real* langgraph install.
+    # Pinning latest stable here lets the `importorskip`-guarded real-package test
+    # (test/integration/test_langgraph_real_package_smoke.py) actually compile and
+    # invoke a real StateGraph, proving the reflection-based patch still matches
+    # LangGraph 1.x's current API surface instead of only mocked SimpleNamespaces.
+    "langgraph>=1.2.9,<2",
 ]
 pre-commit-ci = [
-    "pre-commit>=3.5.0,<5",
-    "pylint>=3.1.0,<5",
+    "pre-commit>=4.6.0,<5",
+    "pylint>=4.0.6,<5",
     { include-group = "lint" },
 ]
 # AAASM-4034: dev/test-only (NOT a runtime dependency) and DELIBERATELY excluded
@@ -127,20 +180,23 @@ pre-commit-ci = [
 #   .venv/bin/python -m pytest test/unit/adapters/langchain/test_getattr_contract_with_langchain.py
 langchain-test = [
     { include-group = "test" },
-    "langchain-core>=0.3.0,<1.5.0",
+    "langchain-core>=1.4.9,<1.5.0",
 ]
 docs = [
     "mkdocs>=1.6.0,<2",
     "mkdocs-material>=9.5.0,<10",
-    "mkdocstrings>=0.24.0,<2",
-    "mkdocstrings-python>=1.10.0,<3",
+    "mkdocstrings>=1.0.5,<2",
+    "mkdocstrings-python>=2.0.5,<3",
     "mike>=2.1.0,<3",
-    "mkdocs-autorefs>=1.0.0,<2",
-    "mkdocs-git-revision-date-localized-plugin>=1.2.0,<2",
+    "mkdocs-autorefs>=1.4.4,<2",
+    "mkdocs-git-revision-date-localized-plugin>=1.5.3,<2",
     "mkdocs-git-authors-plugin>=0.9.0,<1",
     # AAASM-4308: shared metadata macros for docs pages. Docs-only — MUST NOT
     # be added to `[project].dependencies` (runtime install stays lean).
-    "mkdocs-macros-plugin>=1,<2",
+    # AAASM-4434 (manifest transparency follow-up): floor tightened from the
+    # wide-open 1 to the exact resolved version (1.5.0) so the manifest
+    # doesn't understate what's actually locked.
+    "mkdocs-macros-plugin>=1.5.0,<2",
 ]
 
 [tool.uv]

ruff.toml

@@ -14,6 +14,12 @@ exclude = [
     "dist",
     "docs_with_docusarus",
     "docs",
+    # AAASM-4434: generated gRPC/protobuf stubs (scripts/gen_proto.py), committed
+    # verbatim; a CI drift check regenerates them and asserts no diff, so they
+    # must not be hand-edited to satisfy lint (mirrors the mypy.ini ignore for the
+    # same directory). The ruff 0.15 bump newly flags unused-argument findings in
+    # the generated grpc servicer stubs.
+    "agent_assembly/proto/",
 ]
 
 line-length = 120
@@ -42,10 +48,24 @@ ignore = [
 
 [lint.per-file-ignores]
 "__init__.py" = ["F401"]
-"test/**/*.py" = ["S101"]
+# AAASM-4434: ARG002/ARG005 (unused method/lambda argument) are the same category
+# already ignored repo-wide for plain functions as ARG001 — mock/callback test
+# doubles routinely accept an argument only to match the interface they stand in
+# for (monkeypatched hooks, callback signatures, fixture-injected params). The
+# ruff 0.15 bump is what newly enabled these two rule IDs; scope the exception to
+# test/ (unlike ARG001) so a genuinely-unused argument in production code under
+# agent_assembly/ still gets flagged.
+"test/**/*.py" = ["S101", "ARG002", "ARG005"]
 
 [lint.isort]
 known-first-party = ["agent_assembly"]
+# AAASM-4434: the pre-commit `isort` hook (real isort, --profile=black) ships a
+# bundled stdlib module list that includes CPython's own `test` package (Lib/
+# test/), so it always classifies this repo's top-level `test.*` imports
+# (e.g. `from test.bench.conftest import ...`) as standard-library. Ruff's isort
+# implementation does not, so without this the two tools disagreed on where
+# `test.*` imports belong and repeatedly re-sorted each other's output.
+extra-standard-library = ["test"]
 
 [format]
 quote-style = "double"