The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from GitHub issues and local project documentation to drive the implementation workflow.
Ingestion points: The agent reads external content via gh issue view (Phase 4) and scans markdown files in the .specs/ directory (Phase 6).
Boundary markers: The orchestrator uses structured prompt templates and role definitions to separate worker instructions from the provided data, though it does not implement explicit sanitization for the issue content.
Capability inventory: The skill has the capability to modify the file system, execute shell commands, manage Git branches, and interact with GitHub APIs through the CLI.
Sanitization: No explicit filtering or escaping of the GitHub issue body is performed before it is interpolated into sub-agent prompts.
[COMMAND_EXECUTION]: The skill uses a variety of shell commands and CLI tools to manage the repository and coordinate agents.
Evidence: Utilizes git for feature branch management and gh for retrieving issues and creating pull requests.
Evidence: Employs standard Linux utilities like find, ls, grep, sed, yq, and awk for navigating the file system and parsing configuration files.
Evidence: The skill invokes the claude and codex CLIs within a cmux environment, employing automation flags such as --dangerously-skip-permissions and --dangerously-bypass-approvals-and-sandbox to bypass interactive prompts.
[EXTERNAL_DOWNLOADS]: The orchestrator relies on specialized worker skills and coordination scripts.
Evidence: Delegates tasks to worker skills spec-code, spec-review, and spec-test, which are part of the anyoneanderson/agent-skills ecosystem.
Evidence: Depends on a local script agent-delegate.sh to handle cross-AI task routing and peer review coordination.

spec-implement