github.get_pr_to_pass_local_tests

Goal

• Run, monitor, and fix the local unit tests for the current PR

Workflow

Create and Update a Plan

• Create a file plan-github.get_pr_to_pass_local_tests.md with a plan in the form of a bullet list of actions and maintain it updated, by marking each action
  • [.] when something is in progress
  • [x] when something is done
  • [F] when something failed

Get the PR Number and Branch Name

• Get the PR number

  > GH_PR_NUM=$(gh pr view --json number -q .number)
  

• Get the branch name

  > BRANCH_NAME=$(git branch --show-current)
  

Run Local Regressions

• Run full test suite locally:

  > pytest_multi_build.py --target . 2>&1 --timeout -1 | tee tmp.auto_task.pytest_multi_build.log
  

• Wait for the script to terminate

Check Regressions to Complete

• Check the output of the local regressions with:

  > pytest_failed_multi_build.py 2>&1 | tee tmp.auto_task.pytest_failed_multi_build.log
  

Report Status on the PR

• If local regressions are passing:

  > gh pr comment $GH_PR_NUM --body "Local tests passing"
  

• If any failures, document error and post:

  > gh pr comment $GH_PR_NUM --body "Local test failures found: [error summary]. Investigating..."
  

Fix the Failures

• If failures were found, run tmp.pytest_failed_multi_build.repro.sh to understand what failed and use /pytest.fix_local_tests to analyze and fix them
  • If the fix is simple, just fix it and commit
  • If the fix is not clear, stop and ask for the user to help

Never Commit Junk Files

• Never run git add -A or git add . to stage a fix: it sweeps in every untracked file sitting in the working tree (logs, tmp.* scratch files, test-run artifacts like .pkl/.json caches, etc.), not just the files you intended to fix
• Always stage files explicitly by path, e.g. git add <file1> <file2>
• Before committing, run git status --short and review every listed file; drop anything that is not part of the intended fix
• If junk files were already committed and pushed, fix it by amending the commit to contain only the intended files and force-pushing (git push --force-with-lease), rather than leaving the junk in history

Loop

• Keep repeating until the PR is passing all the local tests

Constraints

• This skill runs both when executed locally on a dev computer and remotely on cloud (e.g., on GitHub or Anthropic infrastructure)

Verification

• [ ] pytest_failed_multi_build.py reports no failing tests
• [ ] A PR comment reports the final local test status
• [ ] git status --short shows only the intended files staged before commit