Skip to content
Get packs

Guide · Agent skill

How to use the Agent Reproduce Align agent skill

Updated · written with AI from the project's own documentation and checked against our catalogue data

agent-reproduce-align is an agent skill from the Qwen Code repository. It is for maintainers and contributors who have already built a feature in Qwen Code and want to check it against the same feature in Codex or Claude Code. After following this guide you will know how to install the skill, ask your agent to run a parity check, and read the comparison results.

Agent Reproduce AlignQwenLM/qwen-code · Apache-2.0

Quick start

  1. 1Download and unzip. Download the skill ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder.
  2. 2Confirm prerequisites. Check that you have a Qwen Code checkout, Python 3, Bash, and the reference agent CLI (Codex or Claude Code) installed locally.
  3. 3Make a natural request. Ask the agent to run a parity check for a named feature against codex or claude-code.
  4. 4Run the paired capture. Set REPRO_REFERENCE_AGENT and run run_pair_capture.sh with an output directory and the two commands, or capture manually with tmux for interactive commands.
  5. 5Normalize and compare. Run normalize_trace.py on each http.jsonl, then compare_traces.py on the two normalized files.
  6. 6Patch and rerun. Fix the must-match differences in Qwen Code, rerun the smallest failing scenario, and document any intentional differences.

What is an agent skill and what does this one do?

An agent skill is a folder of instructions and helper files that an AI coding agent loads when a request matches the skill's description. This skill is triggered after a Codex or Claude Code feature has been implemented in Qwen Code.

The skill makes the agent run the same scenario through a chosen reference agent (codex or claude-code) and through Qwen Code, recording the HTTP requests and terminal output of each run. It then normalizes and compares the two recordings. The comparison focuses on tool names, argument schemas, required fields and model settings rather than exact wording, and the agent uses the differences to patch Qwen Code and rerun the failing scenario.

Install the agent-reproduce-align skill

Download the ZIP for this skill from this page and unzip it. You can install it in one of two places:

  • Personal: unzip into ~/.claude/skills/ so it is available across your projects.
  • Project: unzip into the .claude/skills/ folder of a specific project.
  • Claude apps: upload the skill under Settings → Capabilities → Skills.

Trigger the skill with a natural request

You do not need a special command. Describe the parity check in plain language, and name the reference agent and the feature. The skill's description tells the agent when to use it, so requests about comparing a new Qwen Code feature with Codex or Claude Code are the right trigger.

Before running anything, the skill asks the agent to restate the parity target: the feature name and trigger, the selected reference agent, one baseline prompt or interaction script, acceptable differences and must-match fields.

  • Compare the new /help output in Qwen Code with what codex exec '/help' does.
  • Check that our new file-search tool has the same name and argument schema as the Claude Code equivalent.
  • Run the paired capture for the slash-help scenario against claude-code and show me the differences.

Run a paired capture and compare the traces

For simple non-interactive shell scenarios, the skill includes run_pair_capture.sh. Set REPRO_REFERENCE_AGENT to codex or claude-code, then pass an output directory, the reference command and the Qwen Code command. When REPRO_REFERENCE_AGENT is set, the runner also writes reference state-before, state-after and state-diff folders.

The skill then normalizes each trace with normalize_trace.py and compares the results with compare_traces.py. Use the paired runner only when shell quoting is simple. For interactive slash commands, the skill says to run the two captures manually with tmux so each side receives the same keystrokes.

  • Differences are checked in this order: reference-agent state changes, missing tool or function names, schema shape and required fields, model settings and response mode, prompt role and order, then terminal output and exit status.
  • Exact prompt wording, provider endpoints, model names, IDs, timestamps and token counts are generally not chased.
  • The task is done when both traces exist, no must-match differences are unexplained and Qwen Code tests or smoke commands cover the fixed behavior.
REPRO_REFERENCE_AGENT=codex \ .qwen/skills/agent-reproduce-align/scripts/run_pair_capture.sh \ .repro-runs/slash-help \ "codex exec '/help'" \ "npm test -- --runInBand"

Read the normalized comparison

Normalization turns raw mitm JSONL output into stable JSON. Comparison then shows where the two recordings disagree. The skill recommends reading the reference workflow document, references/alignment-workflow.md, before the first comparison pass.

Only redacted, minimal fixtures should be kept in the repository.

  • Normalize: scripts/normalize_trace.py reads .repro-runs/reference/http.jsonl and writes normalized.json.
  • Compare: scripts/compare_traces.py takes the reference and Qwen Code normalized files.
.qwen/skills/agent-reproduce-align/scripts/normalize_trace.py \ .repro-runs/reference/http.jsonl \ > .repro-runs/reference/normalized.json .qwen/skills/agent-reproduce-align/scripts/compare_traces.py \ .repro-runs/reference/normalized.json \ .repro-runs/qwen/normalized.json

Safety notes: scripts, network and sensitive data

This skill runs scripts and uses the network. The captured HTTP traces may contain prompts, keys or account details, and the skill relies on you to redact them before anything is kept in the repository. Do not run it against live credentials you are not comfortable exposing to a capture tool.

The skill assumes a Qwen Code repository checkout and local installs of the reference agent. The default target is the current working directory; a different path is used only when you specify one.

  • Review the scripts before running them, especially if you are using live accounts.
  • Keep raw .repro-runs output out of commits; keep only redacted minimal fixtures.
  • Capture tooling comes from the sibling agent-reproduce-feature skill (run_with_mitm.sh and capture_state.py), which this skill depends on.

Tips for faster parity loops

Use the smallest scenario that shows the failure. The workflow is designed to patch, rerun only the failing scenario and repeat.

Record which reference agent you chose. If the earlier choice is unavailable, the skill asks once and records the answer in the run notes.

  • Write down intentional differences in the task notes or Qwen Code docs when they affect users.
  • Use REPRO_REFERENCE_STATE_ROOT only for tests or custom state directories.

What people use it for

  • Checking that a new Qwen Code slash command accepts the same arguments as the reference agent's equivalent.
  • Confirming that a new tool exposes the same name, argument schema and required fields as a Codex or Claude Code tool.
  • Building small, redacted trace fixtures to use in Qwen Code parity tests.
  • Diagnosing why a feature's model settings or response mode differ from the reference agent.

Frequently asked questions

How do I install the agent-reproduce-align skill?

Download the ZIP from this page and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps, you can instead upload it under Settings → Capabilities → Skills.

What do I need to run agent-reproduce-align?

The source lists Python 3 for normalize_trace.py and compare_traces.py, Bash for run_pair_capture.sh, mitm-based capture tooling from the agent-reproduce-feature skill, an installed Codex or Claude Code CLI as the reference agent, and Qwen Code. Reference agent credentials may also be needed if its CLI requires login. The files do not spell out every requirement in detail.

Is agent-reproduce-align free to use?

The item is listed under the Apache-2.0 licence. The source does not state pricing for the skill itself, so check the licence text for terms.

Is this skill meant for ordinary Qwen Code users?

No. It assumes a repository checkout and local installs of the reference agent. It is intended for maintainers and contributors who are adding features that mirror Codex or Claude Code behavior.

Is it safe to run with my real API keys?

Be careful. Captured HTTP traces may contain prompts, keys or account details, and the skill relies on you to redact them. Avoid running it with live credentials you are not comfortable exposing to a capture tool.

How is this different from just comparing outputs by eye?

The skill records HTTP requests and terminal output from both agents, normalizes them, and compares contracts such as tool names, argument schemas and required fields. It deliberately ignores exact wording, endpoints, timestamps and token counts unless the feature depends on them.

New agent skills, repos and AI tools — every week

The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.

Our verdict

This skill is a good fit if you maintain Qwen Code and need evidence that a new feature matches Codex or Claude Code at the level of tool names, schemas and settings. It is not useful for everyday Qwen Code use, for projects unrelated to parity with a reference agent, or for workflows where you cannot safely redact captured traces.

Download the skill