Skip to content
Get packs

Guide · Agent skill

How to use the Agent Reproduce Feature agent skill

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

The agent-reproduce-feature skill is an instruction bundle for an AI coding agent. It tells the agent how to run a feature from Codex or Claude Code under a local recording proxy, record what that tool sends to its model, and then implement the matching behavior in a Qwen Code repository. This guide explains what the skill does, how to install it, how to trigger it, and what to check before you trust its output.

Agent Reproduce FeatureQwenLM/qwen-code · Apache-2.0

Quick start

  1. 1Install the skill. Download the ZIP from this page and unzip it into ~/.claude/skills/ or a project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
  2. 2Check the requirements. Confirm Python 3 with mitmproxy (mitmdump), tmux, the reference CLI (codex or claude), and a Qwen Code checkout are available.
  3. 3Ask for the feature. Describe the feature and name the reference agent, Codex or Claude Code. The agent will ask once if you do not choose one.
  4. 4Let the agent capture the reference run. The agent runs the reference tool through run_with_mitm.sh or run_tmux_capture.sh and snapshots its local state into a fresh folder under .repro-runs/.
  5. 5Review the trace. Read the prompts, tool schemas, terminal states and state-diff.md, and redact any tokens before keeping excerpts.
  6. 6Implement and test in Qwen Code. The agent implements the smallest compatible behavior in Qwen Code and adds a focused test or smoke command. Hand off to agent-reproduce-align to iterate on parity.

What the skill makes the agent do

The skill turns an observed feature in another coding agent into an implementation task for Qwen Code. The agent treats the current session as the outer harness and runs the reference tool (codex or claude) as a nested process under test.

During a capture, the agent records HTTP request bodies, prompts, tool and function schemas, and terminal output. It also snapshots the reference tool's local config folder before and after the scenario so the changes can be compared. From these recordings it writes down the behavior to match and implements the smallest compatible version in Qwen Code, then adds tests or a smoke command.

  • Selects one reference agent: codex or claude-code
  • Captures request bodies with mitmdump via scripts/run_with_mitm.sh
  • Captures terminal screens with tmux via scripts/run_tmux_capture.sh
  • Snapshots and diffs local state with scripts/capture_state.py
  • Implements the narrow behavior in Qwen Code and adds verification

Install the skill

Download the ZIP for this skill from this page and choose one of two locations. For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder.

In the Claude apps you can instead upload the skill under Settings → Capabilities → Skills.

  • Personal: unzip into ~/.claude/skills/
  • Project: unzip into .claude/skills/ inside your repository
  • Claude apps: Settings → Capabilities → Skills, then upload

Check the requirements before you start

The skill needs a few tools on your machine. It will not run correctly without them. Install and confirm each one before asking the agent to capture anything.

  • Python 3 with mitmproxy installed (the mitmdump command must be available)
  • tmux, for terminal capture
  • The reference CLI installed locally: codex or claude
  • A Qwen Code repository checkout
  • Optional: a mitmproxy CA certificate at ~/.mitmproxy for TLS interception
command -v codex || true command -v claude || command -v claude-code || true

Trigger the skill with a natural request

Ask for the feature in plain language and name the reference agent. The skill description tells the agent to use it when you are reproducing an existing Codex or Claude Code feature in Qwen Code or another agent CLI. If you do not name the reference agent, the agent should ask you once before capture.

The agent works in the current directory by default. Give a path only when you want a different target repo.

  • "Reproduce the Codex slash command behaviour in Qwen Code."
  • "Use Claude Code as the reference and capture how its plan mode shows terminal output, then implement the same in this repo."
  • "Capture the headless output of codex exec for this feature and add a Qwen Code test for it."

Run a capture by hand (optional)

The agent normally runs these commands itself. Knowing them helps you read the output. Each run should use a fresh output folder under .repro-runs/. For headless Codex, the agent wraps the reference command in the mitm script. For Claude Code or interactive slash commands, it uses tmux.

State snapshots are taken before and after the scenario and then diffed. Use --agent claude-code to snapshot ~/.claude instead of ~/.codex.

  • Headless capture writes to a folder you choose, e.g. .repro-runs/slash-command-baseline
  • tmux capture is used for interactive or TUI-visible behavior
  • Use --root PATH only for a custom state directory or tests
mkdir -p .repro-runs/slash-command-baseline .qwen/skills/agent-reproduce-feature/scripts/run_with_mitm.sh \ .repro-runs/slash-command-baseline \ -- codex exec "exercise the Codex feature here" .qwen/skills/agent-reproduce-feature/scripts/capture_state.py \ snapshot .repro-runs/slash-command-baseline/state-before \ --agent codex

Review the output before you use it

Captured request bodies are sensitive local artifacts. They can contain prompts, code and credentials. Redact tokens before you copy any example into docs, commits, issues or pull requests. The state tool redacts common token shapes and omits content for sensitive paths, but you should still review state-diff.md before copying an excerpt into a tracked file.

The skill itself advises against copying all captured prompt text into Qwen Code. The goal is the minimum behavior, schema or test needed for parity.

  • Redact tokens in request bodies before sharing
  • Review state-diff.md before copying any excerpt
  • Keep the first implementation to one feature and one trigger path
  • Prefer tests that check behavior over tests that check exact prompt wording

Safety notes: scripts and network access

This skill runs scripts and uses the network. The capture scripts start a local mitmproxy and run the reference agent through it, so the reference agent's real model traffic passes through the proxy. Run it only on machines without unrelated secrets, and do not share captured traffic publicly without manual redaction.

If TLS fails, the skill's reference notes say to fix certificate trust before treating missing traffic as product behaviour.

  • Scripts: run_with_mitm.sh, run_tmux_capture.sh, capture_state.py, llm_dump.py
  • Network: traffic from the reference CLI goes through the local proxy
  • Do not use on machines holding unrelated secrets
  • Do not publish raw captures; they may contain code and credentials

What people use it for

  • Adding a Codex-style slash command to Qwen Code with the same terminal output
  • Matching how Claude Code builds its tool schemas for a specific feature in a Qwen Code module
  • Studying how a reference agent structures its prompts and request bodies for local analysis
  • Checking what local config files a reference agent writes during one scenario using state diffs

Frequently asked questions

How do I install the agent-reproduce-feature 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 before using this skill?

You need Python 3 with mitmproxy installed so the mitmdump command works, tmux for terminal capture, the reference CLI (codex or claude) installed locally, and a Qwen Code repository checkout. A mitmproxy CA certificate at ~/.mitmproxy is optional and only needed for TLS interception.

Is the agent-reproduce-feature skill free to use?

The skill is published under the Apache-2.0 licence. Check the licence file in the source repository for the terms that apply to your use.

What is the difference between this skill and a general coding helper?

This skill is narrow. It is built for matching a feature from Codex or Claude Code in Qwen Code by recording the reference tool's requests and terminal output. It is not meant as a general coding assistant unrelated to matching another agent's behaviour.

Why does the skill run scripts and use the network?

The capture scripts start a local mitmproxy and run the reference agent through it so its model requests can be recorded. Those requests go over the network, and the scripts run on your machine, so use it only where you are comfortable with that.

What should I do if the capture shows no traffic?

The skill's reference notes say that TLS failures can hide traffic. Fix certificate trust for the proxy first, then run the capture again, rather than assuming the missing traffic reflects the product's behaviour.

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

Use this skill if you maintain Qwen Code or a similar CLI agent and want to match a specific feature from Codex or Claude Code, or if you are studying how those tools build prompts and tool schemas. Pick something else if you want a general coding helper, because this workflow is narrow and needs local tools, scripts and network access to run.

Download the skill