Qwen Code 0.21.12 Introduces Review Witness Gates and Bounded Autofix Budgets
Qwen Code 0.21.12 adds strict evidence gating for AI code review findings and dual line-growth budgets for automated PR repairs. Unwitnessed critical issues are demoted to low confidence, preventing endless automated fix loops.
Impact: Medium
Why it matters
Implement these reference workflow patterns in your GitHub Actions to block hallucinated AI review comments and prevent runaway pull request bloat.
TL;DR
- 01Critical review findings must carry replaying command outputs or execution traces to post on PRs.
- 02Autofix workflows maintain separate 400-line growth budgets for source code and test files.
- 03Non-converging repair cycles automatically halt and escalate decisions to human maintainers.
Key facts
- Default Source Line Budget
- 400 net lines
- Default Test Line Budget
- 400 net lines
- Max Non-Shrinking Rounds
- 2 rounds
Review Witness Gate
Under Qwen Code 0.21.12, high-confidence review findings marked Critical must supply a witness field containing execution output or a valid not run reason. Without execution proof, claims stay off the pull request:
critical_requires: witness_or_not_run_reasonDual Growth Budgets
To prevent pull request bloat during automated fix cycles, the workflow tracks source and test diffs separately against the merge base:
- Source Growth Budget:
QWEN_AUTOFIX_GROWTH_BUDGET_SRC_LINES(default: 400 net lines) - Test Growth Budget:
QWEN_AUTOFIX_GROWTH_BUDGET_TEST_LINES(default: 400 net lines) - Divergence Limit: Pauses after 2 non-shrinking rounds for human intervention
When growth limits are hit, non-critical suggestions stop while blocking fixes remain active.
Try it in 2 minutes
env:
QWEN_AUTOFIX_GROWTH_BUDGET_SRC_LINES: 400
QWEN_AUTOFIX_GROWTH_BUDGET_TEST_LINES: 400yaml
✓ When to use
- When orchestrating automated AI agent repair loops in continuous integration systems.
✕ When NOT to use
- When running simple read-only review scripts without automated write permissions on pull requests.
What to do today
- Adopt witness validation logic in automated PR review pipelines.
- Set separate source and test line growth budgets in your AI autofix workflows.
Sources