Worktree Safe Execution

🔄 Auto-generated mirror. Canonical source: contributing/worktree-safe-execution.md. Edit there — changes here are overwritten by CI.

Worktree-safe execution

Agents and developers increasingly run several checkouts of a repo side by side — git worktrees for parallel branches, agent isolation, or review copies. Display Case is built so that N working copies run independently with no shared or clobbered state. This doc explains the guarantee, the one rule you must follow, and the node_modules setup a fresh worktree needs.

The guarantee

Display Case holds no global or main-repo state. Resolution, the .display-case/ build cache (codegen + bundle cache + local visual-regression baselines + a11y cache), and all repo-relative paths anchor to the package you point it at — never to where the tool itself is installed. So:

• A worktree's run stays entirely inside that worktree.
• Two checkouts never read or write each other's build cache or baselines.
• An agent's component edits show up in the showcase that worktree serves.

This is enforced behavior, specified under openspec/specs/showcase-location/spec.md (the original change: openspec/changes/archive/2026-06-19-make-display-case-worktree-safe).

How the target is resolved

There are exactly two ways to name the target, both anchored where you'd expect:

• . (or no argument) — discovers the nearest display-case.config.ts by walking up from the current directory (bounded). Run from anywhere inside the worktree (any depth) and you get that worktree's package. The bare display-case and display-case . resolve identically — there is no separate "no-argument" behavior to reason about (important under bunx, where the tool's install location and your working directory differ).
• explicit <pkgDir> (display-case apps/foo) — used as given, and validated to contain a display-case.config.ts. A wrong directory fails loudly rather than silently serving an empty or unrelated showcase. Relative paths resolve against the current directory, so a worktree-relative path stays in the worktree.

The one rule: launch it from inside the worktree (cwd within the checkout), or pass a worktree path explicitly — don't rely on a process cwd that points at a sibling checkout. Agent launch configs (.claude/launch.json) pass an explicit package path for exactly this reason.

Ports

A busy port should not hard-fail a second checkout. An explicit --port wins; otherwise DISPLAY_CASE_PORT (set per-worktree) is honored, falling back to the server default (3100). The server treats the chosen port as preferred and bumps off a busy one, so parallel runs coexist. The Playwright e2e config takes the same env overrides (DISPLAY_CASE_PORT, DISPLAY_CASE_A11Y_PORT, …).

Running checks/tests from a fresh worktree

A fresh worktree has no node_modules — it is gitignored, so git worktree add copies none, and tsc/Biome/the test suites can't resolve modules without it. This is a single-package repo, so setup is just:

bun install          # ~2s from bun's global cache; leaves git status clean

Then run the gate from the worktree root:

bun run lint         # biome
bun run typecheck    # tsc --noEmit
bun run check        # display-case check --structure --tokens --ssr (browser-free)
bun test             # unit / type tests
bun run e2e          # Playwright chrome suite (one-time: bun run e2e:install)

The bundled skills automate this: lint-in-worktree, test-in-worktree, and run-e2e (under .agents/skills/ and .claude/skills/) each set up node_modules first, then run their suite.

See also

• coding-best-practices.md — render purity and the determinism rules that make isolated runs reproducible.
• testing-best-practices.md — parallel-safe, read-only e2e isolation and env-overridable ports.
• Product doc: ../docs/cli.md — every command and flag.