numbers-le: Instructions file for Claude Code

CLAUDE.md

AGENTS.md is the technical source of truth for this repo: the engineering standard the code is held to — control flow, error handling, immutability, structure — plus this repo's architecture, invariants, toolchain and release. Read it before writing code. README.md is user-facing and partly generated.

The repo also hosts the Rust CLI in crate/ — read crate/CLAUDE.md and crate/AGENTS.md for that side; the shared corpus is crate/fixtures/.

Where to look

Question	File
How should this code be written?	AGENTS.md — the standard, plus this repo's architecture and invariants
What does the user see?	README.md — Testing and Performance are generated
What changed?	CHANGELOG.md

Gates

bun run typecheck && bun run lint && bun run test

Before a release, also bun run test:integration, bun run package, and bun run test:e2e-vsix — the last is the only test that exercises the artifact users actually install.

Things that will bite you

• Two README sections are generated. Testing and Performance sit between <!-- coverage:start --> / <!-- performance:start --> markers and come from scripts/coverage-readme.js and scripts/perf-readme.js. Edit the code and regenerate; do not type numbers in by hand. CI fails if the coverage figures no longer match a real run.
• Output changes must update the characterization goldens in the same commit, with a CHANGELOG entry describing the behaviour change.
• Every claim must be provable. No feature, metric or format goes in a README, the manifest, or help text unless the code backs it. That governs behaviour and numbers — not availability. Whether something is published, listed or installable is a fact about a registry at a moment in time, and it is false right up until you make it true. Copy for a release you are about to make is staged, never forbidden: write it, and let the release commit be what makes it true.
• This repo is one of ten identical ones. The shared config files, scripts and workflows are byte-identical across the family, and letools-site/scripts/check-fleet.ts is what holds them there rather than memory: run bun run check:fleet ../ from a checkout of the site with the ten beside it, or dispatch its Fleet workflow. It names the file and the repos that drifted, so a missed copy is a report rather than something you find months later. Anything under crate/ is outside the check on purpose — the crates stand on their own.
• Extraction is shared with the Rust CLI, and the corpus under crate/ is the contract. Changing extraction behaviour means running bun scripts/check-extraction-parity.ts and updating the corpus — on both sides, in the same commit. CI fails when either drifts.
• What the contract holds equal is the shared extract_numbers MCP tool, which both servers offer and must answer identically; a difference there is a bug. The surfaces are meant to differ. This one is IDE-first — the active buffer, read by a person. The CLI is terminal-first: a tree walk, exit codes and JSON Lines, none of which has an editor equivalent. That is not drift, and nothing holds them equal — see crate/SPEC.md.
• How a number is printed is part of that contract. The corpus pins the boundaries of JavaScript's Number::toString, because the crate has to reproduce them and Rust's own formatting does not.
• Localization is two mechanisms, and they fail separately. src/i18n/package.nls.*.json covers the manifest; l10n/bundle.l10n.*.json covers runtime strings through vscode.l10n.t(). Twelve locales each, held in exact key parity by the integration test. Never call l10n.t() at module scope, never compare a translated label against an English literal, and use positional {0} placeholders rather than template literals.
• CI narrows itself on a docs-only push. A change touching only *.md and LICENSE runs the Linux leg alone and skips the Zed build; ci-crate.yml runs its policy gate with every Rust job skipped. Nothing that covers the change is skipped — the README coverage gate, the integration suite and the installed-VSIX end-to-end are Linux-only anyway. Anything unrecognised, and an unreadable diff, counts as code and runs everything. A release commit always touches package.json, so a release still sees the full three-OS matrix.
• Coverage floors are a backstop, not a target. They sit well below where the code actually is, and they are not raised to track it — a floor that follows real coverage becomes a tax on writing the next module.