Mike Swanson
d4741e447f
Upgrade the human-flow skill (Gemini-assisted, Claude-reviewed): - scan.mjs rewritten to AST-based (@babel/parser/traverse) with 4 detectors: unlabeled-icon-button, tiny-target, missing-feedback-props, click-without-keyboard; regex fallback on parse failure. - Objective Friction Index (Motor 3.0 / Cognitive 2.5 / Keyboard 2.5 / Feedback 2.0); 0-10 Human Workflow Score. - New heuristics: State-Flow Audit, Precision Rail / Fumble Zones, Restraint-o-Meter (1-5) for the fancy pass. - `fix` command DISABLED for now (advisory only): the AST generator reprints whole files and produces noisy diffs; agents apply surgical fixes from the report. To be revisited with a string-splice editor. - Add @babel/* deps + package-lock.json. - Memory: agy review/review-files is NOT actually read-only (wrote files + ran npm despite documented plan-mode) — diff after every agy review. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
1.4 KiB
name, description, metadata
| name | description | metadata | ||
|---|---|---|---|---|
| feedback-agy-review-not-readonly | agy review/review-files can actually WRITE files + run npm, despite docs claiming read-only plan mode — review Gemini's diffs, don't trust its summary. |
|
The agy SKILL.md documents review / review-files as read-only (--approval-mode plan: "Gemini can read files but cannot modify anything"). Observed 2026-06-05 on GURU-5070: a review-files call asking Gemini to "improve" the human-flow skill resulted in Gemini actually editing 6 repo files, adding babel deps to package.json, and running npm install (created package-lock.json + node_modules). So plan-mode was NOT enforced for that run.
Why: The documented safety contract (read-only review) cannot be relied on. Gemini also over-claims — its final summary said it "delivered/upgraded" the skill as if complete, but the only way to know what truly happened was to git diff and run the code.
How to apply: After ANY agy review* call, git status / git diff the target tree to see what actually changed — never trust the summary. If you need a guaranteed read-only second opinion, copy targets to a scratch dir first, or verify the wrapper's approval-mode. The improvements may be good, but they are a PROPOSAL to review and validate (run it, check repo rules like NO EMOJIS), not trusted output. Related: reference_gitea_internal is unrelated; see agy SKILL.md path gotcha.