mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-09-10 19:01:58 +00:00
f7d39630fe
* fix(autofix): raise the repair and develop-issue agent budgets The repair pass after a deterministic rejection timed out on 48 of 87 attempts this week even at 45 minutes, while the 28 that landed took p50 31m / p90 41m — the successes were brushing the cap. Raise it to 60m and move its step backstop to 70m to keep the 10-minute margin; the review-address job cap follows to 345m (sum of bounded steps + 25m reserve) and the scan's pending-check staleness bound to 375m. The runner default, which only the issue lane's develop leg relies on, goes from 50m to 90m: 6 of 13 develop attempts died at the budget with the branch discarded, against 2 that published a PR. Candidate assessment is pinned to its former 50m explicitly so the issue job's worst case stays under its 180m cap and the always() claim-withdrawal reporter is never cancelled by a job timeout. * fix(autofix): witness assess budget pin, correct calibration prose (#10217) The assess-candidates step's 50-minute QWEN_TIMEOUT_MS pin — the only guard keeping it off the runner's new 90-minute default — had no test witness: deleting the line left the whole suite green. Pin it, and pin the af-101 design-record horizon to the gate's own PENDING_STALE_MIN bound (375m) instead of the review-address job cap, so the doc and the bound cannot drift again. Also correct the idle-watchdog calibration comment: the fleet's shortest leg is now the issue lane's 50-minute assess pass, not the 60-minute repair pass. * test(autofix): witness develop-issue leg inherits runner default budget (#10217) --------- Co-authored-by: qwen-code-dev-bot <qwen-code-dev@service.alibaba.com> |
||
|---|---|---|
| .. | ||
| actions | ||
| assets | ||
| ISSUE_TEMPLATE | ||
| scripts | ||
| workflows | ||
| actionlint.yaml | ||
| CODEOWNERS | ||
| dependabot.yml | ||
| issue-owners.json | ||
| pull_request_template.md | ||
| release.yml | ||
| spam-blocklist.txt | ||