The Agent Skills Directory

[DATA_EXFILTRATION]: The skill instructs the agent to read sensitive configuration files including ~/.claude/settings.json, ~/.codex/config.toml, and ~/.codex/auth.json. These files contain authentication tokens and API keys. The instructions direct the agent to retrieve these secrets and pass them as command-line arguments to a diagnostic script, exposing them to the process list and shell history.
[COMMAND_EXECUTION]: The skill requires the execution of multiple shell commands and scripts to diagnose the environment and apply fixes. This includes running the provided diagnose.py script, using python3 -c for JSON processing, and utilizing shell utilities like find, grep, and awk to inspect logs and session data.
[EXTERNAL_DOWNLOADS]: The skill recommends registering various external MCP (Model Context Protocol) servers from providers such as Exa, Tavily, Aliyun, and Tencent to provide replacement search capabilities. These references are documented to enable core functionality.
[REMOTE_CODE_EXECUTION]: The skill suggests using npx to execute the one-search-mcp package directly from a remote registry. This involves downloading and running third-party code in the local environment.
[PERSISTENCE]: The skill functions by instructing the agent to modify its own persistent configuration files to permanently alter tool availability and permissions. This allows configuration changes to persist across agent restarts and sessions.
[INDIRECT_PROMPT_INJECTION]: The verification workflow involves parsing agent log files (init.jsonl) and session rollout files (rollout-*.jsonl). These files contain output from tools and previous agent turns which could potentially be manipulated by external data sources to influence agent behavior.
Ingestion points: Reads from ~/.claude/settings.json, /tmp/init.jsonl, and session rollout files.
Boundary markers: No explicit boundary markers or instructions to ignore embedded commands are used in the parsing logic.
Capability inventory: Includes shell execution, file system writes, and network operations via diagnose.py and external MCP tools.
Sanitization: Relies on standard JSON parsing and regular expression matching without additional validation or escaping of the parsed content.

agent-web-search-setup