Shell-Gated Bump with Benchmark Regression Guard

This agent gates dependency upgrades on two signals at once: correctness and speed. It runs the test suite and a benchmark in a sandboxed shell, then opens a GitLab MR only when tests pass and performance stays within an acceptable delta of the recorded baseline.

Use it for performance-sensitive services where a silently slower dependency is as dangerous as a broken one. The benchmark guard catches regressions that green tests miss.

1. 1A schedule launches the run.
2. 2The agent pins one package upgrade and, in a sandboxed shell, runs both the test suite and the benchmark script, capturing timing numbers.
3. 3A logic gate compares results: tests must pass and the benchmark delta must stay under the configured threshold.
4. 4If either check fails, the run aborts with a logged reason and no MR.
5. 5On a clean pass, the agent opens a GitLab MR embedding the before/after benchmark table.
6. 6The MR link is returned as the final output.