[experimental-ci-only-validator] Introduce ci_only validator skill
This CL introduces the experimental-ci-only-validator skill
which uses the underlying Python helper script to help triaging
missing coverage bugs on Gerrit.
Bug: 519590223
Test: None
TAG=agy
CONV=19f2d950-e07e-427c-a7a3-de836354df62
[experimental-ci-only-validator] Add ci_only validator tool
This CL introduces the ci_only validator tool and its tests.
The tool checks if a Chromium test suite is barred by `ci_only = True`
on both the coverage-specific trybots and standard CQ coverage trybots.
Bug: 519590223
Test: vpython3 tools/code_coverage/ci_only_validator_test.py
TAG=agy
CONV=19f2d950-e07e-427c-a7a3-de836354df62
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
#### 2. Checking Historical Configuration (at a revision)I chose to not include the time travel code explorer skill since it requires multiple tool calls to explore all the try builder configs at a specific revision.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
- type: contains
value: "ci_only = True"This seems liable to fail because the agent won't report this exact string.
Expecting certain builder names to be reported makes sense, but I don't think this does.
#### 2. Checking Historical Configuration (at a revision)I chose to not include the time travel code explorer skill since it requires multiple tool calls to explore all the try builder configs at a specific revision.
Acknowledged
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
- type: contains
value: "ci_only = True"This seems liable to fail because the agent won't report this exact string.
Expecting certain builder names to be reported makes sense, but I don't think this does.
Would it be better to define what the model should output in the query? Ex. Output "ci_only: True" if the configs contain it. I am pending access to google cloud resources, so I am unable to run evals locally still.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
- type: contains
value: "ci_only = True"Aadarsh BattulaThis seems liable to fail because the agent won't report this exact string.
Expecting certain builder names to be reported makes sense, but I don't think this does.
Would it be better to define what the model should output in the query? Ex. Output "ci_only: True" if the configs contain it. I am pending access to google cloud resources, so I am unable to run evals locally still.
I think the better option would be to update prompt to ask for what we really want. We don't really care about the output format that the agent uses, we only care about the information included in its response.
In this particular case, a prompt like
`Use the experimental-ci-only-validator skill to find all CI-only code coverage LUCI builders that run the 'chrome_junit_tests' suite` seems more fitting and then the asserts can be that the substrings of the builders that we expect.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
- type: contains
value: "ci_only = True"Aadarsh BattulaThis seems liable to fail because the agent won't report this exact string.
Expecting certain builder names to be reported makes sense, but I don't think this does.
Brian SheedyWould it be better to define what the model should output in the query? Ex. Output "ci_only: True" if the configs contain it. I am pending access to google cloud resources, so I am unable to run evals locally still.
I think the better option would be to update prompt to ask for what we really want. We don't really care about the output format that the agent uses, we only care about the information included in its response.
In this particular case, a prompt like
`Use the experimental-ci-only-validator skill to find all CI-only code coverage LUCI builders that run the 'chrome_junit_tests' suite` seems more fitting and then the asserts can be that the substrings of the builders that we expect.
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Code-Review | +1 |
LGTM
+CC bryankim since this adds another eval
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |
19 is the latest approved patch-set.
No files were changed between the latest approved patch-set and the submitted one.
Introduce ci_only validator skill
This CL introduces the experimental-ci-only-validator skill
which uses the underlying Python helper script to help triaging
missing coverage bugs on Gerrit.
TAG=agy
CONV=19f2d950-e07e-427c-a7a3-de836354df62
| Inspect html for hidden footers to help with email filtering. To unsubscribe visit settings. |