Hunt bugs with four-agent investigation swarm

This skill investigates a bug using four read-only sub-agents launched in parallel, then has the main agent rank the likely causes and recommend the fastest path to prove or fix the issue - explicitly diagnosis-first, never editing files or implementing fixes itself. It first builds a bug packet covering symptom, expected versus actual behavior, reproduction steps, scope of impact, and evidence like logs, stack traces, failing tests, screenshots, or diffs, preferring the user's own description, then explicit provided evidence, then recent git history if the bug looks regression-like, then the smallest relevant code path - inferring a minimal problem statement and naming what's still unknown if the report is underspecified, and reading the closest project instructions and docs for the affected subsystem first. It then writes a short investigation brief bounding what's broken, unproven, most likely involved, already evidenced, and what would count as confirmation, using read-only tools such as ripgrep and git diff, log, and show, reading logs, traces, and config, and existing test runs, without editing files or adding instrumentation.

Four parallel read-only investigators each get the same bug packet and brief, are explicitly told they're read-only with no edits, patches, staging, or commits allowed, and must return only a hypothesis, supporting evidence, missing evidence, smallest proof step, and confidence level, avoiding generic code-quality feedback or speculation without evidence. One investigator clarifies the exact failure shape and boundary conditions, such as the narrowest trigger and whether the bug is deterministic or flaky; a second traces the likely execution path and failure seam, checking state transitions, caller and callee mismatches, and data or control-flow breaks; a third investigates recent changes and regressions, checking correlated diffs, configuration or schema drift, and partial multi-entry-point updates; a fourth determines the fastest proof plan, identifying the smallest failing test, the most useful logs or traces, and a minimal non-mutating confirming command. For tiny, obvious issues, investigating locally instead of launching the full swarm is acceptable.

The main agent then synthesizes the results: merging duplicate hypotheses, discarding weak speculation, preferring evidence over elegance, separating root causes from mere contributing factors, and normalizing survivors into a hypothesis, evidence, missing-evidence, proof-step, and confidence shape - saying so directly if the evidence is too weak for a real ranking rather than forcing one. The final output orders the most likely root cause, plausible alternate causes, the fastest proof step, a recommended fix path, and open questions or blockers, optionally grouped into prove-now, fix-next, and follow-up-later buckets, explicitly without implementing any fix.

Use it to investigate bugs, regressions, crashes, flaky behavior, or unexplained failures that need a ranked root-cause diagnosis - not for actually implementing the fix, which is out of scope for this skill by design.

Given a bug description and available evidence, it produces a synthesized, ranked set of root-cause hypotheses with supporting and missing evidence, a recommended smallest proof step, and a suggested fix path - no code changes are made.

Read-only shell and git tooling such as ripgrep and git diff, log, and show, existing test runs, and the sub-agent orchestration mechanism used to launch four parallel investigators in reviewer or explorer roles.

Developers debugging a non-obvious failure who want fast, evidence-backed root-cause hypotheses from multiple investigation angles before committing to a fix, rather than guessing from a single vantage point.