The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the local comet CLI to perform its functions. It executes a wide range of commands for state management (comet state), workspace isolation (comet classic workspace), and phase transitions (comet guard). It also facilitates the execution of build and test suites via comet check run, which is an expected capability for a tool intended to assist with software development tasks.
[DYNAMIC_EXECUTION]: The skill includes instructions to execute commands dynamically generated by the CLI tool. Specifically, the comet classic openspec --agent-json command can return a nextAction object containing an argv and cwd that the agent is instructed to execute directly. This dynamic command assembly is part of the integrated workflow between the agent and the supporting CLI.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection as it processes untrusted data from the repository (e.g., tasks.md, proposal.md, source code) and user requests. 1. Ingestion points: Processes project artifacts like requirement files, design documents, and user utterances to determine workflow routing. 2. Boundary markers: The skill uses structured JSON schemas (e.g., comet.intent.v1) and formal protocol files to define data expectations. 3. Capability inventory: The agent can read/write files and execute local shell commands (CLI and build tools). 4. Sanitization: The instructions emphasize relying on the comet Runtime to validate data and determine routing, reducing the risk of the agent autonomously following malicious instructions embedded in data.

comet-classic