The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill encourages the recording of sensitive system state and I/O operations which can expose secrets.
Evidence in references/non-determinism.md: Recommends recording 'Environment variables', 'File read operations (content)', and 'System state'.
Evidence in references/python-replay.md: Provides code to wrap __builtins__.open and record all read/write operations to log files.
[COMMAND_EXECUTION]: The skill utilizes invasive instrumentation techniques to modify program behavior at runtime.
Evidence in references/python-replay.md: Demonstrates monkey-patching core Python modules and built-ins, including time.time, random.random, and open.
[DATA_EXFILTRATION]: The skill implements logic to capture and store execution data, creating a risk of sensitive data exposure.
Evidence in SKILL.md: Code snippets show recording function results and I/O to a local replay.log file without any sanitization or encryption.
[REMOTE_CODE_EXECUTION]: The skill recommends using unsafe deserialization methods for handling log data.
Evidence in references/replay-tools.md: The ReplayEngine class example uses pickle.load() to load event logs from disk. This is a known security vulnerability that allows for arbitrary code execution if the log file is sourced from an untrusted party.
[PROMPT_INJECTION]: The skill has an indirect injection surface as it is designed to process and replay external data (logs).
Ingestion points: replay.log, recordings.json.
Capability inventory: The skill has the ability to read/write files and execute shell commands (via tools like rr).
Sanitization: No evidence of sanitization or validation of the log file contents before processing.

replay-oriented-instrumentation