multi-perspective-plan
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- Indirect Prompt Injection Surface: The skill's primary function involves gathering context by reading local source files and tests. This process introduces external, potentially untrusted data into the agent's reasoning loop, which could influence the subsequent planning subagents if the source files contain adversarial content.
- Ingestion points: Step 1 ("Gather Context") explicitly reads source files and tests to create a self-contained context block in
.copilot/plans/<slug>/00-context.md. - Boundary markers: The instructions do not specify the use of delimiters or explicit instructions for subagents to disregard embedded directives within the ingested source code.
- Capability inventory: The skill manages the spawning of multiple subagents (speed, usability, conservative, synthesis, and adversarial lenses) and performs extensive file-write operations to the
.copilot/directory for persistence. - Sanitization: There are no mentioned mechanisms for sanitizing or filtering the content gathered from the local filesystem before it is processed by the LLM subagents.
- Local Persistence and File System Access: The skill creates a structured directory hierarchy under
.copilot/plans/. While this involves writing multiple files to the local disk, it is a transparent persistence mechanism aligned with the skill's goal of resumable planning and uses a standard directory associated with the vendor's ecosystem.
Audit Metadata