The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill orchestrates a multi-agent swarm that ingests untrusted data from project files and external research, creating a surface for indirect instructions.
Ingestion points: The orchestrator utilizes knowledge-connect.skill to retrieve external data snippets for task context during Phase S05, and orch-researcher for discovery in Phase S00.
Boundary markers: The framework enforces mandatory "HUMAN APPROVAL REQUIRED" gates at critical SDLC transitions (S00, S01, S02, S07, S08) to prevent autonomous phase advancement without user verification.
Capability inventory: The system has extensive access to the local filesystem for state preservation in agents/memory/, autonomous code remediation tools, and deployment scripts.
Sanitization: The instructions mandate the use of safety-scan.skill and ai-safety-guardrails.skill (Phase S07) to verify inputs and sub-agent security postures.
[DYNAMIC_EXECUTION]: The orchestration loop includes the meta-optimize.skill (Phase S05), which is described as being able to "Rewrite the failing skill/prompt" at runtime as part of a "Self-Evolution" mechanism for failed tasks.
[COMMAND_EXECUTION]: The implementation phase (S05) employs autonomous remediation tools like ci-autofix and self-heal.skill to correct code errors and satisfy compiler requirements during build cycles.
[DATA_EXFILTRATION]: The skill describes "Tactic B" for subagent delegation, which involves generating a delegation-payload.md file intended to be manually transferred by the user to external, isolated LLM instances. While this process is human-mediated, it represents a design pattern for moving internal context to external environments.

loki-mode