Incident loop

Broken prod → plain-English read in your queue → paste-ready fix prompt in Cursor. This is the default path for vibe coders.

Today: classification in seconds; fix brief via MCP tools. Target: sub-10-second end-to-end diagnosis. MCP needs a Mushi account + project key — not a separate OpenAI/Anthropic key.

Prerequisites

A Mushi project + API key (credentials)
At least one report (SDK capture, test report from npx mushi-mushi, or Sentry inbound)
MCP wired into Cursor — MCP quickstart or Connect & Update → Add to Cursor

Step 1 — Capture or pick a report

New project:


npx mushi-mushi
# wizard writes .env.local + optional test report
mushi connect --wait

Existing report: open Admin → Reports and copy a report UUID, or ask the agent: “list recent mushi reports”.

Step 2 — Pull fix context

In Cursor, ask:

Use Mushi MCP: call get_fix_context for report <uuid> and summarize the root cause in plain English.

get_fix_context returns classification, repro steps, blast radius, and inventory spec context in one payload — plus a fixPrompt: a paste-ready, agent-ready fix brief composed server-side (diagnosis + reproduction + suggested fix + relevant code hints). No second LLM key required to build it. For most bugs you can paste fixPrompt straight into Cursor and skip to Step 4; Step 3 below adds blast-radius and similar-bug context when you want a deeper plan.

Step 3 — Generate the fix prompt

Run the MCP prompt summarize_report_for_fix with the same report ID, or ask:

Run the Mushi prompt summarize_report_for_fix for report <uuid>. Give me a paste-ready Cursor prompt to fix it.

The prompt orchestrates get_fix_context, get_blast_radius, and get_similar_bugs into a structured fix plan.

Step 4 — Paste and ship

Copy the agent’s output into a new Cursor chat (or continue in the same thread) and ask it to implement the fix. Optionally dispatch a draft PR later via dispatch_fix when GitHub is linked.

Troubleshooting

Not seeing what you expected?