mirror of
https://github.com/QwenLM/qwen-code.git
synced 2026-08-10 17:27:10 +00:00
253f8b8daf
* fix(ci): harden self-hosted runner workspace ownership recovery Containerised jobs (qwen-triage verify/tmux) leave root- or node-owned files in the runner workspace. When the next job's actions/checkout tries to remove them it fails with EACCES, permanently poisoning the runner for all subsequent jobs. Three-layer fix: 1. qwen-triage.yml: split the ownership-restore (chmod + chown back to the runner user) out of the conditional 'Clean up runner workspace' step into its own 'if: always()' step, so it runs even when the job is cancelled or skipped — while the container still has root. 2. qwen-code-pr-review.yml / ci.yml: after the existing chown + sudo attempts, probe each known problem dir (.qwen, .git) with touch; if unwriteable, rename it aside (mv only needs write on the parent directory, which the runner user owns). ci.yml gains the full 'Restore workspace ownership' step it previously lacked. 3. ci.yml 'Clean stale .qwen before checkout' now also removes any .qwen.stale.* directories renamed aside by the step above. Refs: runs 30339720611 (actions-runner-8), 30480422410 (actions-runner-test-14) — both EACCES on .qwen/agents. * fix(ci): address review feedback on workspace ownership recovery Repoint the verify cleanup regression test at the new 'Restore workspace ownership' step so the chmod-before-chown guard is active again, and remove renamed-aside .git.stale.* / .qwen.stale.* dirs in ci.yml and qwen-code-pr-review.yml so they no longer accumulate on self-hosted runners. * test(ci): guard unconditional ownership restore for verify and tmux (#8115) * test(ci): guard rename-aside ownership recovery in ci.yml and pr-review (#8115) * fix(ci): address review feedback on workspace ownership recovery - Make .stale.* cleanup failure visible with :⚠️: instead of silent || true; add sudo -n rm -rf fallback to ci.yml cleanup step - Add .stale.* sweep to containerised verify/tmux-testing ownership- restore steps (run as root, the only actor that can delete them) - Extend ownership recovery to web_shell_e2e_smoke and integration_cli jobs which share the same ecs-qwen self-hosted runner pool - Probe-first optimization: skip expensive recursive chown/chmod on healthy runs; only pay for the full-tree walk when a probe fails - Use mkdir/rmdir instead of touch/rm for writability probe (mkdir never follows symlinks, avoiding a planted-symlink vector) - Use GITHUB_RUN_ID.GITHUB_RUN_ATTEMPT instead of $$ for unique suffix (PID recycles on long-lived runners) - Add set -uo pipefail and $GITHUB_WORKSPACE/ absolute paths to ci.yml cleanup step - Keep rename-aside blocks byte-identical across ci.yml and qwen-code-pr-review.yml with a NOTE comment explaining why extraction into .github/scripts/ is impossible - Add :⚠️: on chmod failure in verify ownership-restore step - Add tests: new job coverage, byte-identical block assertion, .stale.* sweep assertions, cleanup hardening assertions (#8115) * test(ci): guard all four rename-aside copies in byte-identical assertion (#8115) * fix(ci): drop inert rename-aside, restore unconditional recovery (#8115) Review verification showed the pre-checkout rename-aside fallback never unblocks actions/checkout: checkout deletes every workspace entry (or runs git clean -ffdx), walking straight into the renamed dir, and the rename only fires when the runner does not own the dir — exactly when rm -rf cannot empty it either. Drop it from all four checkout jobs and the qwen-triage root sweeps, restoring the simpler unconditional chown/chmod recovery. The probe-first gating is removed for the same reason: poisoning is workspace-wide (root-owned node_modules/dist with no .qwen/.git), so a probe that only checks .qwen/.git reports "healthy" and skips the chown that main did unconditionally — a regression on runners with passwordless sudo. The layer-1 fix (qwen-triage ownership restore running as root under if: always()) is the actual root-cause fix and is kept unchanged. * fix(ci): address review feedback on ownership tests (#8115) Hoist assertUnconditional helper to module scope and reuse it in the pr-review describe block, add sudo-fallback and stat-based UID discovery assertions to guard the recovery branches that are load-bearing on non-root and containerised runners. * fix(ci): cover triage job, add chown warnings, tmux chmod (#8115) * fix(ci): assert restore-before-checkout ordering, surface sudo chown stderr (#8115) * fix(ci): widen verify restore chmod, guard symlinked .qwen, surface sudo chmod stderr (#8115) --------- Co-authored-by: qwen-code-dev-bot <qwen-code-dev-bot@users.noreply.github.com> Co-authored-by: qwen-code-autofix[bot] <qwen-code-autofix[bot]@users.noreply.github.com> Co-authored-by: Qwen Code Autofix <qwen-code-autofix@users.noreply.github.com> Co-authored-by: qwen-code-ci-bot <qwen-code-ci-bot@users.noreply.github.com> Co-authored-by: Qwen Code Bot <qwen-code-bot@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| ci | ||
| dsw-swe-verified | ||
| auto-minimize-spam.test.mjs | ||
| cap-release-notes.mjs | ||
| cap-release-notes.test.mjs | ||
| check-autofix-contracts.sh | ||
| check-settings-schema.sh | ||
| ci-flaky-rerun.mjs | ||
| classify-release-notes.mjs | ||
| classify-release-notes.test.mjs | ||
| pr-safety-precheck.mjs | ||
| pr-safety-precheck.test.mjs | ||
| pr-triage.sh | ||
| qwen-triage-workflow.test.mjs | ||
| resolve-owning-packages.sh | ||
| resolve-sandbox-image.mjs | ||
| resolve-sandbox-image.test.mjs | ||
| run-autofix-review-verification.sh | ||
| serve-ab-diff.mjs | ||
| serve-ab-diff.test.mjs | ||
| serve-ab-drive.mjs | ||
| web-shell-visuals-compose.mjs | ||
| web-shell-visuals-compose.test.mjs | ||
| web-shell-visuals-publish.mjs | ||
| web-shell-visuals-publish.test.mjs | ||