AGENTS.md

Machine entry point for AI coding agents working in this repo. The authoritative guidance for humans lives in CONTRIBUTING.md and README.md; this file only points at it and states the stable first-action facts. Do not restate evolving prose here — it will drift.

First action

Project: F#-style pattern matching for TypeScript/ESM. Node >=26 (pinned via .node-version), ESM-only (no CommonJS shim).
While iterating: npm run test (check:tsc + the unit suite) for fast feedback on the files you changed.
Optional code intelligence: this repo installs @spences10/pi-lsp (pinned in .pi/settings.json) as a project-local pi extension. It talks to the repo's own TypeScript 7 via tsc --lsp --stdio and exposes read-only tools — lsp_hover, lsp_definition, lsp_references, lsp_find_symbol, lsp_document_symbols, lsp_diagnostics(_many). When you are looking for a symbol, reach for the LSP before rg/grep — lsp_references / lsp_find_symbol / lsp_definition / lsp_document_symbols are semantic and cross-file, so they see shadowing, imports and overloads that a text search cannot; use lsp_hover to read inferred types on generic-heavy code. Use rg for what the LSP cannot see — doc prose, string literals, config, task lists, file discovery — and reconcile the two sets before editing (symbols from the LSP, strings and prose from rg). It has no rename / code-action / apply-edit surface — the write side is pi's edit tool + check:tsc. Treat empty LSP output as inconclusive, not success: npm run test / npm run verify remain the sole authoritative gate (see the next bullet). The server keeps running across that gate with a ~5 min idle timeout and registers no file watchers, so if you change tsconfig.json / package.json mid-session its diagnostics can be stale — when LSP output disagrees with check:tsc, trust check:tsc and restart pi (or wait out the idle timeout) before concluding the LSP is wrong.
Definition of done — run this before you call the work finished: npm run verify. If all green, commit. If red, look at the output, fix the root cause, and re-run.
On commit: write a good message (see CONTRIBUTING.md § Commit messages). Lefthook's pre-commit hook already runs the fast, offline, staged-file checks — don't run them by hand. If the hook fails on style, npm run fix, restage, recommit.
Document decisions where the next maintainer will look: rationale, rejected alternatives and known issues go in development/<category>.md (see development/README.md); the actionable rule stays in CONTRIBUTING.md and links to it. Write each fact once — never copy the rule into development/ or the reason into CONTRIBUTING.md — and change both in the same commit when a rule changes.
npm run maintain is NOT part of the feature loop. Its scans are advisory, never a gate; run them only on an explicit maintenance / update-deps branch.
```
npm run test    # while iterating (fast feedback)
npm run verify  # definition of done: whole-project correctness, one shot
```

Never do

Don't silence the type system to force a green run. As an agent these are forbidden:

// @ts-nocheck, // @ts-ignore, // @ts-expect-error
// oxlint-disable / // oxlint-disable-next-line
editing .oxlintrc.json to silence a finding (e.g. turning typescript/no-floating-promises off)
as casts used to push an expression through (type-aware oxlint already flags unsafe assertions)

Fix the root cause with the type system instead — narrowing, generics, satisfies, conditional / mapped types, utility types (NonNullable, Exclude, …). TypeScript can express it; that's the intended tool. Suppressions — a source oxlint-disable or a .oxlintrc.json entry — are a human last resort, not a tool for you. If the types genuinely cannot express something, stop and surface the conflict (commit message / handover) rather than suppress it.

The same applies to the checks themselves: never git commit --no-verify (or otherwise skip a pre-commit / pre-push hook). The checks are fast and offline, so a redundant run is fine — bypassing a hook to get green is the identical anti-pattern. If a commit already skipped a hook, redo it through one: git reset --soft HEAD~1 && git commit -C <skipped-sha>.

Never start a long-lived / blocking process such as npm run watch. It runs until a human stops it with Ctrl-C, so in an agent turn it hangs forever and floods the context with continuous output. Reach for a one-shot command instead — npm run test (or npm run check) — to get feedback.

Backlog

backlog.tasks uses the vscode-todotasks format (not Markdown). A line ending in : is a project; every other line is a task. Status glyphs: ☐ open, ✔ done, ✘ cancelled; subtasks nest by indentation. Inline @tags carry metadata — @done / @cancelled mark completion, @critical / @high / @low / @today set priority. The (…) timestamp after @done is editor-generated: omit it when checking off by hand.

Required coupling: a ✔ line must also carry @done, and a ✘ line must carry @cancelled. The sandy081.todotasks extension treats the glyph as the completion signal, then unconditionally searches for the matching tag to decorate; a bare ✔/✘ with no tag makes it compute an illegal Range (negative character offset) that throws and kills all highlighting/decoration for the document. A ☐ may stand alone. So check off by hand as ✔ … @done (optionally @done (timestamp)), never a lone ✔.

Working on tasks

Every task — with or without subtasks — goes through the complete branching model:

Create the branch with npm run create:branch -- <prefix>/<desc>, inferring the prefix from the task content (feature/… / fix/… / chore/…) — do not hand-write git switch -c, the command enforces the clean-tree / current-main / green-baseline precondition. Work with commits (each subtask gets one or more), then present a concise handover for the user to review. Use this fixed shape:
```
## Handover — <branch>

**Implemented:** <what was built, and how>
**Judgement calls:** <where the task was unclear, and what you assumed>
**Known problems:** <open issues, caveats, follow-ups>
```
Once the user has no further objections, merge back: npm run create:finish (on the branch — it merges --no-ff, runs npm run verify, and deletes the branch). The branching model is documented in CONTRIBUTING.md § Branching model.

Follow CONTRIBUTING.md § Testing discipline (type-driven) throughout.

Read these

CONTRIBUTING.md § Rules the tools don't enforce — the constraints the linters don't catch; CI/review bounce these. The most important section.
CONTRIBUTING.md § Testing discipline (type-driven) — write the expectTypeOf (type) before the assert (red); the types are the feature.
CONTRIBUTING.md § Script prefix convention — adding an npm run script? reuse an existing prefix or it doesn't belong.
CONTRIBUTING.md § Commit messages — gitmoji + imperative + 50/72.
CONTRIBUTING.md § Feedback tiers — what runs when and at what cost (watch / pre-commit / pre-push / check / verify / fix / maintain / CI).
development/ — the decisions, rejected alternatives and known issues behind the rules; the “why” that CONTRIBUTING.md links to. Read the relevant file before changing an area.
development/tooling.md — the rationale behind each tool choice; read before changing tooling.
package.json #scripts — the source of truth for every command (the LEFTHOOK_FILES convention scopes them to staged files vs. the whole project).

8.2 KiB Raw Blame History