pr-driver-protocol

PR driver protocol — fix-all before push (bd-snx3)

A standalone, callable protocol for any agent (worker, babysit, human) that is driving a PR to green. The single rule: after every push, fix ALL known gate failures in one local pass before the next push. "CI pending" is not an idle state.

When to use this protocol

• You just pushed a commit and need to wait for CI to come back
• You see the same gate failure 2+ times in a row (worker or babysit perspective)
• You are babysitting a worker that appears to be partial-fixing
• You are tempted to do "one quick fix and re-push" — stop, run this protocol first

The 5-step loop

1. ENUMERATE   — read ALL current gate failures, not just the first
2. CLASSIFY    — group by file + root cause (same root cause = one fix)
3. FIX-ALL     — apply every fix locally; do not push until all are staged
4. VERIFY      — run the local equivalent of each failing gate
5. PUSH        — single commit covering all fixes; then back to step 1

Step 1: ENUMERATE

Do not stop at the first failure. Use /green <PR> or read the failure-class log and build a complete list. For each failure, capture:

• Gate name (Green Gate, CodeRabbit, Skeptic, Evidence, Bugbot, inline threads)
• File and line (e.g., packages/cli/src/doctor.ts:142)
• Failure text (the actual error message, not a paraphrase)
• Root cause hypothesis (what is wrong, not just what is failing)

Step 2: CLASSIFY

Group failures by the underlying root cause. Two failures with the same root cause are one fix:

F1: dist/index.js argv shape not recognized in non-canonical check   (cli/doctor.ts:142)
F2: worker binary path resolution fails when global npm is used        (cli/doctor.ts:203)
  → Same root cause: argv path normalization is regex-naive
  → ONE fix: switch to path.basename() comparison

Do not push two commits to fix one root cause. The whole point of this protocol is to stop partial-fix-then-push loops.

Step 3: FIX-ALL