trace_lm

An MVP for a local, reproducible LLM tracing and incident investigation stack. The goal is to orchestrate an agent that pulls context via MCP, queries observability backends, and produces an RCA report with minimal, deterministic infrastructure.

Note

PoC/MVP stage. Agent wiring, MCP servers, and infra are still evolving.

Demo (PoC)

Planned Scope

• Agent orchestration for investigations and RCA report generation
• MCP tool integrations for observability and GitHub context
• Local RAG indexer for org docs and rag/corpus/
• Docker-first infra stack for traces, metrics, logs, and dashboards

Current Status

• MVP/PoC: SvelteKit chat UI -> FastAPI endpoint -> LangGraph agent flow.
• GitHub MCP adapter wired via Docker stdio (read-only).
• APM agent and tools are stubbed for now.
• GitHub MCP server implementation: src/mcp_github
• Prometheus MCP server scaffold: src/mcp_prometheus
• Agent entrypoint and graph: src/agent/
• GitHub MCP smoke test: tests/smoke_mcp_github.py
• Agent schema draft: rag/corpus/schema.yml

What You Can Do Right Now

• Ask questions via the web UI and get a summarized response.
• See tool usage and decisions recorded in agent traces.
• Run the chat flow locally with FastAPI + SvelteKit.

Agent Flow (viz.md)

---
config:
  flowchart:
    curve: linear
---
graph TD;
    __start__([<p>__start__</p>]):::first
    orchestrator(orchestrator)
    github_review(github review)
    apm_review(apm review)
    cove_node(cove node)
    __end__([<p>__end__</p>]):::last
    __start__ --> orchestrator;
    apm_review --> cove_node;
    github_review --> cove_node;
    orchestrator -. apm .-> apm_review;
    orchestrator -. github .-> github_review;
    cove_node --> __end__;
    classDef default fill:#f2f0ff,line-height:1.2
    classDef first fill-opacity:0
    classDef last fill:#bfb6fc

Loading

MCP Servers

• GitHub MCP server: src/mcp_github (read-only by default, guarded issue creation)
• Prometheus MCP server: src/mcp_prometheus (read-only metrics queries)

Getting Started

Nothing to run yet. This section will evolve as the Docker stack and agent wiring land.

Development setup

• Install dev dependencies: pip install ".[dev]"

TODO

• Stand up the MCP server stack for local development and demos.
• Add a real test suite and document how to run it.
• Keep the README flow diagram in sync with src/agent/viz.md.

Contributing

See AGENTS.md for repository guidelines and expected conventions.