qwen-code/AGENTS.md

AGENTS.md

Working Principles

Simplicity First

Minimum code that solves the problem. Nothing speculative. (This is the principle we care about most.)

• No features beyond what was asked.
• No abstractions for single-use code.
• No "flexibility" or "configurability" that wasn't requested.
• No error handling for impossible scenarios.
• If you write 200 lines and it could be 50, rewrite it.

Ask yourself: "Would a senior engineer say this is overcomplicated?" If yes, simplify.

Core Infrastructure Is Maintainer-Only (triage gate, two-tier rule)

Core modules — packages/core/src/**, packages/*/src/auth/**, packages/*/src/providers/**, packages/*/src/models/**, packages/*/src/config/**, packages/*/src/tools/**, packages/*/src/services/**, cross-package changes — are the architectural backbone. External PRs touching them face a two-tier gate (maintainer-authored PRs are exempt):

1. Large-scope refactor changes (500+ production logic lines in core, excluding test and generated/schema files) → hard block. Skip evaluation entirely — the maintainer exemption above is the sole exception. Large-scale core refactors must be maintainer-initiated. When counting lines, exclude files matching *.test.ts, *.test.tsx, *.spec.ts, *.spec.tsx, __tests__/**, *.schema.ts, *.schema.json, *.generated.ts, and **/generated/** — only production logic counts. feat-type and other non-refactor PRs are NOT hard-blocked on size; they escalate to the maintainer for awareness instead. A non-blocking advisory also applies at 1000+ production logic lines. Breadth alone is not size — a low-risk sweep that touches 10+ files but changes a line or two each is escalated to a maintainer for awareness and otherwise judged under Tier 2's 100%-confidence bar, not auto-rejected on file count.
2. Small-scope changes → gate may evaluate, but must be 100% confident. Any doubt at all → escalate to maintainer. "The direction looks correct" is not confidence. The gate must name every downstream consumer; if it cannot, escalate.

When in doubt, escalate. Better to wrongly escalate than to wrongly approve.

Common Commands

Building

corepack pnpm install --frozen-lockfile  # Install all dependencies (the tree CI tests)
npm run build      # Build all packages (TypeScript compilation + asset copying)
npm run build:all  # Build everything including sandbox container
npm run bundle     # Bundle dist/ into a single dist/cli.js via esbuild
                   # (requires build first)

Development

npm run dev        # Run CLI from TypeScript source

Runs the CLI via tsx with DEV=true. Changes to packages/core or packages/cli are reflected immediately without rebuilding. Browser Use reuses its built runtime, like other compiled workspace dependencies. After editing Browser Use, run npm run build --workspace=@qwen-code/browser-use to rebuild and stage its runtime for dev. Normal npm install prepares it automatically.

Unit Testing

Tests must be run from within the specific package directory, not the project root.

Fresh clone or new worktree: packages/cli unit tests import workspace packages (@qwen-code/acp-bridge, @qwen-code/web-templates, packages/channels/*, ...) through their built dist/ output, and packages/core tests import the package's own entry (@qwen-code/qwen-code-core), which also resolves into dist/. A plain corepack pnpm install --frozen-lockfile already builds them via the prepare script, but a worktree that shares the main checkout's node_modules (or a deep-cleaned copy) does not have them. If any prerequisite is missing, a vitest globalSetup guard stops the run and names the fix; build once from the repository root:

npm run build

Installing dependencies: CI, release and every workflow install with the pnpm version pinned in packageManager (corepack pnpm install --frozen-lockfile); scripts still run through npm run. An additional Git worktree can use node scripts/setup-worktree.js, which runs the same install from a shared store (≈ 99 MiB on copy-on-write filesystems such as APFS, btrfs, and XFS with reflink; ≈ 1.2 GiB on ext4) and skips the prepare build, so run npm run build before package tests. When dependencies change, edit the manifest, run corepack pnpm install (or corepack pnpm add), and commit pnpm-lock.yaml; every CI install runs with --frozen-lockfile, so a lockfile that no longer matches the manifests fails the build.

Run individual test files (always preferred):

cd packages/core && npx vitest run src/path/to/file.test.ts
cd packages/cli && npx vitest run src/path/to/file.test.ts

Update snapshots:

cd packages/cli && npx vitest run src/path/to/file.test.ts --update

Avoid:

• npm run test -- --filter=... — does NOT filter; runs the entire suite
• npx vitest from the project root — fails due to package-specific vitest configs
• Running the whole test suite unless necessary (e.g., final PR verification)

Test gotchas:

• In CLI tests, use vi.hoisted() for mocks consumed by vi.mock() — the mock factory runs at module load time, before test execution.

Integration Testing

Build the bundle first: npm run build && npm run bundle

Run from the project root using the dedicated npm scripts:

npm run test:integration:cli:sandbox:none
npm run test:integration:interactive:sandbox:none

Or combined in one command:

cd integration-tests && \
  cross-env QWEN_SANDBOX=false npx vitest run cli interactive

Gotcha: In interactive tests, always call session.idle() between sends — ANSI output streams asynchronously.

Linting & Formatting

npm run lint       # ESLint check
npm run lint:fix   # Auto-fix lint issues
npm run format     # Prettier formatting
npm run typecheck  # TypeScript type checking
npm run preflight  # Full check: clean → install → format → lint → build
                   # → typecheck → test

Code Conventions

• Module system: ESM throughout ("type": "module" in all packages)
• TypeScript: Strict mode with noImplicitAny, strictNullChecks, noUnusedLocals, verbatimModuleSyntax
• Formatting: Prettier — single quotes, semicolons, trailing commas, 2-space indent, 80-char width
• Linting: No any types, consistent type imports, no relative imports between packages
• Core imports in cli: production code in packages/cli/src imports core values from the module that defines them (@qwen-code/qwen-code-core/utils/debugLogger.js), not the package root, which evaluates all of core in every test that reaches the file. Type-only imports are exempt. Files that predate the rule are allowlisted in eslint.legacy-core-barrel-imports.mjs; drop an entry when you move its file off the root, never add one.
• Tests: Collocated with source (file.test.ts next to file.ts), vitest framework
• File naming: PascalCase.tsx for React components, kebab-case.ts for .ts files in packages/core and packages/cli (enforced by ESLint). Existing camelCase files are allowlisted in eslint.legacy-filenames.mjs; rename opportunistically when touching them, updating all imports in the same commit (note: renames lose git blame history).
• Comments: Default to none. Add only when why is non-obvious; don't delete existing ones as cleanup.
• Commits: Conventional Commits (e.g., feat(cli): Add --json flag)
• Node.js: Development and production both require >=22 (Ink 7 + React 19.2 requirement)

Web Shell UI development

• Prefer the shared primitives in packages/web-shell/client/components/ui when developing Web Shell UI. Do not duplicate an existing primitive or rewrite stable CSS Modules solely for consistency.
• If a required primitive is missing, run npx shadcn@latest add <component> from packages/web-shell, then review the generated diff. Do not let the CLI overwrite the existing global CSS, semantic tokens, CSS scoping, or portal-root integration. Keep generated components internal unless a public package API is explicitly required.
• Web Shell supports React 18 and React 19. Generated shadcn components often assume React 19 ref semantics, so wrappers that accept or receive refs — including Radix asChild, Slot, Presence, and portal children — must use React.forwardRef and pass the ref to the underlying DOM or Radix primitive. Add a regression test for any ref-sensitive component path.
• Use unprefixed Tailwind classes and shadcn semantic color tokens such as background, primary, and muted. The package build scopes generated CSS to the Web Shell root and portal root and prefixes global animations and CSS property registrations; changes must preserve that isolation from host-page styles.
• Components with portals, such as dialogs, popovers, dropdown menus, and tooltips, must use useWebShellPortalRoot() as the Radix portal container so themes, scoped CSS, and z-index variables continue to apply. Preserve existing data-web-shell-* attributes and public --web-shell-* CSS variables. See packages/web-shell/README.md for the full conventions.

Development Guidelines

General workflow

1. Design doc for non-trivial work — write one in docs/design/ if the change touches multiple files or involves design decisions. Skip for small bugfixes. Provide both an English <name>.md and a Chinese <name>.zh-CN.md version in the same directory, following the design documentation requirements. Add reciprocal language links and keep both versions complete and synchronized in the same change, including when updating an existing design.
2. Test plan for behavioral changes — write an E2E test plan in .qwen/e2e-tests/ when the change affects user-observable behavior. Dry-run against the global qwen CLI first to confirm the baseline.
3. Build, typecheck, and test before declaring done: npm run build && npm run typecheck, plus unit tests for the files you changed.
4. Self-audit before declaring done — read the full diff you are about to ship, including new untracked files, in open-ended passes, not hunting for anything specific. Then verify each change, and each green test you rely on as evidence, presuming it wrong (a passing test can assert the wrong thing). Stop after two consecutive clean passes — a clean pass is evidence about that pass, not the code. A fix re-runs step 3, resets the clean-pass count, and gets a further pass over the updated diff — never exit on a pass that found something. If five passes bring no convergence, say so instead of declaring done. Scale to the diff: one clean, careful pass suffices for a trivial change.

Feature development

Use the /feat-dev skill for the full workflow: investigate, design, test plan, dry-run, implement, verify, self-audit, code review, and iterate.

Bugfix

Use the /bugfix skill for the reproduce-first workflow: reproduce, fix, verify, test, self-audit, and code review.

Code Review

Project-specific rules for /review. The skill loads this section verbatim (by its ## Code Review heading) and hands it to every review agent, so keep it to things a reviewer of this codebase must check — not general advice.

• Verify a finding against the exact reviewed commit before reporting it. Read the lines you are about to cite. A Critical that quotes code not present at the commit under review is worse than no finding — it blocks the author over nothing. Do not report a defect you have only inferred from a symbol name or a diff fragment.
• A C=0 / APPROVE is a claim, not a default. Before submitting one, take each unresolved Critical already on the PR and check it against the code as it stands: still stands / fixed by this diff / cannot tell. A GitHub thread can read isResolved: false, isOutdated: false for a bug that a later commit fixed on an adjacent line — the flag tracks the anchored line, not the fix.
• For every added field, option, or optional parameter, grep its read sites, including outside the diff. A foo?: boolean that is declared and read but never set by any caller is a dead switch (options.foo ?? true always takes the default). Decide severity at the read site; never explain an unpopulated field with author intent you cannot observe.
• Classify every added or changed daemon route by ownership. Name whether it is process-global, legacy-primary, selected-runtime, live-session-owner, or persisted-workspace scoped, and verify every downstream consumer matches that scope.
• Verify workspace-scoped routes stay inside the resolved runtime. Check the environment, bridge, service, filesystem, trust boundary, and failure paths. Each unknown, untrusted, ambiguous, bootstrapping, draining, or removed state must follow its declared failure semantics and must never fall back to the primary runtime.
• Match the house style when judging. ESM only; no any; no relative imports between packages; kebab-case.ts for .ts in packages/core and packages/cli, PascalCase.tsx for React components; tests collocated as file.test.ts. Comments default to none — flag a missing comment only where the why is genuinely non-obvious, and never fault a diff for deleting a comment that no longer applies.
• A missing test for changed behavior is a Suggestion, not a Critical, unless the untested path is itself the defect.
• Check design documentation in both languages. New or updated designs must include linked English and Chinese versions with matching structure, decisions, constraints, and acceptance criteria. A translation gap alone is a Suggestion, not a Critical.

GitHub Operations

Use the gh CLI for all GitHub-related operations — issues, pull requests, comments, CI checks, releases, and API calls. Prefer gh issue view, gh pr view, gh pr checks, gh run view, gh api, etc. over web fetches or manual REST calls.

Testing, Debugging, and Bug Fixes

• Bug reproduction & verification: spawn the test-engineer agent. It reads code and docs to understand the bug, then reproduces it via E2E testing (or a test-script fallback). It also handles post-fix verification. It cannot edit source code — only observe and report.
• Hard bugs: use the structured-debugging skill when debugging requires more than a quick glance — especially when the first attempt at a fix didn't work or the behavior seems impossible.
• E2E testing: the e2e-testing skill covers headless mode, interactive (tmux) mode, MCP server testing, and API traffic inspection. The test-engineer agent invokes this skill internally — you typically don't need to use it directly.

Submitting PRs

When creating a PR, follow the template at .github/pull_request_template.md. After the PR is submitted, post a separate comment with the E2E test report if applicable.

• PR description: explain the motivation and changes in prose. Avoid referencing file names or function names.
• Reviewer Test Plan (template section): describe behaviors a reviewer should verify and what to expect, not scripted test commands. Use How to verify for reproduction steps; Before/After for TUI evidence when applicable.
• Line wrapping: do not hard-wrap the PR body at a fixed column width. GitHub renders single newlines as <br>, so a wrapped description displays as a narrow column. Write each paragraph or list item as one long line.
• Don't let review rounds balloon the PR. Every accepted change widens the diff and tends to trigger another round, so a PR can drift far past its original intent. Once a PR has been through roughly 5 review rounds, land only Critical fixes — correctness, security, data loss, regressions — and defer remaining Suggestions to a follow-up issue or PR. Record each deferral in the PR thread so nothing is silently dropped.

Project Directories

Design docs and implementation plans are committed under docs/ so they are tracked in version control:

Directory	Purpose
docs/design/	Design docs for planned features
docs/plans/	Implementation plans

Other working artifacts live under .qwen/ (git-ignored):

Directory	Purpose
.qwen/e2e-tests/	E2E test plans and results
.qwen/issues/	Issue drafts before filing on GitHub
.qwen/pr-drafts/	PR drafts before submitting
.qwen/pr-reviews/	PR review notes
.qwen/investigations/	Structured debugging journals
.qwen/scripts/	Utility scripts