The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to capture "durable lessons" from recent work and store them in docs/solutions/ to improve future planning and implementation. This creates a surface for indirect prompt injection.
Ingestion points: The skill reads potentially untrusted content from git log, git diff, gh pr diff, and GitHub issue threads to identify what to capture (SKILL.md, Phase 1).
Boundary markers: The generated solution documents do not use explicit delimiters or "ignore embedded instructions" warnings when incorporating or summarizing external content.
Capability inventory: The skill has the capability to write files, perform git commit/push operations, and create GitHub issues via gh issue create (SKILL.md, Phase 5).
Sanitization: The skill relies on the agent's internal summarization and classification logic rather than explicit sanitization, escaping, or schema validation of the ingested external data.
[COMMAND_EXECUTION]: The skill executes a bundled bash script (references/sidecar.sh) and standard CLI tools to manage project state, configuration, and documentation.
Evidence: Frequent calls to bash "$SC" with various subcommands including base-ref, config, path, and commit.
Context: These commands are used to interact with a "sidecar" repository—a private local git repo used for storing metadata, research archives, and solution entries outside the main project tree.

compound