The Agent Skills Directory

[PERSISTENCE]: The skill includes an unattended operation watchdog mechanism designed to maintain agent sessions over long periods.
Evidence: assets/install-watchdog.sh and assets/watchdog.sh install and manage systemd user units (orchestrator-watchdog.service and orchestrator-watchdog.timer) to periodically nudge the agent session using opencode run.
The installer also suggests using sudo loginctl enable-linger to ensure the persistence mechanism continues to run while the user is logged out.
[COMMAND_EXECUTION]: The framework relies on executing arbitrary system commands defined in a config.json file for operations like setup, formatting, analysis, and running tests.
Evidence: scripts/autonomous/common.py (via the run and git functions), scripts/autonomous/mutation_engine.py (executing test_cmd), and scripts/autonomous/orchestrator.py all use subprocess.run or subprocess.Popen to execute command-line arguments sourced from the project configuration.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a broad attack surface for indirect prompt injection as it is designed to ingest and analyze arbitrary source code and test files from target repositories.
Ingestion points: scripts/autonomous/mutation_engine.py reads source files for mutation, and the adversarial audit logic (Stream B) involves the agent reading test files to identify weaknesses.
Boundary markers: No explicit delimiters or warnings are used to prevent the agent from following instructions embedded in the code comments or strings of audited files.
Capability inventory: The agent can write to files (mutations), execute arbitrary shell commands (test runners), and interact with git remotes or GitHub PRs via the gh tool.
Sanitization: There is no evidence of sanitization or filtering of the repository content before it is processed by the agent.

autonomous-improvement