Skip to content

fix(evals): run unscripted evaluations as the user runs the host - #7

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

phense merged 1 commit into
mainfrom
fix/natural-evals-real-conditions

Conversation

@phense

@phense phense commented Nov 13, 2026

Copy link
Copy Markdown
Owner

Summary

Follow-up to #6, found when the runner was first used for a three-run measurement.

  • The default model is now the adapter's coordinator model. The routing role (a smaller model) loaded skills far less often than the model a user actually runs, so the figures described the wrong condition.
  • On Codex the plugins are used as installed in --auth-home. Repository links under .agents/skills made Codex look for .agents/skills/<plugin>/<skill>/SKILL.md, which does not exist. An installed copy whose skills differ from the source stops the run before any request.
  • A complete shell read of a SKILL.md counts as a load on Claude too.

Verification

  • 183 tests pass; scripts/validate-plugin exits 0.
  • Live: one case per host with Writing Method and Engineering Method; both loaded the expected skill (fable, gpt-6-astra).

🤖 Generated with Claude Code

- default model is the adapter's coordinator model, not the routing role
- Codex uses the plugins installed in the given Codex home and stops when an
  installed copy's skills differ from the source; repository skill links made
  Codex look for paths that do not exist
- a complete shell read of a SKILL.md counts on Claude as well

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
@phense
phense merged commit 91e35da into main Nov 13, 2026
2 checks passed
@phense
phense deleted the fix/natural-evals-real-conditions branch November 13, 2026 19:22
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