The Agent Skills Directory

[State Management via Local Files]: The skill uses a local directory (.copilot/planning/) to store checkpoint files. This allows the planning process to be resumed if interrupted. The file path is derived from a sanitized task slug to ensure predictable behavior.
[Contextual Information Gathering]: To function, the skill reads source files and test definitions. This is a standard requirement for a development-focused planning tool to ensure the generated plans are relevant to the existing codebase.
[Multi-Agent Orchestration]: The process involves spawning multiple sub-agents (lenses) to provide diverse viewpoints. This design inherently includes cross-checks, such as the adversarial review loop, which helps identify potential flaws or edge cases in the synthesized plan.
[Indirect Prompt Injection Surface]: The skill ingests untrusted data from project files into the agent context. While this presents a surface for indirect prompt injection, the architectural design provides guardrails. Ingestion points: Source files and definitions read in Step 0. Boundary markers: The skill uses fresh sub-agents for each lens to minimize context contamination. Capability inventory: The skill performs file reads/writes and orchestrates sub-agents. Sanitization: The task-slug used for directory naming is sanitized to prevent path traversal issues. The mandatory adversarial review in Step 4 acts as a functional mitigation against planning errors or injected biases.

multi-perspective-planning