sai

Staff-engineer-level code review that goes beyond correctness to evaluate architectural alignment, system-level implications, failure modes, performance, scalability, backward compatibility, observability, security, and cross-team impact. Use when reviewing a PR (URL or diff), analyzing code changes for architectural fitness, or when the user asks for a thorough/staff-level/senior review of code changes. Triggers on "review this PR", "review these changes", "staff review", "thorough code review", sharing a GitHub PR URL for review, or asking about the architectural impact of changes.

Produce a daily AI news digest and reuse the source workflow under `claude/ai-daily-digest`. Use when the user wants a curated AI briefing or roundup in Codex.

List, reply to, resolve, unresolve, and create GitHub PR review comment threads using gh CLI scripts.

Clean stale local branches and worktrees by reusing the source workflow under `claude/git-clean-gone`. Use when merged or abandoned work leaves gone branches behind.

Stage selected git hunks non-interactively by reusing the source workflow under `claude/git-stage-hunk`. Use when only part of a file should be committed.

Review Go code using the source material under `claude/go-code-review`. Use when the user wants a focused Go review grounded in known failure patterns.

Rewrite text to sound less formulaic by reusing the source workflow under `claude/humanize`. Use when the user wants writing to sound more natural in Codex.

Draft and refine KubeCon CFPs by reusing the source workflow under `claude/kubecon-cfp`. Use when the user wants to assess or improve a KubeCon proposal in Codex.

Critique implementation plans by reusing the source workflow under `claude/plan-critic`. Use when the user wants a plan stress-tested before implementation.

Turn rough instructions into stronger prompts using the source workflow at `claude/promptgen`. Use when the user wants a task prompt, system prompt, coding-agent prompt, tool description, eval grader, subagent briefing, or prompt-improvement pass rather than direct execution.

Refactoring review through seven sourced refactoring personas (Fowler, Uncle Bob, Feathers, Beck, Metz, Ousterhout, Tornhill) plus an adversarial pass, by reusing the source workflow under `claude/refactor-council`. Use when the user wants a refactoring review of code, a module, an app, or a diff.

Audit `CLAUDE.md` files by reusing the source workflow under `claude/review-claude-md`. Use when the user wants a CLAUDE.md contract reviewed or fixed in Codex.

Diagnose flaky service-mesh tests and connectivity issues by reusing the source workflow under `claude/service-mesh-debug`. Use when the user is debugging mesh flakiness, xDS issues, or mTLS failures.

Review code changes with staff-level depth by reusing the source workflow under `claude/staff-code-review`. Use when the user wants a thorough review beyond line-level correctness.

Coach and tailor staff-level resumes by reusing the source workflow under `claude/staff-resume`. Use when the user wants a staff-oriented resume review or rewrite in Codex.

Write or review behavior-first tests by reusing the source workflow under `claude/test-writer`. Use when the user wants tests added or existing tests reviewed in Codex.

Produce a current AI news digest from a normal Copilot session when the user wants a briefing, roundup, or "what happened in AI" summary. The content depends on current sources, so browse and verify dates before drafting - never write the digest from memory. Not for timeless explanations of how a model or technique works.

Critique an implementation plan from a normal Copilot session when the user shares a plan (Plan Mode output, pasted plan text, or a plan file path) and wants it stress-tested before any code is written. Triages the plan, grounds it against the codebase, delegates to three native persona reviewers (Skeptic, Architect, Verifier), and returns an Approve/Reject/Refine verdict with concrete refinements. Not for executing the plan, reviewing an already-written diff, or a quick lint pass.

Audit and fix a CLAUDE.md against a structured tiered checklist from a normal Copilot session when the user asks to review, score, audit, or improve a CLAUDE.md. Scans the target repo, runs the bundled validation scripts, reports a PASS / NEEDS WORK / FAIL verdict with evidence, then fixes failing checks and re-evaluates against the rubric. Not for generic markdown cleanup, and not for reviewing Copilot or Codex skills/agents.

Run a staff-level code review from a normal Copilot session when the user asks to review a PR, a diff, or a set of changes (a GitHub PR URL, file paths, or a pasted diff). Triages the change, builds a research brief, delegates to seven native dimension reviewer agents plus a code adversary, synthesizes, then a findings adversary red-teams the result before it is returned or posted. Accepts the flag `--no-adversary` (skip both adversary passes). Do not use for a quick lint pass or a single-file style nit.

Generate a lean, high-signal CLAUDE.md for a repository from codebase analysis. Grounded in Anthropic best practices and empirical studies; built to pass review-claude-md (a bundled validator enforces the same critical checks). Use when the user asks to "generate CLAUDE.md", "create a CLAUDE.md", "write a CLAUDE.md", "init CLAUDE.md", "scaffold CLAUDE.md", or "bootstrap CLAUDE.md" for a project.

Generate a lean, high-signal `CLAUDE.md` for a repository by reusing the source workflow under `claude/generate-claude-md`. Use when the user wants a CLAUDE.md created or refreshed in Codex.

Use when the user wants a refactoring review of code, an app, a module, or a diff. Scans the target for code smells and git hotspots, reviews it through seven opposed refactoring lenses (Fowler, Uncle Bob, Feathers, Beck, Metz, Ousterhout, Tornhill), synthesizes a safety-first sequenced refactoring plan, then runs a separate adversarial agent that red-teams the plan before returning it. Triggers on "refactor this", "how should I refactor", "refactoring review", "what should I clean up here", "is this worth refactoring".

Run the refactoring council from a normal Copilot session when the user asks for a refactoring review of code, a module, an app, or a diff. Scans the target for code smells and git hotspots, reviews it through seven opposed refactoring lenses (Fowler, Uncle Bob, Feathers, Beck, Metz, Ousterhout, Tornhill), synthesizes a safety-first sequenced refactoring plan, then a separate adversary agent red-teams the plan before it is returned. Accept flags `--no-scan`, `--no-adversary`, `--since <value>`, `--personas <comma-list>`. Do not use for a quick lint pass or a standalone correctness/security/performance audit.

Use when the user invokes $council, $council:council, Council review, or Council debate. Use loaded SKILL body or one direct installed `skills/council/SKILL.md` read. Direct read path must contain `/.codex/plugins/cache/sai/council/` and end `/skills/council/SKILL.md`. `cd <cwd> && sed -n ... <path>` is valid. Do not use `pwd`, `ls`, `find`, `rg`, `cat`, multiple `&&`, or `;`. Never use repo-local paths. If unavailable, stop exactly `Council not run: skill unavailable.` At most one pre-tool message, exact `Council progress:` line only. Non-final lines start `Council progress:`.

Run the council workflow from a normal Copilot session only when the user explicitly asks for council review, multi-persona critique, debate, design review, code review, architecture feedback, UX review, or tradeoff analysis. Do not use it for commit, stage, merge, approval, or generic pre-commit requests. Accept the same mode syntax as the bundled council reviewers: `core|auto|core-eng|core-ux|core-mix|all|debate <problem|@file>`. During council slash-command use, the current session agent moderates reviewer agents directly. Runs broader than 6 reviewers require explicit AskUserQuestion approval before launch.

Use when the user asks for council review, multi-persona critique, debate, design review, code review, architecture feedback, UX review, or tradeoff analysis. Bare invocations use `core` profile auto-detect; explicit `auto` selects the best-fit 6 personas from the sourced 27-persona engineering and UX roster. Users can still pin `core`, `core-eng`, `core-ux`, `core-mix`, `all`, or `debate`.

Use when turning rough instructions into optimized, evidence-based AI prompts for system prompts, task prompts, coding-agent instructions, tools, eval graders, subagent briefings, or prompt-improvement work. Copies to clipboard.

Daily AI news digest covering technical advances, business news, and engineering impact. Aggregates from research papers, tech blogs, HN, newsletters. Use daily for staying current on AI developments.

Critique a Claude implementation plan before approving execution. Spawns parallel persona subagents (Skeptic, Architect, Verifier) to evaluate the plan against the codebase and return an Approve/Reject/Refine verdict with concrete refinements. Use when a plan has been generated (Plan Mode output, ExitPlanMode, pasted plan text, or plan file path) and needs review before code is written. Triggers on "critique this plan", "review this plan", "is this plan good", "poke holes in this plan", "should I approve this plan", or sharing a plan for evaluation.

Audit and fix CLAUDE.md files using a tiered binary checklist based on official Anthropic best practices and community guidelines. Use when the user asks to "review CLAUDE.md", "audit CLAUDE.md", "score CLAUDE.md", "improve CLAUDE.md", or "fix CLAUDE.md".

Diagnose and fix flaky e2e tests and general connectivity issues in service mesh environments (Kuma, Istio, Linkerd, Consul). Trigger when: a user mentions intermittent test failures, "test is flaky", e2e failures in CI that don't reproduce locally, Ginkgo/Gomega test files that fail sometimes, 503/connection refused errors, mTLS handshake failures, pods not getting traffic, xDS NACKs or warming resources, cert delivery timing issues, or "works locally but not in cluster". Covers Kuma flakiness patterns (timing races, xDS propagation delays, Envoy circuit breakers, mTLS readiness, Gomega misuse) AND universal mesh debugging (control plane connectivity, proxy lifecycle, certificate problems, traffic routing/policy, service discovery).

Write tests that verify behavior (not implementation), use table-driven/parameterized patterns, and minimize mocking. Triggers when asked to write tests, add test coverage, or create test files. Also triggers when reviewing existing tests for quality.

List, reply to, resolve, and create GitHub PR review comment threads using gh CLI scripts. Use when managing code review feedback, replying to reviewer remarks, resolving review conversations, creating reviews with line-level comments, or bulk-processing threads by author.

Auto-review Go code for 100+ common mistakes when analyzing .go files, discussing Go patterns, or reviewing PRs with Go code. Checks error handling, concurrency, interfaces, performance, testing, and stdlib usage.

Interactive KubeCon CFP submission writer. Guides through topic selection, title crafting, abstract writing, and benefits section using acceptance data from 1,100+ talks across 7 KubeCon events (2024-2025). Use when preparing a conference talk proposal for KubeCon/CloudNativeCon, writing a CFP, or asking about KubeCon submission strategy.

Build and refine staff-level engineering resumes through interactive coaching, research-backed best practices, and per-job tailoring. Use when building, improving, or tailoring a resume for Staff/Principal engineer roles.

Stage only specific hunks from the working tree for selective commits. Use when a file has changes from multiple sessions, parallel AI agents, or mixed user and agent edits and you need to commit only the changes made by the current session. Use when git add -p is unavailable or when you need non-interactive partial staging. Lists hunks with IDs, stages by hunk ID, file, regex pattern, or line range.

Identify and remove AI writing patterns to make text sound natural and human-written. Use when humanizing commit messages, PR descriptions, review comments, docs, changelogs, or release notes. Also for de-slopping text that sounds robotic, has AI vibes, or reads like ChatGPT output.

Clean up local branches with deleted remote tracking and their worktrees. Use after merging PRs to remove stale branches, detect squash-merged and rebased branches, and clean up associated worktrees.