SigNoz Observability

Two cooperating halves:

0. Preflight — is SigNoz MCP wired?

Confirm the tools exist before relying on them:

• Look for mcp__signoz__signoz_search_logs, ..._search_traces, ..._aggregate_logs in your tool list.
• If missing: tell the operator to add the SigNoz MCP server (see README.md → SigNoz MCP setup, mirrors the aws-mcp-setup skill pattern). Do NOT fabricate results — say it's not wired and stop.
• Credentials: the key lives in ~/.aish/credentials [profile:signoz] as SIGNOZ_API_KEY, referenced as ${profile:signoz}. Never read or echo it.

1. Instrumentation scan (on repo detect)

When a new repo is opened (or the operator asks), build/refresh its profile:

1. Read the current profile from state/registry.json keyed by the canonical repo root (git rev-parse --show-toplevel).
2. If stale/absent, the bin/scan-repo.sh hook will have populated it — read it back. To do a richer scan yourself, grep for: @opentelemetry/* (Node), opentelemetry-sdk|opentelemetry-instrumentation (Python requirements.txt/pyproject.toml), opentelemetry|tracing-opentelemetry|opentelemetry-otlp (Rust Cargo.toml), go.opentelemetry.io/otel (Go go.mod), io.opentelemetry (Java pom/gradle); plus OTEL_SERVICE_NAME, :4317/:4318, collector URLs.
3. Report a table: language(s) · service name(s) · instrumented? · exporter endpoint(s) · markers.

2. Exception check (post-turn / on demand)

For each service in the active repo's profile (state/registry.json → repos[<root>].services):

1. mcp__signoz__signoz_search_logs with service=<svc>, severity="ERROR", timeRange="30s" (or the operator's window). Use mcp__signoz__signoz_aggregate_logs grouped by service.name + error type for counts.
2. Dedup: skip any (service, error-fingerprint) already noted in state/signoz/seen.ndjson within dedup_ttl_secs (default 300).
3. If nothing fresh → say "clean" and stop. If hits → emit a compact summary: service · count · error type · 1-line stack snippet.
4. Surface per surface config:
   • statusline/both → write the one-liner to state/signoz/exceptions.txt (the statusline segment renders it).
   • memory/both → remember() a durable note tagged signoz,exception with the service, type, and first-seen timestamp so the pattern survives compaction.

3. Deeper triage (optional)

If the operator wants root cause, escalate to the bundled MCP skills:

• debug_service_errors {service, timeRange} — error logs + error-trace aggregation + top ops.
• latency_analysis {service} / incident_triage {alertId} for latency or alert-driven digs.

Guardrails

• Hooks are observe-only and MUST stay silent on stdout except the timer's cache line — never let a poll block or spam the prompt.
• Never invent exception data. If a SigNoz query errors or times out, report the failure and the query you ran.
• Keep every poll inside its timeout budget; a slow SigNoz must degrade to "unknown", not hang the turn.