The Agent Skills Directory

[Command Execution]: The skill makes extensive use of system commands and command-line tools like gh, git, and uv to manage the repository, run tests, and interact with the GitHub API. This is standard behavior for a development automation tool.
[Indirect Prompt Injection]: The skill processes untrusted data from GitHub issues (e.g., ISSUE_BODY_OR_REPRO in assets/triage-prompt.md). This input is used to guide the creation of reproduction scripts and code fixes. This is a common surface area for automation agents.
Ingestion points: External issue descriptions are ingested into subagent prompts (such as triage-prompt.md and strategic-alignment-prompt.md) to guide the automation pipeline.
Boundary markers: The instructions use markdown headers and specific field labels to separate issue data from control instructions.
Capability inventory: Across its scripts, the skill maintains capabilities for repository management, file modification, and test execution using tools like gh, git, and uv.
Sanitization: The agent is specifically instructed to extract minimal reproductions and follow a Test-Driven Development (TDD) process to verify that generated code behaves as expected.
[Dynamic Execution and Testing]: To implement fixes, the skill dynamically generates test code and modifies project files based on the reported issues. It then executes these tests using uv run. This is a core part of its bug-fixing capability and is managed within dedicated worktrees to minimize impact on the main repository state.
[Governance and Access Control]: A notable security feature is the requirement for 'fresh human-scope evidence' retrieved via a local governance script (scripts/governance/eligibility.cjs). This ensures that implementation only proceeds when authorized by a responsible human actor, providing a strong gate against unauthorized automation.
[Assignment and Ownership Gates]: The skill implements a 'sole human assignee' check to prevent unauthorized or conflicting work on issues. It uses the GitHub API to verify that it is the only human actor assigned to an issue before proceeding with any implementation writes.

autopilot-issue-delivery-worker