skills

Perform a deep, scoped review of a working tree, staged diff, commit range, branch diff, pull request, focused file set, or pasted code. Use for `/code-review`, PR or diff review, and merge-safety assessment. Begin with a read-only orchestration-assessment subagent that decides whether one reviewer or multiple specialist subagents are justified by the change scope and risk. Prioritize correctness, regressions, security, privacy, contracts, data, concurrency, migrations, and tests. Persist the canonical Markdown review report and do not edit code or Git state unless the user separately requests fixes.

Consume a code-review Markdown report, PR feedback, or equivalent review comments; re-verify every finding, question, test gap, and uncovered area; formally challenge incorrect or stale review claims with arguments and evidence; and implement confirmed actions through a coding subagent. Use after code review when Codex must decide what still applies, produce a complete disposition ledger, fix proven issues, close coverage gaps, verify changes, and persist the canonical receiving-code-review resolution report. Begin with a re-review assessment subagent that decides whether multiple specialist subagents are warranted by both the change scope and review results. Preserve the user's Git index unless staging or publishing is explicitly requested.

Evidence-first runtime debugging that maximizes the probability of identifying an application bug's root cause in a single reproduction. Use for bugs, regressions, flaky or timing-sensitive failures, expensive or hard-to-repeat reproductions, and requests such as "instrument broadly", "consider every plausible cause", "capture more breakpoints", or "find the root cause in one run". Build a broad but deduplicated causal hypothesis set, map hypotheses to structured high-information probes across boundaries, branches, state transitions, asynchronous work, caches, and external calls, collect correlated NDJSON evidence with the bundled local collector, summarize large logs without loading the full file into context, prove the causal chain, apply the smallest fix, verify with before-and-after evidence, and remove temporary instrumentation.

Use and maintain the `mr` Node CLI for generic Git MR/PR branch workflows. Use when the user asks to create, preview, configure, troubleshoot, install, update, uninstall, or explain MR/PR flows with `mr`, `mrm`, `mrt`, or `mrp`; when handling branches named like `mr/target/current`, strategy flags `--merge`, `--rebase`, `--merge-target`, `--pr`, default detached mode (`--detached`, `--no-detached`, `MR_DETACHED`, `mr.detached`), detached conflict worktree placement (`MR_WORKTREE_DIR`, `mr.worktreeDir`), request providers or commands (`MR_REQUEST_PROVIDER`, `mr.requestProvider`, `MR_REQUEST_COMMAND`, `mr.requestCommand`, CNB/GitHub/GitLab), automatic update notices (`MR_NO_UPDATE_CHECK`, `NO_UPDATE_NOTIFIER`), conflict resume, `--rm-mr`, `--dry-run`, diagnostics flags, or the upstream `JUNERDD/mr` implementation behind this CLI.

Coordinate non-trivial multi-step work with delegation-first subagent scheduling. Use when work can benefit from parallel exploration, implementation, review, verification, queued independent requests, large repositories or monorepos, migrations, dirty worktrees, isolated worktrees or branches; when the user asks to maximize subagent delegation or keep running workers uninterrupted; or when auditing multi-agent orchestration. After triggering, build a dependency graph, dispatch the maximum useful set of ready non-overlapping workers allowed by tools, permissions, isolation, and integration capacity, keep healthy workers running to completion, synthesize evidence, and verify the integrated result. Handle only truly trivial or non-delegable work directly.

Opt-in Cursor SDK delegation workflow for bounded coding tasks, reviewed implementation packets, Cursor internal subagents, hierarchical workstreams, follow-up packets, live Cursor SDK monitoring, and proactive Cursor API-key authorization. Use only when the user explicitly injects or names `$delegate-to-cursor-sdk` / `delegate-to-cursor-sdk`; otherwise do not select it proactively.

Deeply pressure-test a plan or design by asking one question at a time until assumptions, tradeoffs, risks, failure modes, and scope edges are explicit. Use when the user asks to be grilled, wants a plan or design stress-tested, or needs the live conversation kept in sync with a local Markdown Q&A log.

Consume a `thermo-review` Markdown report, PR feedback derived from one, or structural quality-gate feedback involving `Blocker`, `Major`, `Minor`, `Question`, `Complete Findings Index`, `Decomposition Gaps`, recursive coverage, line-count ledgers, candidate sweep logs, `Not covered` rows, or 350-line thresholds. Use when Codex must verify every structural item against the current checkout, build a disposition and behavior-parity ledger before editing, fix, disprove, narrow, waive, or carry forward each item with evidence, avoid user-visible regressions during structural cleanup, and leave Git staging untouched unless explicitly asked to stage, commit, or publish.

Perform an extremely strict, report-writing code quality review focused on structural simplification, abstraction quality, file-size pressure, spaghetti branching, canonical ownership, type boundaries, and exhaustive recursive candidate sweeps. Use for thermo-nuclear code quality review, thermonuclear review, harsh maintainability review, deep code quality audit, structural quality gate, or when asked whether a change is too complex or should be restructured. Writes a Markdown report and does not make code changes unless explicitly asked for fixes.

Decompose a user request into deliverables, workflow phases, tools, domains, and implicit prerequisites, then find and select applicable local skills from Cursor, Claude Code, OpenCode, Codex, and shared Agent Skills roots before analyzing or handling the request. Use when the user asks to find, choose, inspect, route, or reason about available local skills; when a request explicitly says to check all local/available skills first; or when doing skill-aware requirement analysis, planning, or capability routing where deeper request understanding is needed before skill selection.

Consume a coverage-led `regression-review` Markdown report or related PR feedback and turn it into evidence-backed next actions. Use when Codex is given `Block`, `Discuss`, `Watch`, `Intentional Changes`, `Complete Findings Index`, `Coverage Ledger`, `Behavior Graph Deltas`, `Not covered`, or coverage-gap items from a regression-review report and must verify whether each user-visible finding still applies, fix proven regressions, disprove or challenge stale findings, confirm intentional product changes, close coverage gaps, and report a disposition for every item before changing or claiming completion while leaving Git staging untouched unless the user explicitly asks for staging, committing, or PR publication in the current request.

Perform a scoped, coverage-led review of working tree, staged, commit-range, branch, or PR changes to find user-visible behavioral regressions. Use when Codex must audit code changes for broken or degraded user journeys, changed defaults, loading/error/permission/session behavior, stale data, ordering, retries, duplicate/destructive actions, exported output, emails, CLI output, or other visible behavior, and must write a Markdown report that enumerates all distinct findings discovered within the reviewed scope plus coverage gaps, intentional visible changes, and scoped behavior-graph deltas when they clarify the affected path.

Cursor-style Plan Mode for complex, ambiguous, risky, or multi-file work before editing or executing. Use when the user asks for a plan, says to stay in plan mode, asks to design an implementation, asks to write or save a project-local plan file, needs architecture or tradeoff analysis, or when a task requires clarification, codebase research, route/data-flow tracing, broad exploration, or approval before changes. This skill creates or updates a disk-backed editable Markdown plan file by default, researches the codebase, asks clarifying questions, records file/code references and todos, invokes `$grill-me` for non-trivial pressure-testing, then builds only from the approved plan.

Draft a Conventional Commit message from the currently staged Git changes. Use when the user wants a commit message suggestion, asks to summarize staged work into a commit, or needs a Conventional Commit subject/body without actually running `git commit`. This skill must only inspect staged changes and must not stage files, inspect unstaged work, or create the commit.

Identify, prevent, and remediate 重复造轮子 across code, libraries, services, templates, docs, platform workflows, and architecture decisions. Use when asked to audit duplicated implementations, search for existing reusable assets before building, decide build-vs-reuse/buy, consolidate similar components/tools/APIs, create reuse catalogs, write ADR/RFC/migration plans, establish golden paths/paved roads, or improve discoverability, ownership, and governance for reusable assets.

Exhaustive code slimming, code-pruning, and approval-gated architecture cleanup workflow for maximizing removable code while preserving behavior and improving developer experience. Use when the user asks for code cleanup, ruthless simplification, dead-code removal, dependency pruning, refactoring for less code, DX architecture review, exhaustive shrinking, maximal code reduction, redundancy removal, or aggressive but behavior-preserving simplification.

Consume a coverage-led `hack-review` Markdown report, PR feedback derived from one, or a request to address `Block`, `Discuss`, `Watch`, `Intentional Exceptions`, `Complete Hack-Risk Index`, `Ownership Coverage Ledger`, `Not covered`, or coverage-gap items. Use when Codex must verify whether each reported shortcut, ownership problem, intentional exception, and uncovered implementation boundary still applies before changing code, then fix, disprove, narrow, confirm, or carry forward every item with evidence while leaving Git staging untouched unless the user explicitly asks for staging, committing, or PR publication in the current request.

Add or rewrite code comments with calibrated granularity so they explain intent, constraints, data meaning, contracts, and control-flow decisions instead of translating syntax. Use when Codex is asked to document existing code, improve low-value comments, add structured comments to functions, interfaces, classes, types, fields, exported constants, configuration objects, state transitions, or complex internal logic while preserving the repository's existing comment language and style.

Perform a scoped, coverage-led review of a working tree, staged diff, commit range, branch diff, PR, or suspicious implementation to find hack-like implementation risks. Use when Codex must audit brittle shortcuts such as impossible-state fallbacks, masked root causes, duplicate abstractions, hardcoded special cases, boundary bypasses, hidden temporal coupling, write-then-fix-up flows, or other ownership problems, and must write a Markdown report that enumerates all distinct hack-risk findings discovered within the reviewed scope plus coverage gaps and intentional exceptions.

Split a large or mixed Git working tree into multiple focused local commits. Use when changes span unrelated concerns, combine refactors with behavior changes, mix generated files with source edits, or contain separable hunks in the same file, and Codex needs to stage one logical batch at a time, invoke `$git-commit` to draft the staged commit message, present the staged batch for user confirmation, run `git commit` only after the user explicitly approves that batch, and repeat without pushing.