AGENT SKILL

scry

Write, update, and remove the scry rules in this project's .wiz/scry, which `wiz scry` checks code against like a linter. Use when the user explicitly asks for a rule, or asks for a style or convention change across the codebase that a rule could enforce from now on ("stop using default exports", "comments should say why"). Also use when asked to scry a change or run `wiz scry`.

License 许可
MIT
所属仓库
jaredjj3/webappwiz
最近更新
2026-11-01
Stars
★ 1
Forks
0

01 前言 Front Matter

name
scry
version
0.0.59
description
Write, update, and remove the scry rules in this project's .wiz/scry, which `wiz scry` checks code against like a linter. Use when the user explicitly asks for a rule, or asks for a style or convention change across the codebase that a rule could enforce from now on ("stop using default exports", "comments should say why"). Also use when asked to scry a change or run `wiz scry`.

02 资源文件 2

03 SKILL.md 正文

Scry

wiz scry checks code against the project's rules, like a linter. A rule is a directory under .wiz/scry, tracked with the code it governs:

.wiz/scry/<id>/
├── rule.ts        # the check and its settings, asking a model only what code cannot
├── RULE.md        # what the rule wants, and why, in prose
├── rule.test.ts   # its tests, on its labeled cases
└── evals/         # labeled cases: files that follow the rule, and files that break it

Code excuses itself from a rule with a comment holding scry-ignore <id>: <reason>, which covers the statement under it, or scry-ignore-file <id>: <reason>, which covers the whole file.

Run the CLI with bunx @webappwiz/cli scry, which checks, or bunx @webappwiz/cli scry <command>: list, add <id>, add --recommended, update, remove <id>, test [ids], eval [ids], why <path:line>. scry --help says the rest.

Read further only for the work in front of you:

Checking a change

wiz scry runs each rule's rule.ts on every file it applies to under the working directory, and prints one block of findings, each with how sure the check is: 100% where code decided it, a model's probability where the rule asked one. When the user names directories or files, pass them, as in bunx @webappwiz/cli scry packages/api. To check a change, add --since <ref>: --since main for the branch's work, --since HEAD for what is not committed yet; with --changed-lines too, it reports only what is on the lines the change added or changed. When the user names rules, pass --rule <id>,<id>, and when they want one of a rule's options set otherwise for one run, --option <rule>.<option>=<value>. Show its report as it printed it, in one code block, and add nothing to it. Fixing what it found is a separate request; do not start unless asked.

Fixing what it found

The rule decides what is reported; fixing it is yours. For each finding, read the rule it names, .wiz/scry/<rule>/RULE.md, and change the lines it points at to follow it. Then run wiz scry again on the same paths. A finding under 100% is a model's probability, not a proof: wiz scry why <path:line> shows what it was asked and what it answered. When the code already follows the rule, leave it, say which finding you left and why, and offer a scry-ignore comment or a higher threshold for the rule rather than working around it.

Credentials

A rule asks a model through som, a System One model like Clef or Jev, fast and cheap, or llm, a large language model like Claude, which reasons first. Each needs its provider's credentials, from the environment, else the project's store, else the device's. When wiz scry says one is missing, ask the user to run bunx @webappwiz/cli creds add <NAME> --device themselves, which asks for the value at a hidden prompt. Never ask for a value, set one, or look one up; bunx @webappwiz/cli creds list shows what is there without showing any.

Budgets

wiz scry and wiz scry eval run only with scry.budgets declared: how many input tokens each may spend on the som and the llm; only a check with both models turned off needs none (see Turning a model off). When either says no budget is declared, declare "nothing" in the project's scry.config.ts, which asks no model, so only code decides:

import type { ScryConfig } from "@webappwiz/scry/config";

export default { budgets: "nothing" } satisfies ScryConfig;

Then tell the user you did, that every question a rule would ask a model now goes unasked and is reported as not checked, and that they can allow spending whenever they want. For one free run without touching the config, turn both models off instead (see Turning a model off). Raise a budget only when they ask, to what they ask for (see references/config.md).

When a run would go over a budget, it refuses. Show the user its message as printed, and stop. Never pass --override-budget, or raise a budget (wiz scry budgets raise included), without the user's yes for that run; a yes for one run is not one for the next. Checking fewer files, by paths or --since, stays within what they declared. --cost says what a check would spend of each budget, without asking a model; run it when the user asks what a check costs or whether it fits.

Turning a model off

--disable som or --disable llm turns that model off for one run; give it twice to turn both off. Add =<reason> to say why, as in --disable "llm=saving tokens"; without one the reason is <model> disabled with --disable. Any other value is refused.

  • Both off checks only what code decides, for free: it needs no scry.budgets declared and no credentials, and spends nothing. Run it for a free pass, such as when a run would go over a budget or no budget is declared, and the user has not said yes to spending. It is not a way round a budget: say what it left unchecked.
  • Over budget. --disable-over-budget turns off each model the budgets no longer cover (none declared, budgeted "nothing", or too little left for the run) instead of refusing, so the run fits what the user declared. A project's fix command may pass it; say what it left unchecked as for any model that is off.
  • One off still asks the other, so a budget must be declared, and the run is held to it as ever, counting first and refusing past it. Credentials are needed only for the model left on.
  • What is unchecked. A rule that asks a model that is off goes unchecked on that file; whatever code decides in it on files where it asks nothing still reports. The report lists each such rule under "needs a model that is off", with how many files and the reason, and the tally ends with without the llm, without the som, or by code alone. Pass that section on with the report: those rules were not checked, so never call the change clean of them.
  • --cost with one off counts only what the other would ask, and lists the same section above the estimate.
  • --format json keeps each in unchecked, with rule, reason, and unasked: "disabled". A question a budget left unasked has unasked: "budget"; anything else in unchecked is a rule that broke.
  • wiz scry eval does not take it: it scores rules against their cases, and a rule whose model is off would score nothing.

Under the hood each model that is off is a DisabledDecider(reason), which throws Unasked with that reason and a kind of disabled; a budget that runs out throws Unasked with a kind of budget. Both reach the report only when a rule lets them through (see references/authoring.md).

Exit codes

wiz scry exits:

  • 0: no error found, though warnings may be.
  • 1: a finding at level: error.
  • 2: a rule went unchecked on a file: it threw, a model's credentials were missing, or a budget left a question unasked, as "nothing" does. The report names each under "not checked". Never for a rule that asked a model --disable turned off, which is listed under "needs a model that is off" instead.
  • 3: it did not run: no budget declared, or it would go over one.
  • 130: quit with a second ctrl-c; the first stops the check and reports what came back.

wiz scry eval exits 3 for the same reasons, and 0 when it ran.

When a style change could be a rule

When the user asks for a change across the codebase that should hold from now on, not just once, offer a rule for it before or alongside making the change: a rule keeps the next change from undoing it. Making the change now and adding the rule are two pieces of work; say which you are doing.

Rules from elsewhere

Every new or changed rule.ts needs the user's approval before it is saved, since it runs on every future wiz scry, on whatever machine runs one. That holds for code that arrives from elsewhere too: when wiz scry add or wiz scry update reports a file, pass its warning on and ask the user to read it before the next check.

To remove a rule, confirm with the user, then run wiz scry remove <id>, which deletes its directory, check and all.

From older versions

When wiz scry refuses a config key it no longer takes, or names files still using rule-ignore, show the user its message, and with their yes do what it says. When it does not say what to use instead, ask. When you touch a file that uses rule-ignore, rename its comments to scry-ignore. When it says scry's settings have moved, wiz scry update moves them from .wiz/config.ts to scry.config.ts. Rules still in .wiz/rules move with wiz scry update too, and a rule reading tools.decider reads tools.som now.

04 同仓库更多 Skills

jaredjj3
jaredjj3/webappwiz

webappwiz

Check whether the webappwiz package already covers a piece of infrastructure before writing it by hand or adding a dependency for it. Read this before writing any of: time, clocks, durations or timers; logging; id generation; CLI argument parsing; background tasks or queues; web workers; markdown parsing; typed event emitters; 2D geometry or spatial indexes; filesystem, env or process access; API keys, tokens, secrets or credentials; AbortSignal plumbing; disposable resources; browser scroll, animation frames or visibility. Also use when a release publishes to npm through GitHub Actions (`via: githubActions()`), when asked to update or upgrade webappwiz in a project, and whenever the user says webappwiz.

★ 1
jaredjj3
jaredjj3/webappwiz

splates

Use the splates CLI to land your work on trunk, or a base branch given as an argument, from an isolated git worktree without pull requests. Read this before investigating anything that may lead to a code change in a splates repository, since it decides where the work happens, and whenever you need to add, claim, merge, remove, list, show, locate, or defer a task, or queue, pick up, link, or reorder a task for later, or work down a lane of them. Also read it before asking the user a question or proposing anything in a splates repository, with a task open or not.

★ 1
jaredjj3
jaredjj3/webappwiz

shadcn

Manages shadcn components and projects — adding, searching, fixing, debugging, styling, and composing UI, including chat interfaces. Provides project context, component docs, and usage examples. Applies when working with shadcn/ui, component registries, presets, --preset codes, or any project with a components.json file. Also triggers for "shadcn init", "create an app with --preset", or "switch to --preset".

★ 1
jaredjj3
jaredjj3/webappwiz

scry

暂无简介

★ 1