How to use the triage agent skill for Qwen Code GitHub issues and PRs

What the triage skill makes the agent do

The skill turns an AI coding agent into a maintainer's assistant for the Qwen Code repository. For an issue, it classifies the report as bug, feature, docs, unclear or inadmissible, and posts a short bilingual comment with labels and a next step. For a pull request, it runs a staged gate, reviews the change statically, quotes the PR's own CI results as test evidence, and posts three comments: Gate, Review and Test, and Final Decision.

Default stance for issues is admissible. The skill closes only narrow cases such as spam, harassment or content unrelated to Qwen Code. It is written mainly for unattended GitHub Action runs but can also be invoked locally.

• Issues: one comment that is updated in place as Stage 2 adds results.
• Pull requests: three staged comments, each marked so reruns can find them.
• Refactor PRs from forks are reviewed as usual but escalated to a maintainer instead of approved.

Install the triage skill from the ZIP

Download the ZIP from this page and unzip it. Place the resulting triage folder in one of two locations: ~/.claude/skills/ for your user account, or .claude/skills/ inside a project to scope it to that project.

If you use the Claude apps instead of a local folder, upload the skill under Settings → Capabilities → Skills.

• Personal install: unzip into ~/.claude/skills/.
• Project install: unzip into the project's .claude/skills/ folder.
• Claude apps: Settings → Capabilities → Skills, then upload.

Requirements before you start

The skill needs the GitHub CLI (gh) authenticated with access to the repository. The host agent must support Git worktrees, through the enter_worktree and exit_worktree tools. In CI, the workflow needs a GitHub token with write permission for comments and labels.

The pull request workflow reads a reference file, references/pr-workflow.md. That file is referenced by the skill but was not included in the files reviewed for this page, so make sure it is present in your copy before relying on PR stages.

• GitHub CLI (gh), authenticated with repository access.
• Host agent with enter_worktree and exit_worktree tools.
• In CI: a GitHub token with write permission for comments and labels.
• references/pr-workflow.md present alongside SKILL.md.

Trigger the skill with a natural request

Ask the agent in plain language and name the issue or PR number. The skill takes a number and an optional repository. It reads the number from the argument or from the ISSUE_NUMBER or PR_NUMBER environment variables, and the repository from --repo, REPOSITORY or GITHUB_REPOSITORY.

The argument format is shown below. Without --repo, the repository comes from the environment variables listed above.

• "Triage issue 1234 in QwenLM/qwen-code."
• "Run the PR gatekeeper on pull request 567."
• "Review PR 567 for admission, then post the staged comments."

Understand the safety limits

The review is static by design. The skill never runs code from the pull request: it does not run npm, node, npx, build or test commands against a tree containing the PR's changes, and it does not check out or apply the PR. Test evidence comes from the PR's own CI checks through the API. In CI the agent environment holds a write token, so executing PR code could expose it.

Issue text and PR text are untrusted input and must never be placed into a shell command. Labels are applied only if they already exist, and process labels such as welcome-pr and maintainer are left alone. Draft PRs are skipped.

If a scale or scope concern arises, the skill raises it as a question rather than blocking the PR. The skill also includes a duplicate guard: unattended CI events exit if a stage marker already exists, while explicit reruns update prior comments in place.

• No execution of PR code; no gh pr checkout.
• Labels: existing only, never created.
• Drafts are skipped.
• Comments are posted from files with --body-file or gh api -F body=@FILE, never --body @FILE.
• The skill's manifest shows that it does not run scripts and does use the network.

Read the output and rerun safely

Each staged comment includes an invisible marker in the form <!-- qwen-triage stage=N -->, where N is the stage number. The duplicate guard uses these markers, not the headings, to decide whether a stage has already run. Comments are written in English first, with a Chinese version inside a details block.

When the agent reports a missing test result, read the line that says what was not verified and why. The skill is designed to say this rather than hide it. For live behavior of terminal interface changes, the skill notes that a maintainer can trigger the isolated @qwen-code /tmux job.