The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes Git, Codex, and validation scripts using node:child_process (spawn and execFileSync). It mitigates risks by explicitly disabling shell execution and enforcing a strict allowlist of Git subcommands (e.g., rev-parse, worktree, diff).- [DATA_EXFILTRATION]: The skill handles sensitive operational data including plans and intake snapshots. It protects this data using a capability-gated authorization system where tokens are stored in owner-read-only files (mode 0600) in directories isolated from the agents' task worktrees to prevent accidental exposure.- [PROMPT_INJECTION]: The skill processes untrusted feature requests. It mitigates potential indirect prompt injection attacks by verifying that all filesystem changes remain within a defined writeScope (using git-workspaces.verifyScope) and by utilizing a quorum-based adversarial verification panel to confirm defects before they are accepted as actionable items.- [SAFE]: The skill implements several security best practices, including immutable termination sets (budgeting) to prevent resource exhaustion, mandatory human approval for plans before execution, and a deterministic state machine that prevents phases from advancing past blocked required gates.

autonomous-pipeline