/test-branch

Run the test-branch script on all commits in the current branch. When a build failure occurs, use a fresh subagent to fix the error, create a fixup commit, rebase, and retry until all commits pass.

Overview

This command automates the test-fix loop:

1. Run scripts/test-branch to test each commit
2. If a commit fails:
   • Extract the error from the build output
   • Launch a subagent to fix the error
   • Run scripts/test-fix to create fixup commit and rebase
   • Repeat from step 1
3. If all commits pass, report success

Task

Execute the following loop:

1. Start the test run in background:

   • Run scripts/test-branch with run_in_background: true
   • Save the shell ID for monitoring

2. Monitor the output:

   • Use BashOutput with a filter regex to capture only relevant lines:
     • Filter: "FAILURE|SUCCESS|ERROR|error:|warning:|✗|BAD COMMIT|GOOD COMMIT|Recipe.*failed"
   • Poll periodically until the shell completes
   • Accumulate the filtered output

3. When the shell completes:

   • If exit code is 0:
     • Report: "All commits pass!"
     • Exit
   • If exit code is non-zero:
     • Filtered output now contains errors but not the full 10MB+ log
     • Use the Task tool to launch the test-branch-fixer agent with:
       • The filtered error output
       • The current git status
       • The commit SHA that failed (visible in the "BAD COMMIT" line)
     • Wait for the agent to complete

4. Handle the agent result:

   • If the agent reports it cannot fix without user input:
     • Show the error to the user
     • Ask: "I need help fixing this error. Should I: (a) Skip this commit, (b) Wait for you to fix it manually, or (c) Try a different approach?"
     • Exit and let the user decide
   • If the agent successfully fixed the code:
     • Verify there are unstaged changes with git status --porcelain
     • Run scripts/test-fix (NOT in background)
     • Go back to step 1

5. Safety limit: If more than 10 iterations occur, report: "Too many iterations. Please review manually." and exit

Notes

• The BashOutput filter is critical: it extracts errors without consuming the entire build log
• Each fix uses a fresh subagent for clean context
• The agent may report it needs user input; respect that and exit gracefully
• Do NOT commit anything yourself; the test-fix script handles that
• Use KillShell if you need to abort a running background shell