Skip to content

AGENTS.md

Latest commit

 

History

History
68 lines (59 loc) · 3.78 KB

AGENTS.md

File metadata and controls

68 lines (59 loc) · 3.78 KB
Raw

AGENTS.md — GitDesktop

GitDesktop is an AI-native, keyboard-first Git desktop client (Tauri 2 + React 19) with an Astro marketing site in site/. Codex and other agents that follow the AGENTS.md convention auto-load this file whatever their role (Claude sessions load CLAUDE.md instead), so it carries repo context and hard boundaries only.

Binding context — read before substantive work

CLAUDE.md and .claude/skills/gd-conventions/SKILL.md are the binding agent context: the standing brief and the repo playbook of conventions and gotchas. CONTRIBUTING.md holds the human conventions, PRODUCT.md the product intent.

Hard boundaries

  • Git is a whitelist. Permitted: git --no-pager diff, git --no-pager status, git --no-pager log, git --no-pager show, git branch --list — each may be prefixed with -C <path> to address a task worktree. Every state-mutating git command — commit, add/stage, checkout, reset, stash, rm, clean, push, pull, fetch, merge, rebase, tag, branch create/delete, worktree, remote, config — is forbidden, even "just to test": the user commits their own work, possibly in parallel with this run. .claude/skills/gd-conventions/SKILL.md is the canonical playbook and .claude/rules/git-safety.md its always-loaded excerpt; read either in the tree.
  • In a task worktree the sandbox additionally denies the git commands that write the git dir (measured on add, commit, checkout, rm, stash), and a denial there is the policy working, not an obstacle to route around. Commands touching only worktree files (git clean, for one) are not blocked, so the whitelist above is the only thing protecting them.
  • Repo files are written only when the run is executing a delegated work-package spec (CLAUDE.md's delegated-implementation section enumerates the full set of sanctioned write paths). A run handed no spec treats the repo as read-only.
  • Verification claims come from command output in this run, never from memory. Quote failures verbatim.
  • Never edit src/components/ui/ — vendored shadcn/Base UI primitives; fix at the feature or call site (that folder's README.md inventories the sanctioned local deltas).
  • No tree-wide rewrites. pnpm lint is biome check --write over src/ AND site/ — a rewrite, not a check; the check form is pnpm exec biome check ./src/. In a task worktree even that tree-wide check false-fails on CRLF, so the trustworthy gate is the per-file LF-copy biome ci form (gd-conventions), or whatever the spec's Verification field names. Never run repo-wide cargo fmt in src-tauri; rustfmt <file> on individual new files only.
  • No scratch or temp files inside the repo — working files belong in the session scratchpad or C:/temp. The one sanctioned exception: a run that may create files judges formatting with the per-file LF-copy biome ci gate, whose __cigate__<name> sibling lives beside the original for the length of the check and is deleted after. A read-only run creates nothing and reports tree-wide formatting as unverified instead.

Verification commands

  • pnpm build — typecheck plus frontend bundle. Plain tsc --noEmit is a no-op in this repo; only tsc -b (or pnpm build) typechecks.
  • pnpm run checks — the repo guard scripts CI runs.
  • cargo test --manifest-path src-tauri/Cargo.toml
  • cargo clippy --manifest-path src-tauri/Cargo.toml -- -D warnings
  • cd site && pnpm build — the marketing site.

When the run was given a task spec

Stay strictly inside its stated file scope: a spec that looks wrong, or a fix that would need a file the spec did not list, is reported back rather than improvised around. A Permission-denied on a path outside the workspace is an environment limit — report it, don't work around it.