verification-closure

Use when implementing, reviewing, migrating, linting, testing, fuzzing, profiling, optimizing, or hardening Zig 0.16.0 code: .zig files, build.zig/build.zig.zon, std.Io/process.Init migration, C interop, expert comptime/metaprogramming/reflection/codegen, allocator ownership, FFI boundaries, concurrency, dependencies, and measured performance work.

Systematic UX evaluation using Nielsen heuristics and accessibility checks. Use when reviewing UI, "is this usable", improving user experience, or pre-launch.

Use when code smells point to a structural refactor that should ship: flag or state matrices, repeated boundary validation, shared-key agreement checks, branchy policy logic, or syntax mixed with execution. Default to one seam, one smallest honest construction, adapter-first staging, and one proof signal.

Cross-modal diagnostic and review workflow for software systems. Use this skill to understand, explain, compare, critique, debug, profile, review, or refactor code by mapping technical signals into sensory models, then translating those models back into precise engineering language. Best fits include architecture review, readability or maintainability assessment, strange or flaky behavior, performance bottlenecks, API or UX critique, onboarding explanations, and comparing implementations or designs by feel, friction, weight, rhythm, sharpness, smoothness, coupling, or complexity. Also use when prompts ask what a codebase, bug, logs, API, or system feels, sounds, or looks like, or ask to make it lighter, smoother, cleaner, tighter, quieter, or more coherent. Do not use for exact API syntax, compliance or legal interpretation, security sign-off, rote code edits, or terse factual tasks.

Manage persistent task plans in `.step/st-plan.jsonl`, with an explicit first-use choice between repo-committed storage and local-only ignore via `.git/info/exclude`, so state survives turns/sessions and can stay reviewable in git when desired. Use when users ask to "use $st", "resume the plan", "export/import plan state", "checkpoint milestones", "track dependencies/blocked work", "show ready next tasks", "keep shared TODO status on disk", "store backlog tasks on disk without loading them into `update_plan` yet", "select which durable tasks enter the mirrored plan", "map a `$select` plan into durable execution state", "prove `$st` works for implementation tracking", mirror the durable plan into Codex `update_plan` or OpenCode `TodoWrite`, or diagnose/repair `st-plan.jsonl` concerns (for example append-only vs mutable semantics, lock-file ignore policy, or seq/checkpoint integrity).

Shrink and unify code without changing behavior. Use when: simplify, refactor, reduce duplication, remove lines, extract helper, reuse component, DRY, collapse, better abstraction.

Finalize work after validation: confirm a signal, capture proof in the PR description, and open a PR (no merge). Use when asked to run `$ship`, ship/finalize a branch, prepare or open a PR without merging, or publish validation proof before handoff.

Mine Codex sessions JSONL (`~/.codex/sessions`) and file-based memories (`~/.codex/memories`) for explicit `$seq` and artifact-forensics questions, preferring `artifact-search`, then specialized follow-ups such as `skill-blocks`, `plan-search`, `session-prompts`, `session-tooling`, and `orchestration-concurrency`, then `query-diagnose`, and generic `query` only when needed. Opencode mining is explicit-only and requires a literal `opencode` cue in the request.

Discriminately adjudicate PR review comments before implementation. Treat each comment as a claim to test, build the strongest no-change countercase, recover PR rationale with explicit `$seq` when needed, and decide what to act on, rebut, defer, or investigate. Trigger for `$review-adjudication`, review the review, adjudicate PR comments, are these comments relevant, which comments matter, or should we act on these comments. Not for implementing fixes, writing rebuttals only, or final merge closure.

Refine an existing Codex skill in place with minimal diffs, then validate with quick_validate. Trigger when asked to improve a skill's trigger description/frontmatter, workflow text, metadata, scripts/references/assets, or agents/openai.yaml; also for requests to iterate, refactor, rename, or fix a skill using usage/session-mining evidence (for example from $seq).

Deconstruct high-cost abstractions and recommend lower-level primitives that reduce tooling, indirection, and hidden steps. Use when requests ask for fewer layers (for example "too many layers", "remove this framework/plugin/DI", "ditch codegen/task runners", "replace with plain scripts/config/SQL"), or when prompts say "it feels over-engineered", "start simpler", or "reduce the size of the codebase" (analysis-only unless implementation is explicitly requested).

Autonomous multi-turn proof/disproof gauntlet for absolute claims. Default Auto Gauntlet runs exactly one numbered round per assistant turn and auto-continues the same conversation until a terminal proof/disproof certificate or round 10 Oracle synthesis.