Fetch and summarize upcoming unreleased Codex features using a durable local clone synced from GitHub, with source-file mining as primary evidence. Use when asked for latest upcoming/openai-codex features, what is coming next but not in the latest stable release, or a live release-gap summary with links and as-of timestamp.

Mitigate incidental code complexity when control flow is tangled, nesting is deep, names are hard to parse, or reasoning requires cross-file hops. Use when a review stalls on readability, you need an analysis-first refactor plan before edits, or you want essential-vs-incidental verdicts, ranked simplification steps, a small structural sketch, and a TRACE assessment (analysis-only; no edits).

Lateral-thinking playbook that always returns a five-tier strategy portfolio (Quick Win through Moonshot). Use when you need options, alternatives, or trade-offs; when progress is stalled or failing repeatedly; or when you ask to think creatively, reframe constraints, and choose a strategic path before execution.

Generate Deckset markdown presentations from conversation context with high semantic fidelity. Use when requests mention decks, slides, presentations, speaker notes, or Deckset markdown output, especially when converting an existing conversation into a narrative slide flow. This skill refreshes upstream Deckset markdown docs and the configured gist examples on every run before drafting.

Finalize GitHub PRs end-to-end: update branch/PR, monitor CI until green, squash-merge, and clean up local/remote state. Use when asked to $fin or to finish/land/merge/close a PR, watch checks or runs, squash-merge, delete the branch, and sync local state.

Create a language-agnostic ghost package (spec + portable tests) from an existing repo by extracting SPEC.md, exhaustive tests.yaml (operations and/or scenarios), INSTALL.md, README.md, VERIFY.md, and upstream LICENSE files with provenance and regeneration instructions. Use when prompts say "$ghost", "ghostify this repo", "spec-ify/spec-package this library", "ghost library", or ask to extract portable spec/tests for libraries or tool-using agent loops (scenario testing); do not use for implementation work or editing skills.

Clarify ambiguous or conflicting requests by researching first, then asking only judgment calls. Use when prompts say "$grill-me"/"grill me", ask hard questions, request relentless interrogation, pressure-test assumptions, clarify scope/requirements, define success criteria, or request system-design/optimization decisions before implementation; stop before implementation.

Generate provider-agnostic AI agent guardrail blueprints and control matrices from a use case. Use when designing or reviewing agent safety architecture, prompt-injection and tool-misuse defenses, risk-tiered human approval gates, or auditable enterprise guardrail policies using industry patterns across top providers.

Turn 'should never happen' into 'cannot happen' by defining owned inductive invariants and enforcing them at parse/construct/API/DB/lock/txn boundaries with a verification signal. Use when prompts mention invariants, impossible states, validation sprawl, cache/index drift, idempotency/versioning, retries/duplicates/out-of-order events, race/linearization bugs, loop correctness, or hardening another workflow (for example $fix) with invariant checks first.

PR autopilot via `gh` only: create/manage PRs, keep branches current, enforce required CI gates, apply surgical code patches, and publish merge-ready handoff without merging. Use when asked to run or monitor PR automation, fix failing required checks, keep local/remote branch state clean, or prepare branch/PR cleanup for human merge.

Capture, browse, and query evidence-backed execution learnings in `.learnings.jsonl`. Trigger cues/keywords: `$learnings`, browse/recent/search learnings, lessons learned, takeaways, wrap up, handoff, before commit/PR, after tests pass, fail-to-pass, strategy pivot, footgun, retry loop.

Comprehensive, measurement-driven performance optimization for latency, throughput, memory/GC, and tail behavior. Use when the user asks to optimize/speed up, reduce latency (p95/p99), increase throughput/QPS, lower CPU/memory/allocations/GC pauses, profile hot paths, or run a benchmarked perf pass (including JSONL/query-heavy code). Requires before/after measurement on a runnable workload (or an explicit `UNMEASURED` plan) plus a correctness gate. Zig-only CLI iteration where `bench_stats`/`perf_report` are proven before shipping.

Expert analysis and derivation for delimited continuations and defunctionalization, using CPS, shift/reset, prompt/control, abstract machines, effect handlers, and continuation runtime strategy as supporting lenses. Use when Codex needs to compare control operators, explain a witness program, derive CPS or abstract machines, first-orderize higher-order semantics, plan continuation implementations or benchmarks, explain typed control tradeoffs, or build source-backed study and research roadmaps across Racket, OCaml, Haskell, Scala, JavaScript, or language-agnostic settings.

Precision copy edits + naming: compress and clarify without changing meaning. Triggers: `$logophile`, rewrite/reword/rephrase, shorten/tighten/make concise, choose final wording/phrasing, rename titles/labels/headings/skill names. In operational turns, use only when wording or naming output is explicitly requested.

Use `$mesh` for homogeneous leaf-batch execution over `spawn_agents_on_csv`: once planning has shaped repeated independent units, prefer one substantive row per unit with structured results and explicit concurrency.

Create, update, or refactor Codex skills in this repo, including SKILL.md, frontmatter trigger text, agents/openai.yaml, and scripts/references/assets. Use when asked to create a skill or make direct in-place edits to a skill now; do not use for read-only analysis or "should we update this?" turns, and treat $refine as the higher-level entrypoint for evidence-driven refinement of an existing skill.

Analyze a local codebase and infer the architecture it is actually using, including repo kind, best-fit dominant architecture, directly evidenced coexisting patterns, major subsystem exceptions, confidence, docs-vs-code drift, and repo-fit hints for downstream agents. Use when prompts ask what architecture a repo uses, whether it is really hexagonal or just layered, what subtle patterns shape a repo or slice, how a target slice fits a hybrid monorepo, whether documented architecture matches implementation, or when `$tk` needs a repo-dialect preflight before making a minimal change.

Gauntlet for absolute claims (always/never/guaranteed/optimal); pressure-test, then refine with explicit boundaries. Use when users ask to prove or disprove strong certainty claims, request devil's-advocate challenge rounds, or want the $prove-it gauntlet to run all 10 rounds continuously by default.

Launch and manage Codex Cloud tasks from the CLI, including detached background watchers that track completion. Use when users ask to run coding work in cloud/background agents, queue multiple cloud tasks, poll task status, fetch cloud diffs, apply cloud outputs locally, or pair cloud kickoff with `$cas` orchestration.

Continuously evaluate and improve AGENTS.md-style harness instructions through explicit-trigger OpenCode loops with an explicit model. Use when you want recurring harness reliability runs, especially for Gemini 2.5 Pro/OpenCode harness tuning, clean-repo eval cycles, curated exact-output probes, automatic eval-branch commits and PR updates for passing harness/doc changes, and external-blocker detection or regression auto-revert without scheduler/cron automation.