Add gx and gx-pro to the offline benchmark - #1
Open
joelachance wants to merge 1 commit into
Open
Conversation
joelachance
force-pushed
the
gx-submission
branch
from
August 22, 2026 17:36
a807d90 to
8bcdd5c
Compare
joelachance
changed the title
joelachance
force-pushed
the
gx-submission
branch
from
August 22, 2026 18:14
8bcdd5c to
79cdd2d
Compare
joelachance
changed the title
joelachance
force-pushed
the
gx-submission
branch
from
August 22, 2026 18:19
79cdd2d to
38fcc6a
Compare
joelachance
changed the title
joelachance
force-pushed
the
gx-submission
branch
2 times, most recently
from
August 23, 2026 02:50
e1acf1f to
aa89043
Compare
Author
Icon for
|
joelachance
force-pushed
the
gx-submission
branch
2 times, most recently
from
August 24, 2026 02:57
b05c6f5 to
a4eab3d
Compare
joelachance
added a commit
that referenced
this pull request
Aug 24, 2026
gx is an AI code review CLI. Two entries, matching how it ships:
gx two independent Claude Opus 4.6 passes, findings merged
gx-pro a single, more selective pass
Both verify every finding through gx's own Claude Haiku 4.5 verification
stage before reporting it.
Scored through the documented pipeline (extract, dedup, judge) on all three
judge models. Core profile, the dashboard default:
gx gx-pro
prec recall F1 prec recall F1
sonnet-4.5 38.6% 57.0% 46.0 47.7% 46.8% 47.3
opus-4.5 41.2% 57.6% 48.0 49.7% 44.9% 47.2
gpt-5.2 34.3% 55.1% 42.2 43.6% 45.6% 44.6
The two trade coverage against precision. On the 66 Critical and High
severity goldens, gx finds 43 and gx-pro 38.
Both entries follow the documented "adding a new tool" flow. The benchmark
PRs are forked one repo per PR per tool under the naming convention
step1_download_prs.py expects, with the reviews posted on PR #1 of each:
https://github.com/satoricorp-benchmarks
100 repos, 100 PRs, 383 review comments. To regenerate benchmark_data.json
from scratch:
uv run python -m code_review_benchmark.step1_download_prs \
--org satoricorp-benchmarks --tool gx
gx posts under a user account rather than a GitHub App, so both entries are
added to _NON_BOT_TOOLS alongside claude.
No repository-specific configuration was used: none of the five benchmark
repos carries a REVIEW.md, so every review reflects gx's defaults.
Dashboard artifacts are regenerated with `python -m analysis.benchmark_dashboard`.
No existing tool's metrics change; the two new entries are the only additions.
joelachance
force-pushed
the
gx-submission
branch
from
August 24, 2026 04:17
a4eab3d to
532edc6
Compare
gx is an AI code review CLI. Two entries, matching the two configurations
it ships with:
gx full-coverage configuration
gx-pro higher-precision configuration
Every finding is verified before it is reported.
Scored through the documented pipeline (extract, dedup, judge) on all three
judge models. Core profile, the dashboard default:
gx gx-pro
prec recall F1 prec recall F1
sonnet-4.5 38.6% 57.0% 46.0 47.7% 46.8% 47.3
opus-4.5 41.2% 57.6% 48.0 49.7% 44.9% 47.2
gpt-5.2 34.3% 55.1% 42.2 43.6% 45.6% 44.6
The two trade coverage against precision. On the 66 Critical and High
severity goldens, gx finds 43 and gx-pro 38.
Both entries follow the documented "adding a new tool" flow. The benchmark
PRs are forked one repo per PR per tool under the naming convention
step1_download_prs.py expects, with the reviews posted on PR #1 of each:
https://github.com/satoricorp-benchmarks
100 repos, 100 PRs, 383 review comments. To regenerate benchmark_data.json
from scratch:
uv run python -m code_review_benchmark.step1_download_prs \
--org satoricorp-benchmarks --tool gx
gx posts under a user account rather than a GitHub App, so both entries are
added to _NON_BOT_TOOLS alongside claude.
No repository-specific configuration was used: none of the five benchmark
repos carries a REVIEW.md, so every review reflects gx's defaults.
Dashboard artifacts are regenerated with `python -m analysis.benchmark_dashboard`.
No existing tool's metrics change; the two new entries are the only additions.
joelachance
force-pushed
the
gx-submission
branch
from
August 24, 2026 18:14
532edc6 to
33e8c29
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Adds two entries for gx, an AI code review CLI. This PR targets our own fork for review — it is not the submission to withmartian.
Submission requirements
step1_download_prs.pyparses._NON_BOT_TOOLS, since gx posts under a user account rather than a GitHub App.Entries
Two entries, following the existing convention for multiple configurations (
qodo-v2/qodo-extended-v2,greptile-v4-1/greptile-v5):gxgx-proEvery finding is verified before it is reported.
Results — all three judges
Core profile, the dashboard default:
gxgx-progxgx-progxgx-proOn the 66 Critical and High severity goldens,
gxfinds 43 andgx-pro38.Reproducible from the forks
benchmark_data.jsoncan be regenerated rather than taken on trust:Repos are named as the parser expects, with each entry's reviews on PR #1:
Verified: all 100 repo names round-trip through
parse_repo_name→find_golden_urlto the correct golden entry, and live comment counts match the submitted data exactly (383/383, zero mismatches).No repository-specific configuration was used: none of the five benchmark repos carries a
REVIEW.md, so every review reflects gx's defaults.One bug found along the way
find_golden_urlmatches by substring:/discourse-graphite/pull/1is a substring of/discourse-graphite/pull/10, so a repo named…__PR1__…can resolve to PR 10's golden set. It is correct today only becausepull/1precedespull/10in the golden file; reordering that file silently mis-scores. An anchored match would fix it. Not touched here — happy to send it separately.What's in the diff
Same file set as withmartian#20, the maintainer's own tool-addition PR, plus the
_NON_BOT_TOOLSentry:offline/README.md— one vendor rowoffline/code_review_benchmark/step1_download_prs.py—_NON_BOT_TOOLSoffline/analysis/benchmark_dashboard.py—TOOL_COLORSandTOOL_DISPLAY_NAMESoffline/analysis/benchmark_dashboard.{html,json}— regeneratedoffline/results/benchmark_data.jsonand<judge>/{candidates,dedup_groups,evaluations}.jsonfor all three judgesBoth entries went through step 2.5 dedup.
Integrity
gxandgx-profrombenchmark_data.json,candidates.json, andevaluations.jsonreproducesorigin/mainexactly, for all three judges — verified programmatically.benchmark_dashboard.jsonis key ordering, which the generator does not preserve.benchmark_data.jsonis an additions-only diff.ruff checkclean,pytest tests/28 passed.