The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes output from external agents, representing an indirect prompt injection surface. This is effectively mitigated by core architectural invariants requiring the agent to treat all delegated results as proposals and to verify them against local evidence and tests before relaying or using them.\n
Ingestion points: Output from claude, codex, and cursor-agent CLI tools as described in SKILL.md and reference files.\n
Boundary markers: The skill instructs the agent to brief sub-agents as new hires with no context carry-over and enforces a 'verify-before-trust' rule in SKILL.md.\n
Capability inventory: The skill utilizes scoped file system access, network access via external CLIs, and subprocess execution for the delegated agents.\n
Sanitization: Verification involves reconciling claims against local code, spot-checking diffs, and running existing test suites.\n- [DATA_EXFILTRATION]: The skill includes proactive safety measures for secret management, instructing the agent to scan for credentials in files or diffs before piping them to external tools and requiring explicit user sign-off if the risk of secret exposure is unclear.\n- [COMMAND_EXECUTION]: The skill executes trusted CLI tools from well-known services (Claude Code, Codex, Cursor) and a local diagnostic script (scripts/codex-doctor-check.sh). These operations are conducted using safety-oriented flags like --permission-mode plan and --sandbox read-only, while explicitly prohibiting the use of force or bypass flags without user consent.

cli-delegation