Problem
After AI conflict resolution during cherry-pick, the code proceeds to git add -A + cherry-pick --continue without verifying that conflict markers were actually removed from files. If the AI fails silently (e.g. sidecar returns success=True but the model never ran), unresolved conflict markers get committed.
Root Cause
In _resolve_cherry_pick_with_ai() (webhook_server/libs/handlers/runner_handler.py), the only validation after the AI call is:
if not ai_call_result.success:
...
return False, ""The pi_sidecar_client marks results as success=True based on HTTP 200 status, regardless of whether the AI actually performed any work. The AI resolves conflicts through tool calls (edit/write), not through its response text — so empty text alone is not an indicator of failure. However, when the model is never invoked (0 tokens in/out), no tool calls happen and conflict markers remain in the files.
Failure Flow
1. /cherry-pick <branch> → git cherry-pick → CONFLICT in file.py
2. AI conflict resolution invoked
3. Sidecar returns success=True but model never ran (0 tokens)
4. No tool calls executed → conflict markers still in files
↓
5. git add -A ← stages files with conflict markers
6. cherry-pick --continue ← commits broken file
7. Pre-commit fails: SyntaxError (conflict markers)
8. Cherry-pick aborts via "unfixable errors" path
Pre-commit acts as a safety net, but the cherry-pick falls through to the wrong error path ("unfixable errors") instead of the proper manual cherry-pick fallback with instructions.
Expected Flow
1–4. Same as above
↓
5. Scan working tree for conflict markers (<<<<<<< / ======= / >>>>>>>)
6. Markers found → return (False, "") → manual fallback with instructions
Proposed Fix
After the AI call succeeds but before git add -A, scan the conflicted files for remaining conflict markers (<<<<<<<, =======, >>>>>>>). If any remain, treat the resolution as failed and fall back to the manual cherry-pick path.
Done
Problem
After AI conflict resolution during cherry-pick, the code proceeds to
git add -A+cherry-pick --continuewithout verifying that conflict markers were actually removed from files. If the AI fails silently (e.g. sidecar returnssuccess=Truebut the model never ran), unresolved conflict markers get committed.Root Cause
In
_resolve_cherry_pick_with_ai()(webhook_server/libs/handlers/runner_handler.py), the only validation after the AI call is:The
pi_sidecar_clientmarks results assuccess=Truebased on HTTP 200 status, regardless of whether the AI actually performed any work. The AI resolves conflicts through tool calls (edit/write), not through its response text — so empty text alone is not an indicator of failure. However, when the model is never invoked (0 tokens in/out), no tool calls happen and conflict markers remain in the files.Failure Flow
Pre-commit acts as a safety net, but the cherry-pick falls through to the wrong error path ("unfixable errors") instead of the proper manual cherry-pick fallback with instructions.
Expected Flow
Proposed Fix
After the AI call succeeds but before
git add -A, scan the conflicted files for remaining conflict markers (<<<<<<<,=======,>>>>>>>). If any remain, treat the resolution as failed and fall back to the manual cherry-pick path.Done
_resolve_cherry_pick_with_ai()