The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a process for ingesting external data from GitHub issues and pull request comments, creating a potential surface for indirect instructions. However, the instructions provide strong mitigation through a dedicated 'Boundary' section that forbids the inclusion of private data, credentials, or local planning records in public output. \n- Ingestion points: The agent is instructed to read issue descriptions and PR comments via gh pr view. \n- Boundary markers: Explicit instructions define strict limits for public scope and staged diff inspection. \n- Capability inventory: The skill has access to git, npm ci, and gh CLI tools, and executes local scripts within the repository's tools/ directory. \n- Sanitization: The workflow requires verification of committer identity and manual inspection of the staged patch before committing.\n- [DYNAMIC_EXECUTION]: The skill uses inline Node.js scripts for environment validation and configuration parsing. \n- Evidence: It executes node -e to check for Node.js version 24+ and to parse gate-manifest.json for job identification. These are static, hardcoded utility scripts used for local validation.\n- [COMMAND_EXECUTION]: The workflow relies on standard shell commands to manage the contribution lifecycle. \n- Evidence: Commands include git worktree, git add, git commit, npm ci, and various execution scripts like tools/run-node-tests.mjs and tools/run-manifest-gates.mjs. These are typical for software development tasks within a trusted repository context.\n- [EXTERNAL_DOWNLOADS]: The skill automates the retrieval of code and project dependencies. \n- Evidence: It uses git clone to fetch the author's repository and npm ci for package installation from official registries. These actions are transparent and target the project's own infrastructure.

harness-contributing