The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill functions as a pipeline manager that ingests untrusted data from external files to drive subsequent agent actions.
Ingestion points: The orchestrator reads project requirements and task lists from the project-specs/ and project-tasks/ directories (specifically using cat and grep in SKILL.md).
Boundary markers: Instructions for passing context to spawned agents (e.g., PM, Architect, Dev) lack explicit delimiters or instructions to ignore embedded malicious directives within the specs.
Capability inventory: The skill possesses the ability to spawn various specialist agents and execute shell commands (ls, cat, grep) to manage the workflow.
Sanitization: There is no evidence of content validation or escaping when requirements from specification files are interpolated into the instructions for sub-agents.
[COMMAND_EXECUTION]: The skill uses shell commands within its workflow phases to manage state and verify deliverables.
Evidence: The workflow includes ls -la, cat, and grep commands to check for the existence and content of specification files and task lists. While these are used for legitimate process tracking, they represent the skill's capability to interact with the local file system.

agents-orchestrator