How to use the Competition Runtime Routing agent skill

What an agent skill is and what this one does

An agent skill is a folder of instructions and reference files that an AI agent loads when a request matches the skill's description. This skill is a downstream helper: it is meant to run only after the parent skill ctf-sandbox-orchestrator has already set up the challenge context and sent the work here.

Once active, the agent maps how a request travels through reverse proxies, Host and X-Forwarded-* headers, path-prefix rewrites and websocket upgrades to reach the right backend. It looks for the smallest change to a request that alters routing, then records the evidence as one compact host-to-proxy-to-service-to-process map.

• Builds a route map: client host and scheme, proxy rule, service or container, process, then any downstream store or worker.
• Changes one routing input at a time (Host, forwarded headers, Origin, path prefix, base URL) and compares the results.
• Separates route resolution from application authentication logic.
• Replies in Simplified Chinese unless you explicitly ask for English.

Install the skill from the ZIP

Download the skill ZIP from this page and unzip it. The folder contains SKILL.md, references/runtime-routing.md and agents/openai.yaml. Place the folder in one of the locations below so your agent can find it.

You can also upload it through the Claude apps: open Settings, go to Capabilities, then Skills, and upload the skill there.

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

Install the parent skill too

This skill does not work on its own as intended. Its SKILL.md says to use it only after $ctf-sandbox-orchestrator is already active and has set sandbox assumptions. The parent skill is referenced by name and is not bundled in this download, so you need to have it available separately.

Trigger the skill with natural requests

The skill's description lists the kinds of questions it handles. You do not need special syntax; ask the way you would ask a colleague, once the parent skill has set up the challenge. The routing skill may also point you to a different skill if the problem is about parser differentials or request smuggling.

Good triggers focus on which node serves a route, or how a header changes behavior.

• Which container serves this route on the sandbox?
• Why does this public-looking domain still belong to the sandbox?
• How does changing X-Forwarded-Host affect which backend answers?
• How does this route resolve across the proxy, the container and the worker?
• Does the websocket upgrade go to the same service as the normal HTTP path?

Example: tracing a route in a challenge

Suppose a challenge exposes a site on a hostname that looks like a public cloud domain. Ask the agent which service answers the request for the /app prefix. The skill tells the agent to treat the hostname as a sandbox fixture unless the evidence disproves it, then to build the route map from the client host to the proxy rule, to the container and process.

Next, it re-sends the same request with one input changed, for example the Host header or a path prefix. If the decisive backend changes, the agent reports the exact request shape that flips the route, with the proxy snippet and headers that support it.

• Provide proxy configs, container details or traffic traces if you have them, since the agent needs them for analysis.
• Expect a host-to-proxy-to-service-to-process map as the output.

Safety notes and limits

This skill runs no scripts and makes no network calls, according to its metadata. It works from the files and evidence you give it, such as proxy configs, container details and traffic traces.

The skill tells the agent to treat every hostname as a sandbox fixture unless evidence disproves it. That is appropriate for CTF practice but can mislead a real investigation, so do not use it on production systems or real third-party domains. Parser-differential and request smuggling problems are out of scope and belong to a different skill.

• Do not use on production systems or real third-party domains.
• Scripts: none. Network: none.
• Requires the parent ctf-sandbox-orchestrator skill to be active first.

Tips for better results

Give the agent one complete request and its response before asking broader questions. The reference checklist advises proving one route end to end before widening to other hosts or prefixes.

Keep checked-in config separate from live config when they differ, and say which one you are asking about. Capture a real request as evidence rather than relying only on reading the code.

• Start with one request that you can reproduce.
• Change one routing input at a time.
• Note whether a failure is a routing problem or an auth problem; they are different.