Observability

Mushi Mushi integrates Langfuse (LLM traces) and Sentry (error tracking). Both are optional and configured as Edge Function secrets.

Langfuse (LLM tracing)

When the Langfuse keys are set, LLM calls are traced to Langfuse with the model, token counts, latency and metadata such as the project and report IDs (string metadata is PII-scrubbed first). Every call is also recorded in the llm_invocations table, Langfuse or not.


cd packages/server
npx supabase secrets set LANGFUSE_PUBLIC_KEY=pk-lf-…
npx supabase secrets set LANGFUSE_SECRET_KEY=sk-lf-…
npx supabase secrets set LANGFUSE_BASE_URL=https://cloud.langfuse.com

LANGFUSE_BASE_URL is the canonical name (matches root .env.example). Docker Compose and older snippets may use LANGFUSE_HOST — edge code accepts both.

Replace LANGFUSE_BASE_URL with your self-hosted Langfuse URL if needed.

Which key paid for a call is recorded in llm_invocations.key_source (byok or env), not in Langfuse. See BYOK.

What gets traced

Trace	Function
Report classification	`fast-filter`, `classify-report`
Fix generation	`fix-worker`
Classification grading	`judge-batch`
Intelligence report	`intelligence-report`
PDCA iteration	`pdca-runner`

Viewing traces

In Langfuse: Traces, then filter by the project ID in the trace metadata. The Health page in the Mushi Mushi admin console lists recent LLM calls from llm_invocations, each with its Langfuse trace ID.

Sentry (error tracking)

Every Edge Function wraps its handler with withSentry() from _shared/sentry.ts. Unhandled exceptions are forwarded to Sentry automatically.


cd packages/server
npx supabase secrets set SENTRY_DSN_SERVER=https://…@sentry.io/…   # SENTRY_DSN also works
npx supabase secrets set SUPABASE_ENV=production                     # Sentry environment
npx supabase secrets set SENTRY_RELEASE=$GIT_SHA

Admin SPA Sentry

The Vite SPA also ships a Sentry integration. Pass the DSN at build time:


VITE_SENTRY_DSN=https://…@sentry.io/… pnpm --filter "@mushi-mushi/admin..." build

OpenTelemetry

When OTEL_EXPORTER_OTLP_ENDPOINT is set, classify-report exports an OTLP span for each classification, as a child of the ingest request’s traceparent when there is one. Point it at your collector (OTEL_EXPORTER_OTLP_HEADERS adds auth headers):


npx supabase secrets set OTEL_EXPORTER_OTLP_ENDPOINT=https://your-collector/v1/traces

Spans carry report.id and project.id attributes so you can join them back to Mushi Mushi data.

Health dashboard

After deploying, open Mushi Mushi health (/health) in the admin console. It shows the AI calls behind diagnoses and fixes (count, error rate, fallback rate, latency and cost) and the background jobs’ last runs. Integration probes (Anthropic, OpenAI, Langfuse, Sentry, GitHub) are on the Integrations page.