Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

1.9 KB
Raw Blame

Repository and code work, end to end, in the action sandbox (action_dir). Clone repos and write build output there; the internal workspace is off limits.

  1. Locate. grep / glob / list for symbols, error strings or files, then file_read the top hits. Two rounds of searching, then start editing.
  2. Edit. edit for a small exact replacement, apply_patch for multi-hunk changes, file_write only to create a new file. Never rewrite an existing file through a shell heredoc.
  3. Run. Use shell for builds, focused tests, and scripts. Node.js and Python commands use the host's node and python3 on PATH. curl is for a raw HTTP call when web_fetch is not enough. Only stdout/stderr come back: a script that prints nothing returned nothing, not success. If you create a virtualenv, use its own pip/python.
  4. Verify. run_tests and run_linter in the working tree; fix what fails before reporting.
  5. Review before you finish. read_diff, then check it in this order: security (injection, secrets, unsafe input) → correctness (edge cases, off-by-one, error paths, races) → tests cover the new paths → style matches the surrounding code. Name file and line for each problem.
  6. Git. git_operations (or shell with git) for status, diff, branch, commit and push in the working tree. GitHub state (issues, PRs, reviews, checks, comments) goes through the connected GitHub integration: tool_search for the action. Never shell gh.

Shell hygiene: plain commands, pipes and redirects are fine; avoid $(…), backticks and background &; run the inner command as its own step instead. No sudo or system package installs unless the user granted it: say what is missing and propose an alternative instead of looping on installers.

Report what you changed (paths), how you verified it (commands and results) and anything left undone.