The Agent Skills Directory

[COMMAND_EXECUTION]: The skill extracts a next_command field from the JSON output of the specgate change status command and explicitly instructs the agent to "Run it verbatim". While the skill attempts to mitigate risk by requiring the command to be a single specgate CLI call without shell operators, it still allows a remote service to dictate which local commands the agent executes at runtime.
[REMOTE_CODE_EXECUTION]: Because the instructions to execute specific commands are provided by the specgate service's state data, this creates a mechanism for remote logic to drive local agent behavior. Additionally, the --run-checks flag in section 5 indicates that the tool executes shell commands defined in a completion scaffold.
[DATA_EXFILTRATION]: The skill collects and transmits local project metadata to the specgate service. This includes repository remote URLs, branch names, current head revisions (git rev-parse HEAD), and file change statuses (git status --short). This data is sent via the submit and report subcommands.
[PROMPT_INJECTION]: The skill processes untrusted external data in the form of guidance and missing fields from the CLI status. These fields are used to determine the agent's next steps and logic flow. 1. Ingestion point: Data enters via specgate work context and specgate change status. 2. Boundary markers: None provided for the interpolated guidance text. 3. Capability inventory: Subprocess calls to specgate and shell-based check execution. 4. Sanitization: No sanitization is performed; the agent is told to carry the guidance verbatim through later steps.

specgate-work-delivery