The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes a local Python script (diagnose.py) and a replay analysis script to process session data and generate reports. This is intended functionality for diagnosing plugin performance.
[DATA_EXFILTRATION]: The skill reads session logs, outcome archives, and state files from the /tmp/drupal-workflow-states/ directory. This data is used locally for analysis; no network transmission or access to sensitive credentials like SSH keys or API tokens was observed.
[PROMPT_INJECTION]: The skill possesses an indirect prompt injection surface because it processes session logs that could contain content influenced by external inputs.
Ingestion points: The diagnose.py script reads interventions.log, outcomes.jsonl, and state.json from state directories.
Boundary markers: Output is presented as structured data, but there are no explicit delimiters to isolate untrusted text within the logs.
Capability inventory: The agent can modify local scripts (task-classifier.sh, autopilot-monitor.sh) and policy files based on the analysis.
Sanitization: No explicit sanitization of log content is performed. However, the requirement for user approval before applying changes is a significant mitigation.

autopilot-tuner