Features/code mode experiments by srimon12 · Pull Request #2 · veristamp/AgentGovernor

srimon12 · 2026-03-07T15:42:56Z

No description provided.

GATE 1 - Static Auditor (Python): - AST parsing to extract tool/skill calls before execution - Policy check against allowed tools - Detects loops, conditionals, nesting depth - Supports both mcp.use() and skill.method() patterns GATE 2 - MCPClientManager (Bun + TypeScript): - Connects to 5 MCP servers (31 tools) - ABAC Policy Engine with scope matching - Auth SDK with JWT validation and kill switch - Audit logging with structured entries - Windows named pipe + Unix socket support Binding Pattern (L1 Skills): - Skills use _binding proxy (LLM never sees raw tool names) - BindingProxy routes all calls through Policy Gate - Skill loader injects _binding at runtime - from skills import filesystem syntax Sandbox Runtime: - NsJail config for Linux kernel isolation - Fallback unsafe mode for Windows development - Python runner with skill package installation Examples: - skill_workflow.py - Uses skills layer - fetch_and_save_docs.py - Fetches from context7, saves locally

TypeScript SDK for MCP agents and resource servers: src/auth/ ├── types.ts - Type definitions (MCPCredentials, MCPToken, ValidationResult, etc.) ├── errors.ts - MCPError, MCPRegistrationError, MCPAuthError, MCPRateLimitError ├── jwt.ts - JWT decoding and validation utilities ├── agent-client.ts - MCPAgentClient for registration and token acquisition ├── resource-server.ts - MCPResourceServer for token validation ├── admin-client.ts - MCPAdminClient for admin operations ├── helpers.ts - Convenience functions (registerAgent, getAccessToken, etc.) ├── index.ts - Barrel export ├── demo.ts - Demo script similar to Python e2e_demo.py └── README.md - SDK documentation Features: - Agent registration with REG_JWT (Budgeted DCR) - Token acquisition with scope and audience (RFC 8707) - JWT validation (stateless, ~0.1ms) - Introspection validation (~35ms) - Kill switch support (requireActiveCheck) - Scope and audience enforcement Updates: - MCPClientManager now uses MCPResourceServer instead of old AuthSDK - Policy Engine updated to handle optional Identity fields - Removed old src/policy/auth.ts in favor of src/auth/ module

…Secure Tool Search - **Registry Architecture**: Unified ToolRegistry, SkillRegistry, and WorkflowRegistry to use a shared PostgreSQL backend (via Drizzle ORM) with FTS5 support. - **Database**: Replaced local file scanning/SQLite with asynchronous, high-performance Postgres connections. Added src/registry/db.ts and schema definitions. - **Tool Search**: Implemented RegistrySearchTool (BM25-style) for dynamic, auth-aware tool discovery. - **Agent**: Updated Agent to support an interactive SEARCH('query') loop, enabling it to discover tools on-demand while enforcing RBAC permissions. - **AuthZ**: Propagated AgentIdentityScope through the search and discovery layer to ensure users only see tools permitted by their roles. - **Cleanup**: Removed legacy compatibility code and standardized ingest()/search() APIs across all registries. - **Tests**: Updated E2E tests to validate the new async registry and search workflows.

…Agents - **Registry Overhaul**: Migrated Tool, Skill, and Workflow registries to use PostgreSQL with Drizzle ORM for high-performance persistence and FTS5 search. - **Async I/O**: Refactored the entire codebase to use Bun's native asynchronous file APIs ('Bun.file', 'Bun.write') replacing synchronous 'fs' calls for better performance. - **Agent Discovery**: Implemented a robust 'Tool Search' loop in the Agent, allowing it to dynamically discover tools via natural language queries (BM25) while enforcing RBAC permissions. - **Testing**: Fixed and stabilized end-to-end tests ('agent_e2e_llm', 'gcm_workflow_reuse') by introducing fake LLM harnesses and isolating test data with unique Organization IDs. - **Security**: Strengthened Authorization propagation in the discovery layer ('RegistrySearchTool') to ensure users only access tools permitted by their roles. - **Cleanup**: Standardized registry interfaces ('ingest', 'search') and removed legacy compatibility layers.

Introduce a structured tool_call/final loop with registry discovery, Gate 1-in-loop validation, and gpt-4o-mini wiring to support UI-driven workflow building.

- Added `types.ts` to define Engram-related types including NodePointer, EngramNode, and various result interfaces. - Introduced `EngramService` interface with methods for searching, inspecting, exploring, and loading content. - Created `test_lifecycle.ts` for testing MCPClientManager lifecycle operations. - Developed `service.ts` for Mission management, including creation, retrieval, and context updates. - Added demo scripts for orchestrating missions and recursive discovery. - Implemented message handling in `message.ts` for session message storage and compaction.

…n and run management - Updated MissionService to allow optional IDs for missions and sessions. - Improved session creation logic to preload context from missions. - Enhanced run creation to handle optional run IDs and associate with sessions and missions. - Introduced capability registry for managing tools, skills, and workflows with caching and permission checks. - Refactored runtime loop and sub-agent to accommodate new session and run structures. - Added capability loading logic to dynamically manage tool and skill capabilities.

…l creator, and related runtime files - Deleted demo files for governance, grand orchestrator, recursive discovery, and skill creator to streamline the codebase. - Updated the loop and sub-agent files to improve session management and tool execution. - Introduced a new session manager for better handling of mission sessions and runs. - Refactored agent management to support workflow agents and improved agent runtime creation. - Enhanced tests to reflect changes in agent structure and functionality.

- Updated JWT claims interface to support both string and array for scopes. - Improved JWT decoding with a base64 URL decode helper function. - Added rotateSecret method to MCPAgentClient for client secret rotation. - Implemented secret rotation flow in demo application. - Introduced safeJson utility function for robust JSON parsing from responses. - Updated OAuth demo to include secret rotation and improved user experience. - Added new OAuth demo script for user flow demonstration. - Updated .gitignore to exclude docs directory.

- Added ContextBuilder interface and StandardContextBuilder implementation for building runtime contexts. - Implemented buildRuntimeContext function for convenience in creating contexts. - Created Skill Creator Executor for phase-based skill creation, including discovery and generation phases. - Added agent configuration types and interfaces for managing agent execution and dependencies. - Introduced analytics and memory management modules for tracking performance and optimizing memory usage. - Implemented caching mechanisms for improved performance in LLM responses.

…rypto - Add grep_search tool for regex content search across files using ripgrep with Node.js fallback - Replace Node.js crypto.randomBytes with Bun-native crypto.getRandomValues for temporary file generation - Replace Node.js crypto.createHash with Bun.crypto.subtle.digestSync for SHA256 hashing (significantly faster on Bun) - Update stitch_file tool description to mark as ADVANCED with byte offset warnings - Update documentation and schema files for new grep_search tool - Remove additionalProperties restriction from context7 schemas

@ts-ignore

- Convert indentation from spaces to tabs - Standardize quote style to double quotes - Reorganize and sort import statements - Break long lines for readability - Use Number.isNaN() instead of global isNaN() - Replace @ts-ignore with @ts-expect-error - Add trailing commas for cleaner diffs - Remove empty api.md file

kb/llm/providers/bedrock_provider.py

+import json
+import boto3
+from botocore.exceptions import ClientError
+, load_dotenv


kb/services/pr_scanner/core.py

+    COMMENT = "comment"  # Neutral - just observations
+
+
+class PRRiskLevel(Enum):


kb/judgment/core.py

+    DRY_RUN = "dry_run"
+
+
+class RiskLevel(Enum):


sandbox/runtime/skill_loader.py

+            raise ImportError(f"Failed to load skill: {skill_name}")
+
+        module = importlib.util.module_from_spec(spec)
+        bindings_obj = bindings
