The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands to determine the project state.
Evidence: Uses git log --oneline -20 and git describe --tags --abbrev=0 to extract versioning and recent change information.
[DATA_EXFILTRATION]: The skill reads project files and metadata to perform its analysis.
Evidence: Accesses ROADMAP.md, package.json, pyproject.toml, Cargo.toml, and various source code files via glob patterns. All data remains within the local context, with no network operations detected.
[EXTERNAL_DOWNLOADS]: The skill mentions an external Node.js utility in its documentation.
Evidence: References npx roadmapsmith@0.14 in a warning message (Step 2c) as a suggestion for users with legacy blocks, but the skill does not automate this execution.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect injection via untrusted repository data.
Ingestion points: Reads ROADMAP.md, source code files via glob, and git log entries (file: SKILL.md, Step 2 & 3).
Boundary markers: None explicitly defined for interpolated data.
Capability inventory: Uses Edit tool for file writes and shell execution for git commands.
Sanitization: None specified for handling external commit messages or task text. However, the risk is mitigated by the mandatory human-in-the-loop requirement (Step 6) where the user must approve a diff before any write occurs.

roadmap-update