The Agent Skills Directory

[DATA_EXFILTRATION]: The skill reads local agent transcripts from the agent-transcripts/ directory to facilitate session resumption, auditing, and evaluation. It includes explicit safety instructions to prevent the agent from globbing across unrelated projects or workspace boundaries, effectively mitigating accidental data exposure.
[COMMAND_EXECUTION]: The skill leverages internal repository scripts to automate maintenance tasks. Examples include scripts/worktree-audit.sh for disk space management, scripts/watch-pr/watch-pr for monitoring forge status, and scripts/orch/orch.ts for managing complex, multi-day programs.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted external content, such as pull request comments and review-bot findings. It mitigates this risk through the bugbot-triage.md reference and babysit.md playbook, which instruct the agent to triage such data skeptically and never treat it as direct instructions.
[DYNAMIC_CONTEXT_INJECTION]: The host adapter documentation for OpenCode mentions support for !code shell blocks in commands. This is a platform-specific feature for dynamic content generation, documented here to ensure correct interaction with the host environment.

poteto-mode