How to use the Competition Queue Worker Drift agent skill

What does the skill make the agent do?

When active, the skill guides the agent to follow a task from the moment something is put onto a queue, through the background worker that processes it, to the final effect. It covers queues, async workers, cron jobs, delayed tasks, retry behaviour and worker-only configuration drift.

The agent is asked to record queue names, payload shapes, retry and dead-letter behaviour, and any settings that exist only in the worker environment. It then reduces these findings to one minimal enqueue-to-side-effect reproduction chain rather than a long list of observations.

• Map the enqueue point, queue payload, worker consumer, retries and final side effect.
• Keep request-time state separate from worker-time state.
• Compare the synchronous path with the async path when behaviour diverges.
• Reproduce the smallest flow that proves the async difference.

What you need before installing

This skill does not work on its own as described. The parent skill $ctf-sandbox-orchestrator must be installed and active, and it must have established the sandbox assumptions first. The skill tells the agent to return to that parent if this has not happened.

You also need access to the CTF sandbox target and its source or logs. The skill itself does not name specific software, so what you need depends on your challenge.

• Parent skill: $ctf-sandbox-orchestrator, installed and active.
• Access to the sandbox target and its source code or logs.
• A sandboxed competition target you are authorised to test.

How do you install the skill?

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

If you use the Claude apps instead, you can upload the ZIP under Settings → Capabilities → Skills. Either way, the skill folder should contain SKILL.md, references/queue-worker-drift.md and agents/openai.yaml.

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

How do you trigger the skill with natural requests?

You do not need to type the skill name. Describe the problem in plain language and the agent should match the request to the skill. The skill description names these situations: tracing a queue payload, inspecting async job execution, explaining worker-only behaviour, following retries or dead-letter handling, and connecting an enqueued item to a later file, cache, email or privilege-bearing side effect.

Remember that the parent skill must be active first. Starting with the parent gives the sandbox assumptions the downstream skill expects.

• "Trace what happens to this queue payload after it is enqueued."
• "Why does this job succeed when run by the worker but not in the request?"
• "Follow the retry and dead-letter path for this task."
• "Connect this enqueued item to the file it eventually writes."

What does a typical analysis look like?

Suppose a challenge only changes state when a delayed job runs. You ask the agent to trace the payload. The skill has the agent record the queue name, the message shape, the worker process and its environment, and the retry policy, then identify the downstream store where the effect lands.

The output should be a single chain such as enqueue, then worker runtime, then retry or branch, then resulting effect. The skill also asks the agent to state whether the key difference is in the payload shape, the worker config, the retry path or the downstream consumer.

What safety rules apply?

The skill is written for sandboxed competition targets. Its own listing says it is not for production systems or real third-party infrastructure, because tracing privilege-bearing effects on systems you do not own could be unauthorised access.

According to the listing, the skill runs no scripts and uses no network. It works by directing the agent's reasoning and the evidence you provide, so the access you give the agent to the sandbox is the main control you have. By default it replies in Simplified Chinese unless you explicitly ask for English.

• Use it only on sandboxed competition targets you are authorised to test.
• Do not point it at production systems or third-party infrastructure.
• The skill runs no scripts and uses no network according to its listing.
• Replies default to Simplified Chinese; ask for English explicitly if you need it.

Tips for better results

Give the agent the enqueue point, the worker identity and any worker-only environment variables, mounts or credentials up front. The skill's checklist treats these as the evidence most often missed. Ignoring worker-only environment or mount differences is a listed common pitfall.

Do not stop at the request path. The reference checklist warns against explaining only the request path without proving the worker branch.

• Keep enqueue request, stored payload and worker identity tied together.
• Record retry count, backoff, dedupe keys and idempotency behaviour.
• Separate immediate request success from eventual worker success or failure.
• If the real problem is a file parser run by the worker, the skill says to switch to the tighter file-parser skill.