The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts run_with_mitm.sh and run_tmux_capture.sh execute user-provided commands (using the $@ pattern) to exercise and record the behavior of reference agent implementations.
[DATA_EXFILTRATION]: The skill snapshots local agent state from sensitive directories including ~/.codex and ~/.claude using capture_state.py. Although the script implements extensive regex-based redaction for tokens, certificates, and API keys (e.g., OpenAI, GitHub, AWS), it captures local file metadata and content for analysis in the .repro-runs/ directory.
[EXTERNAL_DOWNLOADS]: The capture workflow documentation recommends installing the third-party utility mitmproxy via the standard Python package registry to facilitate HTTP traffic interception.
[REMOTE_CODE_EXECUTION]: The run_with_mitm.sh script starts mitmdump with the --set ssl_insecure=true flag, which disables TLS verification for intercepted traffic during the capture session to allow for local debugging.
[PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface as it processes untrusted data (prompts, tool schemas, and terminal outputs) captured from external agents to guide the reproduction of features.
Ingestion points: Captured logs in http.jsonl, tmux-pane.txt, and state-diff.md described in references/capture-workflow.md.
Boundary markers: Not present in the reproduction instructions; the agent is directed to extract behavioral facts directly from the traces.
Capability inventory: The skill can execute shell commands via capture scripts and perform local filesystem operations.
Sanitization: Robust regex-based redaction of tokens and keys is implemented in capture_state.py and llm_dump.py to strip sensitive information from captured records.

agent-reproduce-feature