agent

Loom Agent

Run a local Ollama agent task. The agent operates entirely on your machine via Ollama, using tool calling to autonomously read, edit, and search code through PowerShell MCP tools.

Capabilities

• Dual-model routing: Fast model (qwen3:4b) for tool calls, smart model (deepseek-coder-v2:16b) for analysis
• 7 tools: read_file, read_file_lines, edit_file, write_file, search_code, find_files, run_powershell
• Git safety: Auto-creates loom/agent-<timestamp> branch before file writes, shows diff on completion
• Think-then-act: Qwen3 models use built-in chain-of-thought; others get a planning turn
• Retry + caching: Failed tools retry once; file reads are cached within a task
• Syntax validation: Auto-validates .py files after writes
• Session memory: Stores/retrieves task context via Graphiti knowledge graph

Workflow

1. Call the local_agent_task MCP tool with the user's task description
2. Present the agent's AgentResult:
   • response: The agent's final analysis or review
   • files_changed: List of modified files
   • git_branch / git_diff: Branch name and diff summary
   • tool_log: Detailed log of every tool call made
   • validation_results: Syntax check results for written files
   • tool_calls_made / turns_used: Execution stats

Usage Examples

/loom:agent Review src/loom/server.py for security issues
/loom:agent Add input validation to the execute_powershell function
/loom:agent Refactor kan_engine.py to reduce method complexity
/loom:agent Find all TODO comments and summarize them

Constraints

• All file operations pass through the 5-tier safety pipeline (cache + KAN + blocklist + path + Gemma review)
• The agent runs up to 15 tool-calling turns by default
• Gemma safety review fails open (CAUTION, not BLOCKED) if Ollama offline
• Session memory requires Neo4j running (graceful degradation if unavailable)
• The agent works within LOOM_ALLOWED_ROOT only — path safety prevents external access

Alternative: PSKit MCP Tools

If the local agent isn't needed (no autonomous multi-turn flow), use PSKit tools directly:

• mcp__pskit__read_file, edit_file, search_code — same underlying PS functions
• PSKit returns structured TypedDicts — easier for Claude to parse
• PSKit has readOnlyHint annotations — safe tools auto-approve in clients
• For quick reads/edits/searches, PSKit is faster than spawning a full local agent