talkops-opscloud 0.3.0b7__tar.gz → 0.3.0b9__tar.gz

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: talkops-opscloud
-Version: 0.3.0b7
+Version: 0.3.0b9
 Summary: Extensible Terminal Multi-Agent Framework for Cloud Operations & DevOps Coding
 License-File: LICENSE
 Requires-Python: >=3.12

talkops_opscloud-0.3.0b9/docs/issues/MCP_PERFORMANCE_AND_CONTEXT_EXPLOSION_AUDIT.md

@@ -0,0 +1,275 @@
+# Engineering Issue & Architectural Audit: MCP Tool Schema Explosion & Extreme Latency in Plugin Skills Flow
+
+**Issue ID:** `ISSUE-OPSCLOUD-MCP-001`  
+**Date:** October 8, 2026  
+**Severity:** **P0 — Blocker**  
+**Component:** Deep Agent Core / Subagent Flow / MCP Session Manager / Plugin Adapters  
+**Target Environments:** `src/opscloud` (OpsCloud Orchestrator) vs. `reference/dcode` (Reference Implementation)  
+**Authors:** AI Architecture & Core Systems Audit  
+**Audience:** Core Engineering Team, Agent Platform Engineers, DevOps & Infrastructure  
+
+---
+
+## 1. Executive Summary & Impact Analysis
+
+During testing of skill plugins equipped with MCP (Model Context Protocol) servers (e.g., `aws-containers@talkops-devops-plugins`), the Deep Agent experienced catastrophic response latency and token usage degradation:
+
+* **OpsCloud Turn 1 Performance:** **315.79 seconds (5.22 minutes)** total turn duration, processing **211,300 input tokens** ($0.1531 cost) on a simple user command.
+* **Reference `dcode` Performance:** **15.10s P50** across 4 turns (Turn 1: 0.73s, Turn 2: 25.39s [45.37K tokens], Turn 3: 15.09s [46.6K tokens], Turn 4: 40.15s [254.8K tokens]).
+* **Discrepancy Ratio:** **~21x Latency Explosion**, **~4.7x Initial Prompt Context Explosion**.
+
+```
+┌───────────────────────────────────────────────────────────────────────────────┐
+│                      INITIAL TURN LATENCY & CONTEXT COMPARISON                │
+├──────────────────┬─────────────────────────────┬──────────────────────────────┤
+│ Metric           │ OpsCloud (Current)          │ reference/dcode (Benchmark)  │
+├──────────────────┼─────────────────────────────┼──────────────────────────────┤
+│ Turn 1 Latency   │ 315.79s (5m 15s) ⚠️ BLOCKER │ 0.73s - 25.39s (P50: 15.10s) │
+│ Turn 1 Tokens    │ 211,300 tokens (Context full)│ ~45,370 tokens               │
+│ Tool Injection   │ Eager / Global / Unscoped   │ Scoped / Filtered            │
+│ Schema Strategy  │ Recursive $ref Dereference  │ Protocol Native (Compact)    │
+│ Session / Loop   │ Ephemeral Thread Loop + Stdio│ Server Loop Router Client    │
+└──────────────────┴─────────────────────────────┴──────────────────────────────┘
+```
+
+The system became completely unresponsive for 5+ minutes because **the entire LLM context window was inundated with over 211K tokens of recursively inlined JSON schemas for dozens of un-scoped Kubernetes/AWS tools**, which then had to be re-evaluated across every internal graph step over disjointed stdio subprocess loops.
+
+---
+
+## 2. Trace & Incident Evidence Breakdown
+
+### 2.1 OpsCloud Trace (Image 1: `media_1791459244211.png`)
+* **Execution Status:** 1 Turn, Duration: **315.79s**, Context: **211.3K tokens**, Cost: **$0.1531**.
+* **Observed Bound Tools:**
+  * Core: `create_temp_artifact`, `delete_temp_artifact`, `fetch_url`.
+  * Finch MCP: `plugin__aws-containers_ta_finch_build_container_ima_df47730a85e8`, `plugin__aws-containers_talkops-dev_finch_push_image_6b02bcde1560`, `plugin__aws-containers_talkop_finch_create_ecr_repo_a8f88f0a1f9f`.
+  * EKS / Kubernetes MCP: `plugin__aws-containers_talkops-d_list_k8s_resources_5aa672efc864`, `plugin__aws-containers_talkops-devops-_get_pod_logs_a03a76775bc1`, `plugin__aws-containers_talkops-devop_get_k8s_events_465d1e587238`, `plugin__aws-containers_talkops-de_list_api_versions_c0a261010b5e`, `plugin__aws-containers_talkops-_manage_k8s_resource_b877895269a7`, `plugin__aws-containers_talkops-devops-pl_apply_yaml_34e08bc5a89a`, `plugin__aws-containers_talkop_generate_app_manifest_555c8188f2bb`...
+* **Symptom:** Every single tool from all MCP servers declared in the plugin was injected directly into the main agent's LLM prompt, expanding schemas recursively.
+
+### 2.2 Reference `dcode` Trace (Image 2: `media_1791459309416.png`)
+* **Execution Status:** 4 Turns, P50: **15.10s**, Cost: **$0.1014**.
+  * Turn 1: **0.73s** (Goal initialization)
+  * Turn 2: **25.39s** (45.37K tokens)
+  * Turn 3: **15.09s** (46.6K tokens)
+  * Turn 4: **40.15s** (254.8K tokens after 4 rounds of conversation history accumulation)
+* **Observed Bound Tools:**
+  * Core: `update_goal`, `ask_user`, `js_eval`, `create_temp_artifact`, `delete_temp_artifact`, `fetch_url`, `get_current_thread_id`.
+  * Compute MCP (`aws-compute`): Only 6 focused tools (`aws__get_regional_availa`, `aws__search_documentatio`, `aws__read_documentation`, `aws__retrieve_skill`, `aws__list_regions`, `aws__get_tasks`).
+* **Symptom:** Fast execution, minimal initial token overhead, and precise tool availability.
+
+---
+
+## 3. Deep Root-Cause Analysis (Forensic Audit)
+
+Our cross-codebase audit identified **5 compounding architectural flaws** in OpsCloud that do not exist in `reference/dcode`.
+
+---
+
+### Root Cause 1: Recursive `$ref` Dereferencing in `_clean_mcp_schema`
+**Location:** [`src/opscloud/mcp/session_manager.py:68–85`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/mcp/session_manager.py#L68-L85)
+
+#### What OpsCloud Does:
+```python
+def _clean_mcp_schema(schema: Any) -> dict[str, Any]:
+    if not isinstance(schema, dict):
+        return {}
+
+    from langchain_core.utils.json_schema import dereference_refs
+
+    try:
+        resolved = dereference_refs(schema)
+    except Exception as exc:
+        logger.debug("Failed to dereference MCP schema refs: %s", exc)
+        resolved = schema
+
+    cleaned = dict(resolved) if isinstance(resolved, dict) else dict(schema)
+    for key in ("$schema", "$id", "$defs", "definitions", "additionalProperties"):
+        cleaned.pop(key, None)
+    ...
+```
+
+#### Why This Explodes Context to 211.3K Tokens:
+* When `awslabs.eks-mcp-server` loads, tools like `apply_yaml`, `manage_k8s_resource`, and `generate_app_manifest` define parameters that accept complete Kubernetes manifests.
+* The underlying JSON schema contains deep OpenAPI definitions (`#/$defs/ObjectMeta`, `#/$defs/PodSpec`, `#/$defs/Container`, `#/$defs/EnvVar`, etc.) with multiple recursive and cross-referencing branches.
+* `dereference_refs()` replaces every `{"$ref": ...}` pointer with a **deep, full copy** of the referenced definition.
+* If a single schema references sub-models 15 times, inlining them causes exponential duplication. A single tool's schema grows from a compact 1.5KB reference graph into a **60KB–100KB JSON behemoth**.
+* Multiplying this by 10+ Kubernetes tools in `aws-containers` results in **150,000+ tokens generated from tool schema definitions alone**.
+
+#### What `reference/dcode` Does Differently:
+**Location:** [`reference/dcode/mcp_tools.py:1514–1520`](file:///Users/structbinary/Documents/work/talkops/opscloud/reference/dcode/mcp_tools.py#L1514-L1520)
+* `dcode` passes the raw MCP schema directly to `langchain.mcp.as_langchain_tool(routed_tool, client)`.
+* It **never calls `dereference_refs`**. The schema remains compact and standard, allowing the LLM provider's API client to handle references natively without exponential duplication.
+
+---
+
+### Root Cause 2: Global Eager Tool Injection of Vertical Plugins into Main Coordinator Agent
+**Locations:**
+* [`src/opscloud/plugins/models.py:174–185`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/plugins/models.py#L174-L185)
+* [`src/opscloud/plugins/adapters/mcp.py:200–203`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/plugins/adapters/mcp.py#L200-L203)
+* [`src/opscloud/agent/factory.py:604–660`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/agent/factory.py#L604-L660)
+
+#### The Fault Chain:
+1. In `src/opscloud/plugins/models.py`:
+   ```python
+   @property
+   def is_agent_plugin(self) -> bool:
+       """Return True if this plugin provides autonomous agents (agent-based plugin)."""
+       if self.inventory.agents:
+           ...
+       if self.root and isinstance(self.root, Path):
+           agents_dir = self.root / "agents"
+           return agents_dir.is_dir() and any(agents_dir.glob("*.md"))
+       return False
+   ```
+2. In `src/opscloud/plugins/adapters/mcp.py`:
+   ```python
+   def plugin_mcp_configs(plugins, *, include_subagents: bool = False):
+       for plugin in plugins:
+           if not include_subagents and plugin.is_agent_plugin:
+               continue  # Dynamic subagent plugins isolate their own MCP servers
+   ```
+3. Vertical plugins (such as `aws-containers`, `aws-networking`, `aws-web-and-mobile`) and skill plugins do **not** define an `agents/*.md` directory. Therefore, `plugin.is_agent_plugin` evaluates to **`False`**.
+4. Consequently, when `factory.py` calls `discover_mcp_configs()` to build the **Main Coordinator Agent**, all MCP servers declared by `aws-containers` (`finch`, `eks`, `ecs`, `aws-mcp`) are treated as global coordinator tools!
+5. In `factory.py` lines 654–657:
+   ```python
+   discovered_tools = build_mcp_tools_from_server_infos(coordinator_infos, mcp_manager)
+   if discovered_tools:
+       mcp_tools_list.extend(discovered_tools)
+       all_tools.extend(discovered_tools)  # <-- Injected into all_tools!
+   ```
+6. In `factory.py` line 1081:
+   ```python
+   graph = create_deep_agent(
+       model=active_model,
+       tools=all_tools,  # <-- ALL vertical plugin tools bound to Main Agent!
+       ...
+   )
+   ```
+7. **The result:** The main agent is inundated with 40+ specialized container/Kubernetes tools that should have been isolated to a specialized subagent or dynamically mounted skill.
+
+---
+
+### Root Cause 3: `ToolFilterMiddleware` Does NOT Filter Tools from the Model Request
+**Location:** [`src/opscloud/middleware/tool_filter.py:412–431`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/middleware/tool_filter.py#L412-L431)
+
+#### The Missing Hook:
+* `ToolFilterMiddleware` only implements:
+  - `wrap_tool_call(self, request: ToolCallRequest, handler)`
+  - `awrap_tool_call(self, request: ToolCallRequest, handler)`
+* It **does not implement** `wrap_model_call`, `awrap_model_call`, or `modify_request`.
+* In LangGraph / DeepAgents, `tools` are passed to `model.bind_tools(self.tools)`.
+* Because `ToolFilterMiddleware` does not intercept the model call request to prune `request.tools`, **100% of all registered tools are sent in the system prompt / tool definitions to the LLM on every single turn**.
+* The filter only functions as an execution-time authorization gate (blocking the tool *after* the LLM decides to call it). It provides **zero context-reduction benefits**.
+
+---
+
+### Root Cause 4: Multi-Step Compounding Turn Latency (315 Seconds TTFT)
+* Processing **211,300 input tokens** on remote LLM endpoints (Gemini 2.5 / 3.7 Flash, Claude 3.5 Sonnet) incurs an extreme Time-To-First-Token (TTFT) penalty: ~30 to 70 seconds per call.
+* When the agent runs an agentic step (e.g., executing a command or asking for user clarification), LangGraph's Pregel engine executes sequential model nodes (`model` -> `tools` -> `model` -> `tools` -> `model`).
+* Because the tool schema footprint remains at 211K tokens, **each model call must re-send and the LLM must re-process the full 211K tokens**.
+* 4 sequential steps × 75 seconds per step = **300+ seconds (5.22 minutes)**.
+
+---
+
+### Root Cause 5: Event Loop Mismatch & Stdio Process Locking/Churn
+**Locations:**
+* [`src/opscloud/agent/factory.py:643–646`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/agent/factory.py#L643-L646)
+* [`src/opscloud/mcp/session_manager.py:530–547`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/mcp/session_manager.py#L530-L547)
+
+#### Thread & Event Loop Lifecycles in OpsCloud:
+1. In `factory.py`:
+   ```python
+   if loop is not None and loop.is_running():
+       with concurrent.futures.ThreadPoolExecutor(max_workers=1) as pool:
+           new_infos = pool.submit(asyncio.run, preload_mcp_metadata(missing_configs)).result()
+   else:
+       new_infos = asyncio.run(preload_mcp_metadata(missing_configs))
+   ```
+   `asyncio.run()` creates a temporary event loop. Preloading connects via stdio, fetches tool names, and when `asyncio.run()` finishes, **that event loop is destroyed**. Any stdio pipes or process handles bound to that loop become dead or detached.
+2. Later, when an agent turn executes on the server's `uvloop`, `get_session(server_name)` finds that no live session exists for that loop and calls `connect_server()` to start a brand new stdio connection via `uvx`.
+3. Running multiple `uvx` instances (`uvx awslabs.eks-mcp-server@latest`, `uvx awslabs.finch-mcp-server@latest`, `uvx ecs-mcp-server`) concurrently causes:
+   - Python virtual environment creation / cache resolution latency for each `uvx` invocation.
+   - Stdio process blocking and mutex lock contention under `_get_server_lock(server_name)`.
+   - Thread stalls while waiting for child processes to initialize their IPC pipes.
+
+#### What `reference/dcode` Does Differently:
+**Locations:**
+* [`reference/dcode/server_graph.py:161–168`](file:///Users/structbinary/Documents/work/talkops/opscloud/reference/dcode/server_graph.py#L161-L168)
+* [`reference/dcode/mcp_tools.py:2247–2251`](file:///Users/structbinary/Documents/work/talkops/opscloud/reference/dcode/mcp_tools.py#L2247-L2251)
+* `dcode` initializes a single, shared `MCPSessionManager` directly on the server's running event loop via `_get_mcp_session_manager()`.
+* Backends are mounted onto a single FastMCP router client.
+* Preflight and connection setup are executed concurrently with bounded concurrency:
+  ```python
+  preflight_results = await _gather_bounded(
+      [functools.partial(_preflight_and_connect, name, cfg) for name, cfg in server_items],
+      limit=_MCP_LOAD_CONCURRENCY,
+  )
+  ```
+* All tools route through this unified client. There are no detached thread pools or orphaned event loops.
+
+---
+
+## 4. Architectural Comparison: OpsCloud vs. `dcode`
+
+| Architectural Dimension | OpsCloud (`src/opscloud`) | Reference Implementation (`reference/dcode`) |
+| :--- | :--- | :--- |
+| **Tool Adapter Layer** | `langchain_mcp_adapters` with custom `StructuredTool` wrapping | `fastmcp.Client` + `langchain.mcp.as_langchain_tool` |
+| **JSON Schema Handling** | **Recursive `dereference_refs()`** expanding all definitions inline | Native schema pass-through without ref dereferencing |
+| **Plugin MCP Scoping** | Only checks `plugin.is_agent_plugin` (`agents/*.md`). Vertical/Skill plugins dump all tools to Main Agent | Strict config discovery; plugins inject scoped configs with user trust gating |
+| **Model Prompt Filtering** | **None.** `ToolFilterMiddleware` only acts at tool execution time | `_apply_tool_filter` prunes tools before agent construction via `allowedTools`/`disabledTools` |
+| **Session Lifetime** | Eager thread preloading with `asyncio.run()`, on-demand reconnect on separate `uvloop` | Singleton `MCPSessionManager` bound to server event loop |
+| **Connection Pooling** | Separate `_MCPSessionEntry` per server with individual locks | Single unified FastMCP router client multiplexing all backends |
+| **Teardown & Cleanup** | `session.invalidate()` per server | Bounded 5s teardown via `AsyncExitStack` |
+
+---
+
+## 5. Engineering Action Plan & Implementation Guidance
+
+To permanently resolve this blocker and achieve the ~15s response latency of `reference/dcode`, the engineering team must implement the following four-phase remediation:
+
+### Phase 1: Eliminate Recursive Schema Expansion (Immediate Fix)
+**Target:** [`src/opscloud/mcp/session_manager.py:68–103`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/mcp/session_manager.py#L68-L103)
+1. **Remove `dereference_refs()`**: Cease using `langchain_core.utils.json_schema.dereference_refs` on raw MCP tool schemas.
+2. **Schema Sanitization**: Instead of deep inlining, clean schemas by stripping provider-incompatible root keys (`$schema`, `$id`) while preserving local `$defs` or flattening top-level properties shallowly.
+3. **Verify Schemas**: Ensure parameters for tools like `apply_yaml` and `manage_k8s_resource` pass concise schemas to the model.
+
+### Phase 2: Restrict Plugin MCP Tool Scope to Subagents / Skills
+**Target:** [`src/opscloud/plugins/adapters/mcp.py`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/plugins/adapters/mcp.py) & [`src/opscloud/agent/factory.py`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/agent/factory.py)
+1. **Update Plugin Classification**: Plugins that provide skills with MCP servers must not automatically dump their MCP tools into the Main Coordinator Agent (`all_tools`).
+2. **Scope Vertical Plugins**: Tools from vertical plugins (`aws-containers`, `aws-networking`, etc.) should only be attached to their respective specialized subagents or mounted when their specific skill is activated.
+3. **Main Agent Tool Budget**: Ensure the Main Coordinator Agent is initialized with only core tools (`fetch_url`, `web_search`, `read_file`, `run_command`, `task`, `ask_user`), keeping initial context below **15,000 tokens**.
+
+### Phase 3: Implement Model-Level Tool Filtering in `ToolFilterMiddleware`
+**Target:** [`src/opscloud/middleware/tool_filter.py`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/middleware/tool_filter.py)
+1. Implement `wrap_model_call` and `awrap_model_call` (or `modify_request`):
+   ```python
+   async def awrap_model_call(
+       self,
+       request: ModelRequest,
+       handler: Callable[[ModelRequest], Awaitable[ModelResponse]],
+   ) -> ModelResponse:
+       if self._allowed_patterns and request.tools:
+           filtered_tools = [
+               t for t in request.tools
+               if self.is_tool_allowed(getattr(t, "name", str(t)), tool_obj=t)
+           ]
+           request = request.override(tools=filtered_tools)
+       return await handler(request)
+   ```
+2. This guarantees that subagents with restricted tool lists (e.g., `allowed_tools: ["aws-finops-*"]`) only send their authorized tools to the LLM prompt.
+
+### Phase 4: Align Session Lifecycle with `dcode` Pattern
+**Target:** [`src/opscloud/mcp/session_manager.py`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/mcp/session_manager.py) & [`src/opscloud/server/server_graph.py`](file:///Users/structbinary/Documents/work/talkops/opscloud/src/opscloud/server/server_graph.py)
+1. **Bind Session Manager to Server Loop**: Ensure `MCPSessionManager` is instantiated on the server's primary event loop, eliminating cross-thread `asyncio.run()` preloading.
+2. **Router Client Multiplexing**: Adopt FastMCP router client mounting so all stdio/http transports share a single client context.
+3. **Graceful Subprocess Management**: Ensure bounded subprocess shutdown (5s timeout) to prevent dangling `uvx` / `docker` processes on server reload.
+
+---
+
+## 6. Verification & Acceptance Criteria
+
+The fix will be considered complete and verified when:
+1. **Context Benchmark:** Turn 1 prompt context with `aws-containers` installed does not exceed **45,000 tokens** (comparable to `dcode`).
+2. **Latency Benchmark:** Turn 1 execution latency drops from **315 seconds to under 25 seconds** on the same test query.
+3. **No Schema Inlining Bloat:** Inspection of LangSmith tool definitions confirms that Kubernetes OpenAPI definitions are not duplicated into thousands of lines of JSON.
+4. **Subagent Isolation:** Tools from `aws-containers` only appear in turns where the container subagent/skill is explicitly dispatched.

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "talkops-opscloud"
-version = "0.3.0b7"
+version = "0.3.0b9"
 description = "Extensible Terminal Multi-Agent Framework for Cloud Operations & DevOps Coding"
 readme = "README.md"
 requires-python = ">=3.12"

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/src/opscloud/agent/factory.py

@@ -415,9 +415,9 @@ def _subagent_cli_middleware(
 
     server_keys: list[str] = []
     if mcp_config:
-        server_keys.extend(mcp_config.keys())
+        server_keys.extend(str(k) for k in mcp_config.keys())
     if mcp_server_info:
-        server_keys.extend(getattr(s, "name", str(s)) for s in mcp_server_info)
+        server_keys.extend(str(getattr(s, "name", s)) for s in mcp_server_info)
 
     if allowed_tools or capabilities:
         middleware.append(
@@ -603,13 +603,16 @@ def create_opscloud_agent(
 
     # 5. MCP Discovery
     mcp_tools_list: list[BaseTool] = []
-    cached_infos: list[Any] | None = None
-    mcp_configs: dict[str, Any] | None = None
+    cached_infos: list[Any] | None = kwargs.get("mcp_server_info")
+    mcp_configs: dict[str, Any] | None = kwargs.get("mcp_config")
     if mcp_tools is not None:
         mcp_tools_list = list(mcp_tools)
         all_tools.extend(mcp_tools)
+        if mcp_configs is None:
+            mcp_configs = discover_mcp_configs(project_scope_root)
     else:
-        mcp_configs = discover_mcp_configs(project_scope_root)
+        if mcp_configs is None:
+            mcp_configs = discover_mcp_configs(project_scope_root)
         from opscloud.mcp.preload import (
             get_cached_mcp_server_infos,
             preload_mcp_metadata,
@@ -707,6 +710,7 @@ def create_opscloud_agent(
     from deepagents.middleware.subagents import GENERAL_PURPOSE_SUBAGENT, SubAgent
 
     compiled_subagents: list[SubAgent] = []
+    base_subagent_tools = [t for t in all_tools if t not in mcp_tools_list]
     for name, subagent_meta in subagent_by_name.items():
         model_spec = subagent_meta.get("model")
         has_explicit_model = bool(model_spec)
@@ -723,6 +727,18 @@ def create_opscloud_agent(
         subagent_path = subagent_meta.get("path")
         custom_mw = subagent_meta.get("middleware")
 
+        # Build per-subagent MCP config and tools for this subagent
+        from opscloud.plugins.adapters.mcp import prepare_subagent_mcp
+
+        servers, subagent_mcp_tools, sub_mcp_server_infos = prepare_subagent_mcp(
+            subagent_name, subagent_meta, project_scope_root
+        )
+
+        if subagent_mcp_tools:
+            subagent_dict["tools"] = [*base_subagent_tools, *subagent_mcp_tools]
+        else:
+            subagent_dict["tools"] = list(base_subagent_tools)
+
         sub_middleware = _subagent_cli_middleware(
             has_explicit_model=has_explicit_model,
             assistant_id=assistant_id,
@@ -738,6 +754,9 @@ def create_opscloud_agent(
             worktree_root=effective_cwd,
             subagent_path=subagent_path if subagent_path else None,
             custom_middleware=custom_mw if isinstance(custom_mw, list) else None,
+            mcp_server_info=sub_mcp_server_infos or None,
+            mcp_config=servers or None,
+            mcp_tools=subagent_mcp_tools or None,
             backend=composite_backend,
             model=active_model,
         )
@@ -748,9 +767,10 @@ def create_opscloud_agent(
 
         compiled_subagents.append(cast(SubAgent, subagent_dict))
         logger.info(
-            "Registered declarative subagent '%s'",
+            "Compiled subagent '%s' (tools: %d)",
             subagent_name,
-            extra={"subagent": subagent_name},
+            len(subagent_dict.get("tools", [])),
+            extra={"subagent": subagent_name, "tools_count": len(subagent_dict.get("tools", []))},
         )
 
     if not any(sub.get("name") == GENERAL_PURPOSE_SUBAGENT["name"] for sub in compiled_subagents):
@@ -773,6 +793,7 @@ def create_opscloud_agent(
             "name": gp_name,
             "description": gp_description,
             "system_prompt": gp_system_prompt,
+            "tools": list(base_subagent_tools),
             "middleware": gp_middleware,
         }
         if interrupt_on is not None:
@@ -832,12 +853,22 @@ def create_opscloud_agent(
         agent_middleware.append(AskUserMiddleware())
 
     # 10.5 MCP tool arguments and error handling
-    if mcp_tools_list:
-        if cached_infos or mcp_configs:
+    if mcp_tools_list or cached_infos or mcp_configs:
+        mcp_cfg_for_mw = mcp_configs
+        if not mcp_cfg_for_mw and not cached_infos and mcp_tools_list:
+            synth_cfg: dict[str, Any] = {}
+            for t in mcp_tools_list:
+                srv = getattr(t, "metadata", {}).get("_mcp_server") if isinstance(getattr(t, "metadata", None), dict) else None
+                if srv and srv not in synth_cfg:
+                    synth_cfg[srv] = {"command": "custom", "args": []}
+            if synth_cfg:
+                mcp_cfg_for_mw = synth_cfg
+
+        if cached_infos or mcp_cfg_for_mw:
             from opscloud.middleware.mcp_context import MCPContextMiddleware
 
             agent_middleware.append(
-                MCPContextMiddleware(mcp_server_info=cached_infos, mcp_config=mcp_configs)
+                MCPContextMiddleware(mcp_server_info=cached_infos, mcp_config=mcp_cfg_for_mw)
             )
         agent_middleware.append(MCPToolMiddleware())
 

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/src/opscloud/mcp/discovery.py

@@ -117,7 +117,11 @@ class MCPDiscovery:
         """
         self._store = store
 
-    def discover(self, project_root: Path | None = None) -> dict[str, MCPServerConfig]:
+    def discover(
+        self,
+        project_root: Path | None = None,
+        include_subagents: bool = False,
+    ) -> dict[str, MCPServerConfig]:
         """Discover and merge MCP server configs across global, user, plugin, and project files.
 
         Precedence (lowest to highest):
@@ -158,7 +162,10 @@ class MCPDiscovery:
         try:
             from opscloud.plugins.adapters.mcp import discover_plugin_mcp_configs
 
-            plugin_configs = discover_plugin_mcp_configs(project_dir=effective_project_root)
+            plugin_configs = discover_plugin_mcp_configs(
+                project_dir=effective_project_root,
+                include_subagents=include_subagents,
+            )
             if isinstance(plugin_configs, dict):
                 for name, config in plugin_configs.items():
                     if isinstance(config, dict):
@@ -224,6 +231,9 @@ class MCPDiscovery:
         return results
 
 
-def discover_mcp_configs(project_root: Path | None = None) -> dict[str, MCPServerConfig]:
+def discover_mcp_configs(
+    project_root: Path | None = None,
+    include_subagents: bool = False,
+) -> dict[str, MCPServerConfig]:
     """Discover active MCP server configs from disk."""
-    return MCPDiscovery().discover(project_root)
+    return MCPDiscovery().discover(project_root, include_subagents=include_subagents)

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/src/opscloud/mcp/preload.py

@@ -77,42 +77,63 @@ def _filter_tool_names(
     return True
 
 
-def _clean_mcp_schema(schema: Any) -> dict[str, Any]:
+def _clean_mcp_schema(schema: Any, max_ref_depth: int = 3) -> dict[str, Any]:
     """Clean JSON Schema for compatibility with LLM provider tool schemas (Gemini, Anthropic, OpenAI).
 
-    Dereferences all `$ref` pointers (such as `#/$defs/...` or `#/definitions/...`)
-    using LangChain's `dereference_refs` so that referenced models (like PricingFilter)
-    are inlined directly into properties before removing metadata keys.
+    Safely inlines local `$ref` pointers (such as `#/$defs/...` or `#/definitions/...`)
+    up to a bounded ref depth, strips provider-incompatible metadata keys (`$schema`, `$id`,
+    `additionalProperties`), and prevents recursive bloat or corrupted empty keys.
     """
     if not isinstance(schema, dict):
         return {}
 
-    from langchain_core.utils.json_schema import dereference_refs
-
-    try:
-        resolved = dereference_refs(schema)
-    except Exception as exc:
-        logger.debug("Failed to dereference MCP schema refs: %s", exc)
-        resolved = schema
-
-    cleaned = dict(resolved) if isinstance(resolved, dict) else dict(schema)
-    for key in ("$schema", "$id", "$defs", "definitions", "additionalProperties"):
-        cleaned.pop(key, None)
-
-    properties = cleaned.get("properties")
-    if isinstance(properties, dict):
-        new_props = {}
-        for prop_name, prop_val in properties.items():
-            if isinstance(prop_val, dict):
-                p_clean = dict(prop_val)
-                p_clean.pop("additionalProperties", None)
-                p_clean.pop("$schema", None)
-                new_props[prop_name] = p_clean
-            else:
-                new_props[prop_name] = prop_val
-        cleaned["properties"] = new_props
-
-    return cleaned
+    defs: dict[str, Any] = {}
+    if isinstance(schema.get("$defs"), dict):
+        defs.update(schema["$defs"])
+    if isinstance(schema.get("definitions"), dict):
+        defs.update(schema["definitions"])
+
+    def _resolve(node: Any, ref_depth: int, seen_refs: set[str]) -> Any:
+        if isinstance(node, dict):
+            if "$ref" in node:
+                ref = node["$ref"]
+                if (
+                    ref_depth >= max_ref_depth
+                    or not isinstance(ref, str)
+                    or not ref.startswith("#/")
+                    or ref in seen_refs
+                ):
+                    return {"type": "object"}
+
+                parts = ref.lstrip("#/").split("/")
+                target = None
+                if len(parts) >= 2 and parts[0] in ("$defs", "definitions"):
+                    target = defs.get(parts[1])
+                elif len(parts) == 1 and parts[0] in defs:
+                    target = defs.get(parts[0])
+
+                if isinstance(target, dict):
+                    merged = dict(target)
+                    for k, v in node.items():
+                        if k != "$ref":
+                            merged[k] = v
+                    return _resolve(merged, ref_depth + 1, seen_refs | {ref})
+                return {"type": "object"}
+
+            res = {}
+            for k, v in node.items():
+                if k in ("$schema", "$id", "additionalProperties", "$defs", "definitions"):
+                    continue
+                res[k] = _resolve(v, ref_depth, seen_refs)
+            return res
+
+        if isinstance(node, list):
+            return [_resolve(item, ref_depth, seen_refs) for item in node]
+
+        return node
+
+    result = _resolve(schema, 0, set())
+    return result if isinstance(result, dict) else {}
 
 
 def _clean_stderr_diagnostic(stderr_text: str | None) -> str | None:

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/src/opscloud/mcp/session_manager.py

@@ -65,42 +65,63 @@ def _normalize_mcp_arguments(arguments: dict[str, Any], input_schema: Any) -> di
     return cleaned
 
 
-def _clean_mcp_schema(schema: Any) -> dict[str, Any]:
+def _clean_mcp_schema(schema: Any, max_ref_depth: int = 3) -> dict[str, Any]:
     """Clean JSON Schema for compatibility with LLM provider tool schemas (Gemini, Anthropic, OpenAI).
 
-    Dereferences all `$ref` pointers (such as `#/$defs/...` or `#/definitions/...`)
-    using LangChain's `dereference_refs` so that referenced models (like PricingFilter)
-    are inlined directly into properties before removing metadata keys.
+    Safely inlines local `$ref` pointers (such as `#/$defs/...` or `#/definitions/...`)
+    up to a bounded ref depth, strips provider-incompatible metadata keys (`$schema`, `$id`,
+    `additionalProperties`), and prevents recursive bloat or corrupted empty keys.
     """
     if not isinstance(schema, dict):
         return {}
 
-    from langchain_core.utils.json_schema import dereference_refs
+    defs: dict[str, Any] = {}
+    if isinstance(schema.get("$defs"), dict):
+        defs.update(schema["$defs"])
+    if isinstance(schema.get("definitions"), dict):
+        defs.update(schema["definitions"])
+
+    def _resolve(node: Any, ref_depth: int, seen_refs: set[str]) -> Any:
+        if isinstance(node, dict):
+            if "$ref" in node:
+                ref = node["$ref"]
+                if (
+                    ref_depth >= max_ref_depth
+                    or not isinstance(ref, str)
+                    or not ref.startswith("#/")
+                    or ref in seen_refs
+                ):
+                    return {"type": "object"}
+
+                parts = ref.lstrip("#/").split("/")
+                target = None
+                if len(parts) >= 2 and parts[0] in ("$defs", "definitions"):
+                    target = defs.get(parts[1])
+                elif len(parts) == 1 and parts[0] in defs:
+                    target = defs.get(parts[0])
+
+                if isinstance(target, dict):
+                    merged = dict(target)
+                    for k, v in node.items():
+                        if k != "$ref":
+                            merged[k] = v
+                    return _resolve(merged, ref_depth + 1, seen_refs | {ref})
+                return {"type": "object"}
+
+            res = {}
+            for k, v in node.items():
+                if k in ("$schema", "$id", "additionalProperties", "$defs", "definitions"):
+                    continue
+                res[k] = _resolve(v, ref_depth, seen_refs)
+            return res
 
-    try:
-        resolved = dereference_refs(schema)
-    except Exception as exc:
-        logger.debug("Failed to dereference MCP schema refs: %s", exc)
-        resolved = schema
-
-    cleaned = dict(resolved) if isinstance(resolved, dict) else dict(schema)
-    for key in ("$schema", "$id", "$defs", "definitions", "additionalProperties"):
-        cleaned.pop(key, None)
-
-    properties = cleaned.get("properties")
-    if isinstance(properties, dict):
-        new_props = {}
-        for prop_name, prop_val in properties.items():
-            if isinstance(prop_val, dict):
-                p_clean = dict(prop_val)
-                p_clean.pop("additionalProperties", None)
-                p_clean.pop("$schema", None)
-                new_props[prop_name] = p_clean
-            else:
-                new_props[prop_name] = prop_val
-        cleaned["properties"] = new_props
+        if isinstance(node, list):
+            return [_resolve(item, ref_depth, seen_refs) for item in node]
 
-    return cleaned
+        return node
+
+    result = _resolve(schema, 0, set())
+    return result if isinstance(result, dict) else {}
 
 
 def _filter_tool_name(

{talkops_opscloud-0.3.0b7 → talkops_opscloud-0.3.0b9}/src/opscloud/middleware/server_hooks.py

@@ -919,7 +919,7 @@ def _mcp_server_from_tool(tool: object | None) -> str | None:
     metadata = getattr(tool, "metadata", None)
     if not isinstance(metadata, Mapping):
         return None
-    for key in ("mcp_server", "mcp_server_name", "server_name"):
+    for key in ("_mcp_server", "mcp_server", "mcp_server_name", "server_name"):
         value = metadata.get(key)
         if isinstance(value, str) and value:
             return value