Holmesgpt Catalyst — Development skill for Claude Code

HolmesGPT Durable SRE Agent

A durable SRE Agent that uses the open-source [HolmesGPT](https://holmesgpt.dev) agent, as a durable workflow on [Diagrid Catalyst](https://www.diagrid.io/catalyst). A single investigator agent is used to investigate a set of sample failing apps and HolmesGPT decides which tools to call (Kubernetes, ArgoCD, Grafana, Prometheus, GitHub, YugabyteDB, Pulsar, Bash, …). Diagrid Catalyst wraps every LLM call and tool invocation in a durable workflow activity, providing durable execution to the entire system, ensuring that when calls fail or pods go down, the investigation runs to completion without expensively retrying the LLM.

This sample is using Catalyst Cloud, where Catalyst runs as a managed service and the workflow engine and state store is hosted by Diagrid, and nothing additional is running on your Kubernetes cluster.


All application code lives in [`holmes-app/`](./holmes-app/).

How it works

• holmes-app/app_holmes.py is the Chainlit handler. It instantiates a single DaprWorkflowHolmesRunner at startup and streams events from the workflow's event tape into the UI as Chainlit steps.
• holmes-app/holmes_config.yaml configures HolmesGPT's toolsets, MCP servers, and custom skill paths.
• holmes-app/skills/ holds per-incident runbooks as SKILL.md files. HolmesGPT matches incoming questions to a skill by description, fetches the body via the fetch_skill tool, and follows the workflow it prescribes.
• Workflow state, conversation memory, and the event tape are all persisted via Catalyst managed state stores (agent-workflow, agent-memory, agent-registry, type state.diagrid) — provisioned in the Catalyst project, not in-cluster.
• LLM access goes directly through LiteLLM (env-var credentials). The llm-provider Catalyst conversation component is retained for non-investigation LLM flows (e.g. the planned post-investigation summary in Phase 5)