Repository navigation
Add Smallest complete change rule to /go - #267
Merged
Merged
Conversation
/go's single pass follows the rule, and the stage loop passes it to each stage implementer alongside the judgment-call rule. review-code --fix agents still get neither. ai/evals/smallest-change is an eval-only plugin for `claude plugin eval`. The with arm loads a SessionStart hook that prints rule.md, a copy of the rule that ai/tests/test-eval-rule-sync.sh keeps in sync. The without arm is the tool's no-plugin baseline, so the rule is the only difference. Three scaffolded Python cases grade with hidden tests, a production-line limit, a dependency check, and (in two cases) a check that the change uses the stdlib module or existing helper. Claude-Session: https://claude.ai/code/session_01AovXvYri48wFhqGV3epaMi
Judgment calls and Smallest complete change go to the single pass and each stage implementer only. The test writer and review-code --fix agents get neither. Claude-Session: https://claude.ai/code/session_01AovXvYri48wFhqGV3epaMi
claude plugin eval ignores the workspace's .claude/settings.json, and its no-plugin ablation arm drops every plugin, so neither can grade the baseline. run.sh runs two single-arm evals instead: this plugin, whose hooks print the rule and grade the workspace, and a baseline plugin it builds from a copy of the cases and the grading hook alone. compare.py prints each grader's pass count in both arms. Claude-Session: https://claude.ai/code/session_01AovXvYri48wFhqGV3epaMi
Each case now has a grader per size limit (5, 10, 20, and 40 production lines), so the pass counts show roughly how many lines each arm changed. retry-reuse replaces permalink-reuse, which both arms passed every time: it asks for webhook retries in a codebase that already has a retry helper. Claude-Session: https://claude.ai/code/session_01AovXvYri48wFhqGV3epaMi
haacked
marked this pull request as ready for review
October 29, 2026 01:40
haacked
changed the title
haacked
added a commit
that referenced
this pull request
Oct 29, 2026
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.
/go doesn't tell its implementers to prefer the smallest change that fully works. They sometimes hand-roll what the standard library or the codebase already has, or add options nobody asked for. The Ponytail skill targets this, and JetBrains measured it independently at 10% lower cost and 15% less code across 80 paired tasks with no quality difference. I didn't want the plugin itself, though. It injects its rules into every subagent, code reviewers included, and writes a per-task ledger by default.
This adds a Smallest complete change rule to /go's Implement step, adapted from Ponytail's ladder. It leaves out Ponytail's testing, comment, and end-of-reply rules, because
unit-test-writer,comment-cleanup, and the## Flagged callslist already handle those. The single pass follows the rule, and the stage loop passes it to each stage implementer with the judgment-call rule. The test writer andreview-code --fixagents get neither.I also ran a small
claude plugin evalcomparison (3 cases, 3 runs per arm). Every hidden test passed with and without the rule, and code size and cost came out about even. The cases turned out too easy to show Ponytail's effect, so I left the eval out of this PR.Test plan
## Flagged callslist padded with trivial omissions from the rule's "flag each one left out" stephttps://claude.ai/code/session_01AovXvYri48wFhqGV3epaMi