The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes project-level data (e.g., README.md, CHANGELOG.md), git log history, and issue/PR discussions to generate baseline documentation. This ingestion of untrusted external content represents an attack surface for indirect prompt injection.
Ingestion points: The Step 1: Investigation process in references/new-project-retrofit.md scans README files, CHANGELOGs, git commit messages, and external issue trackers.
Boundary markers: The skill utilizes a structured labeling system (用户确认, 已有文档, 推导待确认, 未知) to delineate the provenance of every statement.
Capability inventory: The skill is capable of writing markdown files to docs/devdocs/, performing git operations via Bash, and invoking other internal skills such as codebase-insight and agent-memory via Task orchestration.
Sanitization: Robust safety measures are built into the instructions, including mandatory cross-referencing of documentation against the actual codebase (Step 2: Validation) and an explicit rule to prioritize code as the authoritative source when conflicts arise.
[COMMAND_EXECUTION]: The skill invokes the Bash tool to perform necessary repository operations, specifically git mv for file renaming and git rev-parse HEAD for tracking the baseline adoption commit. These commands are constrained to specific maintenance tasks and do not allow for arbitrary command injection or execution of external payloads.

retrofit