Skip to content

fix(evals): count a skill read inside a compound command - #8

Merged
phense merged 1 commit into
mainfrom
fix/natural-evals-compound-reads
Nov 13, 2026
Merged

phense merged 1 commit into
mainfrom
fix/natural-evals-compound-reads

Conversation

@phense

@phense phense commented Nov 13, 2026

Copy link
Copy Markdown
Owner

Summary

The unscripted runner counted a Codex skill read only when the command itself was a plain cat, sed or head. Codex usually reads a skill inside a compound command (pwd; rg ...; cat <skill>), so real loads were scored as misses. A skill now counts when the complete SKILL.md is in a successful command's output, whatever the command looked like. A path, a listing or a partial read still does not count.

Verification

  • 183 tests pass (one new assertion for a compound command).
  • Re-scoring the saved transcripts of 36 Codex runs: 8 runs scored as misses contained a complete read of the expected skill; the result moved from 27 to 35 of 36.

🤖 Generated with Claude Code

Codex reads a skill as part of commands such as `pwd; rg ...; cat <skill>`.
The unscripted runner now counts a skill when the complete SKILL.md is in a
successful command's output, whatever the command looked like. Re-scoring the
saved transcripts of 36 Codex runs turned 8 apparent misses into loads.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@phense
phense merged commit 64d9a02 into main Nov 13, 2026
2 checks passed
@phense
phense deleted the fix/natural-evals-compound-reads branch November 13, 2026 19:51
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant