The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local developer tools including git, bun, and gh to analyze repository state and manage the submission pipeline. The execution is handled via node:child_process.spawnSync and node:child_process.execFile using array-based arguments, which prevents shell command injection. Input parameters such as branch references are validated against a specific whitelist (origin/dev, origin/minor, origin/major).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the repository, such as project manifests (package.json) and release notes (.changeset/*.md), to inform its planning logic.
Ingestion points: Files are read in scripts/change-plan.ts (manifest name/privacy checks) and scripts/changeset-plan.ts (workspace discovery and changeset frontmatter parsing).
Boundary markers: The workflow is governed by structured reference documentation (references/*.md) that separates analysis, verification, and submission phases.
Capability inventory: The agent has capabilities to read/write repository files, perform Git operations (show, diff, rebase, push), and execute project CLI scripts.
Sanitization: Paths are normalized to prevent traversal, and critical branch names are strictly validated before being passed to Git commands.

seed-change