JawsomeJason/jawsome-agents

Standards

Reviewers enforce the same conventions by hand, comment after comment. This skill mines those comments into written standards, then checks diffs against them. Each standard pairs a rule an agent follows when it writes code with a lens a reviewer applies to a diff. Each numbered § cites the review comments behind it, so anyone can trace a rule back to the reviewers who enforced it.

Core principles

• Evidence first. A § needs review comments from at least 3 distinct changes and 2 distinct reviewers, because many comments on one change, or from one person, are a single reviewer's taste.
• Every rule carries its why. A reason still guides an agent after the code around the rule changes shape.
• One home per convention. Two copies drift apart, and an agent follows whichever one it read first.
• A check beats a §. A convention a linter, type or test can catch belongs in that tool, because a check runs every time.
• Delete what agents already do. A rule followed unprompted costs context on every load and changes nothing.
• Scripts are shortcuts. Each output file has a schema in references/schemas/, so when a script is missing or fails, write the same file by hand.
• git is the only hard dependency. A forge's CLI, such as gh or glab, adds review comments, and git log stands in without one.

Commands

Command	What it does	Output	Reference
mine	Pulls review comments, labels each by convention, ranks the clusters by frequency times severity	.standards/clusters.json	references/mine.md
write	Turns ranked clusters into a standards skill in the repo, each § with its why and evidence	the standards skill	references/write.md
review	Applies every lens a diff triggers and reports findings with the § cited	findings in the reply	references/review.md
prune	A/B tests each rule across harnesses and deletes the ones agents follow unprompted	prune.json	references/prune.md
doctor	Flags drift: globs that match nothing, cited files that are gone, lenses that never fire, stale prune data	.standards/doctor.json	references/doctor.md

Routing

Invoke as standards [command] [scope]. A scope is a path or glob, an area slug from the generated index, or a § number (§2, §2.3). Resolve it per references/scope.md. mine takes only a path, because areas exist only after write.

• No argument. Show the menu below and wait for a choice. Never auto-run a command, because each one reads history, edits the repo or spends many agent runs.
• A scope with no command. Run review of that scope without asking, because review only reads.
• An explicit or clearly implied command. Load its reference and follow it step by step.
• Two commands fit. Ask once which one, naming both and what each would do. Then follow the answer.
• Standards from nothing. Start at mine, because write needs clusters.json. Offer write when mine finishes.
• A question about the skill itself. Answer from this file without running a command.

Menu

standards: what should I do?
  mine    Mine review comments into ranked conventions
  write   Write standards from the ranked conventions
  review  Review a diff against the standards
  prune   Test each rule and delete the ones agents already follow
  doctor  Check the standards for drift
Every command takes an optional scope: a path, an area or a §.

When generated standards exist, list their areas under the menu as § number and slug.

Pipeline

mine feeds write, and write produces what review applies. prune and doctor maintain the standards after that.

Work files live in .standards/ at the target repo's root. Keep that folder out of version control, because review comments quote code and people. prune.json is the exception. It lives beside the generated standards, because doctor reads it there.

<skill-dir> in the references is the folder that holds this file. The scripts in <skill-dir>/scripts/ speed up mine and need only a JavaScript runtime's standard library.

Judgment model

mine, prune and doctor ask many small typed questions. A judgment model is a fast model that answers typed questions (pick one of a set, a yes/no probability, a score) with calibrated probabilities. If your environment gives you a judgment model, answer the questions file with it; otherwise answer them yourself. Your environment's own instructions decide which model that is.

Generated standards

• The default output is one skill with an index of one-liners, each linking to a § file. Templates: index and section.
• One skill per area is an option, settled per repo by the shape experiment in references/write.md.
• Generated standards follow the same rules as this skill: spec frontmatter only, no URLs, no dates, and a why for every rule.