Repository navigation
Add an eval that measures the impact of loadout - #6
Merged
Merged
Conversation
The eval runs real headless sessions with and without the plugin. It compares cost, time and quality for five arms: baseline, loadout, effort, model and all. - evals/impact.mjs runs the sessions and writes results.jsonl and report.md. The report compares each arm with the baseline per task. - evals/fixture.mjs has a fixed setup: 12 fake MCP servers with 80 tools and 40 project skills. Use --env real for your own setup. - evals/tasks.mjs has 8 tasks with deterministic checks. - The plugin writes the token use of each classifier call to its trace file, because the session cost does not include these calls. - tests/session.mjs gets options for the eval. It can answer the permission asks and the questions of the plugin through the control channel of the SDK. Code review: two findings are not fixed. - installedLoadout reads only the user settings. A copy from managed settings cannot be turned off by the eval. - Some answer checks are loose. The prompts ask for the number only. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-eval-dfe58b # Conflicts: # .claude-plugin/plugin.json
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
This PR adds an eval that measures the cost, speed and quality of Claude Code sessions with and without loadout. It also measures the effort and model switches.
evals/impact.mjsruns real headless sessions in five arms:baseline,loadout,effort,modelandall. It writesresults.jsonlandreport.md. The report compares each arm with the baseline per task (geometric mean over the tasks).evals/fixture.mjsgives a fixed setup: 12 fake MCP servers with 80 tools and 40 project skills.--env realuses your own setup.evals/tasks.mjshas 8 tasks with deterministic checks (server calls, files, answers).tests/session.mjsgets new options for the eval. It answers permission asks and plugin questions through the SDK control channel.docs/development.mdx("Impact eval").First result
Sonnet, effort
high, fixed setup, 2 runs, 80 sessions, $2.58, 4.5 minutes. All checks pass.The first prompt goes from 40452 to approximately 33k tokens.
Verification
/simplifyand/code-review: the review found 4 problems. 2 are fixed. The commit message gives the reasons for the 2 that are not fixed.claude plugin validate .,claude plugin test .(49 pass),node tests/docs.check.mjs: all pass./verify: the smoke test passes. A driven session shows oneclassifiertrace line for each Haiku call, with the model and the token use.🤖 Generated with Claude Code