Bug type

Regression (worked before, now fails)

Summary

When openclaw update runs preflight and the candidate commit fails (e.g. lint), the updater walks backward through earlier commits to find a "good" one. Every one of those fallback checkouts fails with:

error: Your local changes to the following files would be overwritten by checkout:
    src/canvas-host/a2ui/.bundle.hash
Please commit your changes or stash them before you switch branches.
Aborting

The updater then bails with Reason: preflight-no-good-commit, even though there almost certainly is a good commit further back — it just can't reach it.

Related symptom-level issue: #58544 (different root cause — deps install — but same final error).

Root cause

Two things combine:

1. src/canvas-host/a2ui/.bundle.hash is tracked in git again, despite being listed in .gitignore:67 and explicitly untracked by c568142af9 (chore: stop tracking a2ui bundle hash, Jan 27). It was re-introduced accidentally in 74e7b8d47b fix(cycles): bulk extract leaf type surfaces and has been updated several times since (most recently 370efaa4a0 build(canvas): refresh a2ui bundle).

2. scripts/bundle-a2ui.sh writes to that file as a side effect of the build (line 104: echo "$current_hash" > "$HASH_FILE"). It only skips writing if the previous hash matches AND the output bundle exists — otherwise it always rewrites.

So the preflight worktree sequence is:

1. preflight checkout <candidate> — file matches tracked blob.
2. preflight build <candidate> — bundle-a2ui.sh rewrites the hash with a different value (e.g. when inputs differ, or when the worktree's build cache doesn't line up with the committed hash).
3. preflight lint <candidate> — fails (for whatever unrelated reason).
4. preflight checkout <older-commit> — refuses, because the working copy's .bundle.hash differs from the index and the older commit tracks a different blob at that path.
5. Every subsequent walkback checkout hits the same wall → preflight-no-good-commit.

Reproduction

Easy to reproduce whenever main has any transient preflight failure. To force it cleanly on a currently-good tree:

cd /path/to/openclaw
echo "deadbeef" > src/canvas-host/a2ui/.bundle.hash
git checkout HEAD~1 -- .   # or any older commit that tracks the file
# -> "Your local changes... would be overwritten"

Real-world repro I hit today: a lint regression around 370efaa4 triggered the bisect walkback, and every fallback checkout (aa092045, 3e013342, 81535d39, 355794c2, 7c5b42e4, 7315914e, 43bd5545, c159d22b, ed356d74) failed with the .bundle.hash conflict. Update bailed with preflight-no-good-commit, leaving the install stuck on an older version until upstream landed a fix.

Fix options

Either (or both) of these would eliminate the class of failure:

1. Re-untrack the file (mirrors c568142af9):

   git rm --cached src/canvas-host/a2ui/.bundle.hash
   git commit -m "chore: stop tracking a2ui bundle hash (again)"

   This is the cleanest fix and matches the clearly-intended gitignore state. Worth also hunting for whatever tooling keeps re-adding it — 74e7b8d47b looks like an unrelated commit that accidentally staged the file.

2. Harden scripts/bundle-a2ui.sh to avoid rewriting the hash file when the content would be unchanged:

   if [[ ! -f "$HASH_FILE" || "$(cat "$HASH_FILE")" != "$current_hash" ]]; then
     echo "$current_hash" > "$HASH_FILE"
   fi

   Defense in depth — handles any future re-track of the file.

3. Make the preflight updater resilient to checkout conflicts — e.g. git reset --hard or git stash inside the preflight worktree before each fallback checkout. The preflight worktree is disposable, so force-resetting it is safe.

Option 1 is the minimal upstream patch; option 3 is the most robust since it also protects against similar issues with other generated-but-accidentally-tracked files.

Environment

• OpenClaw 2026.4.6 → 2026.4.11 (openclaw update path)
• macOS (Darwin 25.5.0)
• Git install