From f0e10d6d2648e9d74587026f6bb2cf4c4ee9d06c Mon Sep 17 00:00:00 2001
From: tiennm99 <tiennm99@outlook.com>
Date: Thu, 14 May 2026 15:46:13 +0700
Subject: [PATCH] fix(ci): narrow conflict resolution to bot-generated files
 only

The blanket git pull --rebase -X theirs in the push-retry loop silently
discarded conflicting changes in any file, so a maintainer push to
data/agents.yml racing with the daily run could be dropped without
warning. Replace it with a selective resolver: plain rebase first, then
inspect git diff --name-only --diff-filter=U; if every conflicting path
is data/history.jsonl or README.md (both bot-generated), resolve those
with --theirs and continue; otherwise abort the rebase and fail the run
loudly so the human edit is preserved and visible.

Also leave a comment at the permissions block warning future maintainers
not to swap GITHUB_TOKEN for a PAT, since the no-recursion invariant is
what currently prevents the auto-commit from triggering its own workflow.
---
 .github/workflows/update.yml | 28 +++++++++++++++++++++++++---
 1 file changed, 25 insertions(+), 3 deletions(-)

diff --git a/.github/workflows/update.yml b/.github/workflows/update.yml
index 74cd3c2..cb0fa36 100644
--- a/.github/workflows/update.yml
+++ b/.github/workflows/update.yml
@@ -14,6 +14,9 @@ on:
       - '.github/workflows/update.yml'
 
 permissions:
+  # Uses GITHUB_TOKEN intentionally — a PAT would cause infinite trigger loops
+  # on the auto-commit ("chore: daily ranking refresh") because GITHUB_TOKEN-authored
+  # pushes do NOT re-trigger workflows, whereas a PAT would.
   contents: write
 
 concurrency:
@@ -46,11 +49,30 @@ jobs:
             exit 0
           fi
           git commit -m "chore: daily ranking refresh"
-          # rebase + retry guards against the race where two runs commit near-simultaneously.
-          # -X theirs makes our run's data/history.jsonl line win conflicts (our snapshot is the freshest).
+          # Rebase + retry guards against the race where two runs commit near-simultaneously.
+          # Strategy: plain rebase first; only force-resolve bot-generated files (README.md,
+          # data/history.jsonl) when they are the sole conflicts. If any human-maintained file
+          # (e.g. data/agents.yml) conflicts, abort and fail loudly so a human can investigate.
           for attempt in 1 2 3; do
             if git push; then exit 0; fi
             echo "push attempt $attempt rejected, rebasing on origin/main"
-            git pull --rebase -X theirs origin main
+            if git pull --rebase origin main; then
+              # Clean rebase — no conflicts; loop back to try push again.
+              continue
+            fi
+            # Rebase stopped on conflicts. Inspect which files are unresolved.
+            conflicted=$(git diff --name-only --diff-filter=U)
+            echo "conflicted files: $conflicted"
+            # Only auto-resolve if ALL conflicts are in bot-generated files.
+            non_bot=$(echo "$conflicted" | grep -v -E '^(README\.md|data/history\.jsonl)$' || true)
+            if [ -n "$non_bot" ]; then
+              echo "ERROR: conflict in human-maintained file(s): $non_bot — aborting rebase"
+              git rebase --abort
+              exit 1
+            fi
+            # Safe to force-resolve: keep our (freshest) bot-generated content.
+            git checkout --theirs README.md data/history.jsonl 2>/dev/null || true
+            git add README.md data/history.jsonl
+            GIT_EDITOR=true git rebase --continue
           done
           exit 1