The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill implements a workflow that relies on external data ingested through a local collector server (scripts/debug-server.mjs).
Ingestion points: The agent is instructed to run a server that accepts JSON payloads via POST /log. The data collected is then used in the 'Analyze' step to confirm or reject hypotheses.
Boundary markers: While the skill defines a structured JSON schema for logs, it lacks robust boundary markers to differentiate between intended log data and malicious instructions that might be present in the application's output during debugging.
Capability inventory: The agent has the ability to modify source code ('Fix' step) and execute arbitrary reproduction commands ('Feedback command' step), making the consequences of an injection potentially impactful.
Sanitization: The collector server performs basic JSON validation but does not sanitize the contents of the log fields, leaving the agent vulnerable to instructions embedded in the 'observed' data.
[DATA_EXFILTRATION]: The debugging process involves logging application state to a local file, which creates a risk of sensitive data exposure.
The skill instructs the agent to capture context 'without leaking secrets,' but this is a manual guideline prone to error during complex debugging sessions.
Sensitive data such as API keys, environment variables, or PII could be captured in the <debug-events.jsonl> file.
[COMMAND_EXECUTION]: The skill requires the agent to execute a local Node.js script (scripts/debug-server.mjs) and potentially arbitrary commands for reproduction scripts.
While these are part of the intended debugging workflow, they provide a broad surface area for command execution based on agent-generated or user-provided parameters.

debug-mode