Debug and Resolve Code Issues

Debugging Toolkit Smart Debug is an AI-assisted debugging specialist skill covering modern debugging tools, observability platforms, and automated root cause analysis. It runs a ten-step workflow: (1) Initial Triage using the Task tool's debugger subagent for error-pattern recognition, stack-trace analysis, dependency analysis, severity assessment, and 3-5 ranked hypotheses; (2) Observability Data Collection from error trackers (Sentry, Rollbar, Bugsnag), APM tools (DataDog, New Relic, Dynatrace), distributed tracing (Jaeger, Zipkin, Honeycomb), log aggregation (ELK, Splunk, Loki), and session replay (LogRocket, FullStory); (3) Hypothesis Generation scoring each hypothesis 0-100% with supporting evidence, falsification criteria, and expected symptoms, across categories like race conditions, stale cache, integration failures, resource exhaustion, configuration drift, and data corruption; (4) Strategy Selection among Interactive (VS Code/Chrome DevTools), Observability-Driven (Sentry/DataDog/Honeycomb), Time-Travel (rr/Redux DevTools), Chaos Engineering (Chaos Monkey/Gremlin), and Statistical/delta debugging; (5) Intelligent Instrumentation suggesting breakpoint/logpoint placement at entry points, decision nodes, state mutations, and integration boundaries; (6) Production-Safe Techniques - OpenTelemetry dynamic instrumentation, feature-flagged debug logging, sampling-based profiling (Pyroscope), read-only debug endpoints, and canary traffic shifting; (7) Root Cause Analysis via full execution-path reconstruction, variable-state tracking, and similar-bug-pattern identification; (8) Fix Implementation with code changes, impact assessment, risk level, and rollback strategy; (9) Validation via test suite run, performance comparison, and canary deployment monitoring; and (10) Prevention through AI-generated regression tests, knowledge-base updates, new monitoring/alerts, and runbook documentation.

Use this skill when working on debugging toolkit smart-debug tasks or needing guidance, best practices, or checklists for AI-assisted debugging - particularly for production or staging issues that benefit from observability-data correlation and structured hypothesis testing. It is not for tasks unrelated to debugging or for work in a different domain or tool outside this scope.

Input is the issue description passed as $ARGUMENTS, parsed for error messages/stack traces, reproduction steps, affected components, performance characteristics, environment, and failure pattern (intermittent vs consistent). Output is a structured report: Issue Summary (error, frequency, impact), Root Cause (detailed diagnosis with evidence), Fix Proposal (code changes, risk, impact), Validation Plan, and Prevention (tests, monitoring, documentation). A worked example walks a "checkout timeout errors" case from initial AI analysis through observability queries, trace-based hypothesis confirmation of an N+1 query pattern, instrumentation, canary rollout, and a validated fix that cut latency 70% and query count from 15 to 1:

const sentryData = await getSentryIssue("CHECKOUT_TIMEOUT");
const ddTraces = await getDataDogTraces({
  service: "checkout",
  operation: "process_payment",
  duration: ">5000ms"
});

Names a wide observability and debugging toolchain: Sentry, Rollbar, Bugsnag for error tracking; DataDog, New Relic, Dynatrace for APM; Jaeger, Zipkin, Honeycomb for distributed tracing; ELK, Splunk, Loki for log aggregation; LogRocket, FullStory for session replay; VS Code and Chrome DevTools for interactive debugging; rr and Redux DevTools for time-travel debugging; Chaos Monkey and Gremlin for chaos engineering; OpenTelemetry for dynamic instrumentation spans; and Pyroscope for sampling-based continuous profiling. Initial triage delegates to the Task tool's debugger subagent.

Engineers debugging production or staging incidents - especially intermittent or hard-to-reproduce failures - who want an AI-assisted, observability-driven workflow that generates ranked, falsifiable hypotheses and production-safe instrumentation strategies rather than guessing from a stack trace alone.