Use this skill when the user asks to audit, review, find bugs, security-check, or validate the correctness of existing code. Runs a subagent-driven audit-then-fix loop — first subagent (auditor) finds P1/P2/P3/low findings, subsequent subagents (one per blocking finding) each return a changeSet fix, applied atomically with test-gated rollback, until only low findings remain or iteration cap hit. Default 7 iterations. Not for polishing working code (use xo-polish), not for perf (use xo-improve).
Use this skill when the user asks to research, explore alternatives, propose a different implementation, question the approach, or find a better way to do something. Runs a subagent-driven Champion-vs-Challenger tournament — current implementation is Champion; each round spawns a "Challenger" subagent proposing a radically different approach. Both must pass tests; a judge subagent picks the winner on a 1-5 rubric. Converges when Champion wins twice in a row. Default 5 rounds. Not for polishing within the existing paradigm (use xo-polish).
Use this skill when the user asks to benchmark, measure, profile, or lock deterministic output of code via SHA-256-hashed expected results. Runs entry-point command, captures stdout, parses JSON output, matches against expected. This mode is deterministic — no subagents or API required. Not for iterative improvement toward a benchmark (use xo-improve).
Use this skill when the user asks to build, create, implement, scaffold, or write a NEW feature from scratch. Runs serialized TDD via TWO subagents — Agent A (Spec Writer) writes failing tests from acceptance criteria, Agent B (Builder) writes implementation that makes tests pass. Fixed 2-iteration sequence (spec → implementation), red→green validated between them. Not for modifying existing code (use xo-polish), not for fixing bugs (use xo-audit).
Use this skill when the user asks to generate, update, refresh, or sync documentation for the repo. Scans public API surface + tests + existing docs, proposes JSDoc/docstrings/README/CHANGELOG updates, then runs two polish passes for tightness and link/example correctness. NOT iterative — fixed 3-round pipeline (generate + polish + polish). Can only edit docblocks in source files, never logic; can create/update doc files freely. Not for writing new code (use xo-build), not for refining source logic (use xo-polish).
Use this skill when the user asks to finalize, split, squash, or turn a noisy xoloop run into clean reviewable branches. Reads .xoloop/session.jsonl, groups kept proposals by non-overlapping file sets, and creates one independent branch per group starting from the merge-base. Each branch is a standalone reviewable PR. Not for running the loop itself (use xo-polish / xo-audit / etc.).
Use this skill when the user asks to fuzz, stress-test, property-test, or find crashes in existing code. Runs fast-check property-based fuzzing on module exports. This mode is deterministic — no subagents or API required. Distinguishes AdapterErrors (expected rejections) from real bugs (TypeError/RangeError/etc). Writes crash corpus for CI replay.
Use this skill when the user asks to improve code AGAINST A SPECIFIC BENCHMARK — make it faster, reduce memory, lower cost, hit a target metric. Runs a subagent-driven iteration loop — each round a subagent proposes an optimization, the benchmark runs (not the test suite), keeps if metric improves without regressing correctness. Default 7 iterations. Not for open-ended refinement (use xo-polish), not for finding bugs (use xo-audit).