[sferik]

A review of the same Duel workflow in another project turned up a couple of things worth fixing here too, so this brings perf.yml in line with what duel init generates now.

The Duel job checks out and runs the pull request's code, benchmarks included, and it also had pull-requests: write so it could post its report. Now it only reads, and its checkout keeps no credentials. It uploads the report as an artifact, and a separate Comment job downloads it and posts the comment. That job checks nothing out and runs none of the branch's code, so it's the only one that can write. A pull request from a fork still gets no comment, since its token can't write one, but the job's summary has the report.

The other fix is in .duel.rb. Duel only compares a change that touches lib/ or ext/, so a pull request that changed only the benchmarks or the config passed with "nothing to compare", and a broken benchmark could slip through. It now watches duel/, .duel.rb and the workflow too. That also means this pull request runs a real comparison of the code with itself, which checks that the new workflow works.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-26T01:32:14.810521Z	0ed9b95	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

Duel passed: none of 23 benchmarks slower

• Baseline: 2822344, where HEAD left origin/main
• Candidate: b976316

23 benchmarks in 28s. A timing is slower when a 99.0% confidence interval for its ratio, the candidate's to the baseline's, is above 1.02.

Across the board

• cpu 0.999 [0.995, 1.003] same (limit 1.0050, 23 benchmarks)

23 benchmarks unchanged

Benchmark	Timings (ratio [interval])
backend: libxml, 20 records	cpu 0.999 [0.994, 1.014]
backend: libxml SAX, 20 records	cpu 0.999 [0.990, 1.018]
backend: nokogiri, 20 records	cpu 0.997 [0.986, 1.008]
backend: nokogiri SAX, 20 records	cpu 1.000 [0.958, 1.011]
backend: oga, 20 records	cpu 0.993 [0.790, 1.010]
backend: rexml, 20 records	cpu 1.001 [0.989, 1.014]
boot: require multi_xml	cpu 1.002 [0.979, 1.008]
parse: a tiny document	cpu 0.994 [0.966, 1.009]
parse: 10 records	cpu 1.009 [0.970, 1.019]
parse: 120 records	cpu 1.002 [0.992, 1.011]
parse: a large catalog	cpu 0.997 [0.973, 1.016]
parse: 450 shallow fields	cpu 1.002 [0.988, 1.013]
parse: a tree 180 levels deep	cpu 0.995 [0.983, 1.013]
parse: attribute-dense nodes	cpu 1.003 [0.992, 1.016]
parse: mixed content	cpu 1.006 [0.980, 1.015]
parse: a feed, namespaces stripped	cpu 0.999 [0.978, 1.016]
parse: a feed, namespaces preserved	cpu 0.993 [0.980, 1.004]
parse: dashed keys	cpu 1.003 [0.990, 1.014]
parse: 120 records, symbolized names	cpu 0.998 [0.983, 1.020]
parse: 120 records, no typecasting	cpu 0.996 [0.982, 1.016]
parse: 10 records from an IO	cpu 1.008 [0.989, 1.018]
typecast: one typed record	cpu 0.993 [0.984, 1.015]
typecast: 50 typed records	cpu 1.000 [0.981, 1.011]