The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data, including code diffs, acceptance criteria, and PR metadata, which are interpolated into instruction blocks for the main agent and several spawned subagents.
Ingestion points: Code diffs are mapped via Explore subagents (Phase 1); story documentation and acceptance criteria are read for traceability (Phase 2A); and PR comments are parsed for status selection (Step 6).
Boundary markers: The skill lacks strong delimiters to segregate external data from instructions in subagent prompts, although it uses structured JSON for some internal state and result reporting.
Capability inventory: Spawns multiple parallel subagents using the Task and Agent tools; executes shell commands via spawnSync and execFileSync; performs network operations via curl and the gh CLI; modifies filesystem artifacts like reports and gate files.
Sanitization: Implements an allow-list of safe shell commands in qa-execute-snippets.mjs to filter executed code snippets.- [DYNAMIC_EXECUTION]: The skill is designed to extract and execute bash code snippets found within markdown fences in SKILL.md and other prompt documents (Phase 1.7). It also uses node -e for logic execution and dynamic variable resolution in shell scripts.- [COMMAND_EXECUTION]: Extensive use of shell commands through system utilities such as gh, git, jq, node, sed, and awk to automate QA processes and update document metadata.

qa-story