The Agent Skills Directory

[COMMAND_EXECUTION]: The skill manages external processes to facilitate remote logging. Specifically, it references stopping 'cloudflared' processes, which are used to establish network tunnels to Cloudflare's well-known service infrastructure.
[DATA_EXFILTRATION]: The skill is designed to capture runtime state and variables ('vars') and transmit them to an external endpoint defined by 'LOG_SERVER_URL/log'. While intended for debugging, this pattern involves sending potentially sensitive program data to a remote server.
[DYNAMIC_EXECUTION]: The workflow requires the agent to modify the application's source code at runtime to insert 'probes' (temporary logging functions). These modifications are used to verify hypotheses and are intended to be removed during the cleanup phase.
[INDIRECT_PROMPT_INJECTION]: The skill processes data from 'LOG_FILE', which contains NDJSON records generated from instrumented code. This introduces a vulnerability surface where malicious runtime data could influence the agent's analysis.
Ingestion points: The agent reads the 'LOG_FILE' in 'references/prove-fix-verify.md' to analyze bug causes.
Boundary markers: Uses 'ATTEMPT_START_LINE' to isolate records belonging to the current session.
Capability inventory: Includes source code modification, network logging, and process management (e.g., cloudflared).
Sanitization: Instructions in 'references/probes.md' advise conducting a 'privacy-review' of events and redacting sensitive values before logging.

debug-mode