mirror of
https://github.com/LearningCircuit/local-deep-research.git
synced 2026-06-13 18:44:51 +03:00
bc45bbf2e7
* fix(ci): make LDR research workflow honestly fail on Python crash A real run (job 77511717371, PR #4225) crashed with glibc 'double free or corruption (!prev)' but the workflow reported success and the caller posted a hollow PR comment. Two cooperating defects: the script's exit code was discarded inside set +e / set -e, and `jq .` exits 0 on a zero-byte response.json so the JSON-shape check passed on empty input. Capture the exit code, harden the validation order (exit -> non-empty -> jq -e shape -> .error -> .research non-empty), tee stderr to a log surfaced in the ::error:: annotation, upload the artifact with if: always() so failed runs leave debuggable evidence, and flush stdout in a finally block in the script so a SIGABRT during interpreter shutdown after json.dumps can't drop the otherwise-completed output. Matches the house pattern from dockle.yml and the jq -e idiom from release-gate.yml. * chore(ci): enable faulthandler in ldr-research.py Dumps a Python traceback to stderr on SIGABRT/SIGSEGV/SIGFPE/SIGBUS/SIGILL before the signal re-fires. Pairs with the stderr-capture plumbing earlier in this PR: on the next glibc abort the ::error:: annotation will show "Fatal Python error: Aborted" plus the actual Python stack frame, making the deps-level investigation possible without a re-run. Verified locally: a deliberately aborted child process emits its frame through faulthandler before exiting on signal 6.