Qwen Code 0.26.5 caps /autofix-pr at three fix pushes, then hands the PR to a person
Alibaba's open-source coding agent released v0.26.5 on November 1. /autofix-pr now stops after three fix pushes that leave a check failing, and posts a PR comment saying what it tried and what a human needs to do.
Qwen Code v0.26.5 landed on GitHub early on November 1. The change we'd point out is a retry budget for /autofix-pr: if a CI check is still failing after three fix pushes, the agent stops and hands the PR to a human, and the handoff shows up in /autofix-pr status. When a round is blocked, it now posts a comment on the PR saying what it attempted, what's blocking it and what a person needs to do.
A few related changes: PR watch merges a stale base branch before trying fixes when CI failures might come from the branch being behind, the PR-watch skill gets an --in-place mode that works in your own checkout and replies on the PR, and the bundled code-review skill refuses retries and re-anchors when the PR head moves.
Outside the PR tools, Generated UI is on by default, the Context Usage panel lets you click memory files like AGENTS.md to preview them, and there's an experimental one-command Helm deploy for the Managed Agent.
Why it matters for builders
An agent left alone on a failing PR can push fix after fix and burn CI minutes and tokens all night. A hard stop with a written handoff is the behavior you want from any unattended fix loop, whichever tool you use.
Try this
If you use Qwen Code, update to v0.26.5 and run /autofix-pr status on an open PR to see the new retry budget. If you use another agent, add one line to your AGENTS.md: after three failed fix attempts on the same check, stop and leave a PR comment saying what you tried and what's blocking.
Sources
Get stories like this every Monday
One tested AI workflow with the prompts, plus 3 news items that matter if you build. Free.
More AI news
- Pi 1.5.0 adds custom self-update commands, extension changelogs and GPT-6.1 Terra
- Anthropic renames Max 20x to Ultra and drops the five-hour limit for verified accounts
- SemiAnalysis launches GDPx, an agent benchmark where the best model scores 12%
- DeepSeek open-sources dsh-superagent, a research prototype of a coding agent that takes work from GitHub mentions