Testing LocalPilot

Fast checks, every change

npx tsc --noEmit && npx eslint src         # types and lint (what `npm run compile` runs before bundling)
node --check media/chat.js                  # webview scripts are plain JS and not type-checked
npm run smoke                               # offline: mock Ollama/OpenAI servers, agent loop, context, pool, worktrees

Mocha suites in VS Code

npm test compiles to out/, bundles, lints and runs every out/test/**/*.test.js in an Extension Development Host. .vscode-test.mjs points telemetry and the skill catalog at scratch folders, and VS Code builds are cached in .vscode-test/.

Write tests in src/test/<area>.test.ts using suite/test and assert. Put logic in pure functions or classes with injected dependencies (a runner, a clock now, a file path) so tests need no real models or timers. src/test/autonomy.test.ts shows the pattern: a controllable runner, a fake clock, await q.flush() before teardown.

Run one suite, or run while another VS Code test host is open

The test CLI refuses to start while another instance using the same user-data folder is running, and it resolves packages relative to its config file. So create a temporary config in the repo root with its own user-data folder, run it, then delete it:

npm run compile-tests && node esbuild.js
d=$(mktemp -d /tmp/lp-test-XXXX)
cat > .vscode-test.tmp.mjs <<EOF
import * as fs from 'fs';
export default {
  files: 'out/test/autonomy.test.js',            // or 'out/test/**/*.test.js'
  workspaceFolder: fs.mkdtempSync('$d/ws-'),
  launchArgs: ['--user-data-dir', '$d/ud'],
  env: { LOCALPILOT_TELEMETRY_DIR: '$d/telemetry', LOCALPILOT_SKILL_CATALOG: '$d/cat.json', LOCALPILOT_WORKTREES_DIR: '$d/worktrees' },
  mocha: { timeout: 60000 },
};
EOF
npx vscode-test --config .vscode-test.tmp.mjs 2>&1 | grep -E "✔|passing|failing|^\s+[0-9]+\) |Error|expected|actual"
rm -f .vscode-test.tmp.mjs; rm -rf "$d"

• The telemetry folder must end in telemetry, because telemetry.test.ts asserts it.
• Don't kill VS Code processes you didn't start. Another session or the user may own them.
• The full suite takes about 4 minutes. Run it in the background and keep working.

End-to-end tests with a real model

src/test/e2e.test.ts needs Ollama on localhost:11434 with a tool-capable model, and skips itself without one. Set LOCALPILOT_E2E_MODEL=ollama/<tag> to choose the model. A timeout like session … still running after 200000 ms with only the first user message in the transcript means the model never answered. Check curl -s localhost:11434/api/ps for large models loaded and other test runs sharing Ollama before suspecting the code. Say so in your report instead of calling the run green.

Report

Give pass and fail counts. For each failure, say whether it comes from your change, from the environment (Ollama load, a config artifact), or is unknown, with the evidence. Don't call anything verified that you didn't run.