eval-audit
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per SKILL.md's Global UX Rules. Dimension labels: see the canonical table in SKILL.md.
$ npx -y skills add Evol-ai/SkillCompass --agent claude-codeHow it fires
How this command gets triggered: by you, by Claude, or both.
/eval-rollbackContext preview
What this command does when you run it.
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per SKILL.md's Global UX Rules. Dimension labels: see the canonical table in SKILL.md.
> **Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per SKILL.md's Global UX Rules. Dimension labels: see the canonical table in SKILL.md.
Use the **Read** tool to load `.skill-compass/{skill-name}/manifest.json`.
If not found: output a locale-appropriate message, e.g.:
Then present a choice (skip if `--internal` or `--ci`):
[Evaluate this skill now / Cancel]
If the user chooses **Evaluate this skill now**, invoke `/eval-skill <skill-name>` and stop. If **Cancel**, stop.
Show all tracked versions:
Version History: sql-optimizer | # | Version | Score | Verdict | Trigger | Dimension | Date | |---|--------------|-------|---------|--------------|-----------|------------| | 1 | 1.0.0 | 38 | FAIL | initial | — | 2026-01-15 | | 2 | 1.0.0-evo.1 | 52 | CAUTION | eval-improve | security | 2026-01-16 | | 3 | 1.0.0-evo.2 | 62 | CAUTION | eval-improve | trigger | 2026-01-17 | | 4 | 1.0.0-evo.3 | 71 | PASS | eval-improve | functional| 2026-01-18 | ← current
If `--to` was specified: proceed to Step 3 with that version. If not: prompt the user to enter the row number (`#`) from the table above — do **not** ask them to type the version string. Example prompt:
Enter the row number to rollback to (e.g. 2):
Map the entered number to the corresponding version via the table. Re-prompt on invalid input.
Before rollback, snapshot the current version: 1. Compute content hash of current SKILL.md 2. If not already in snapshots: save using the **Write** tool
Use the **Read** tool to load the target version from `.skill-compass/{skill-name}/snapshots/{version}.md`.
If snapshot not found: output `"Snapshot missing for version {version}. Cannot rollback."` and stop.
Use the **Write** tool to overwrite the SKILL.md file with the snapshot content.
Update `current_version` in manifest to the restored version. Do NOT delete later version records (they remain in history for future reference).
Use the **Write** tool to save the updated manifest.
Log the rollback event to the audit chain so that the Skill Inbox `undo-2x` rule can detect repeated rollbacks:
Using Node.js (or instruct Claude to execute):
const { AuditChain } = require('./lib/audit-chain');
const crypto = require('node:crypto');
const auditChain = new AuditChain(skillName, 'cc');
auditChain.log({
type: 'rollback',
severity: 'WARN',
message: `Rolled back from ${currentVersion} to ${targetVersion}`,
skillHash: crypto.createHash('sha256').update(restoredContent).digest('hex'),
findings: []
});The second argument `'cc'` routes the audit log to `.skill-compass/cc/{skill-name}/audit.jsonl` (platform-specific path). If `lib/audit-chain.js` is not accessible from the command context, manually write a JSON line to `.skill-compass/cc/{skill-name}/audit.jsonl` with `type: "rollback"` and the current timestamp.
Output a locale-appropriate confirmation, e.g.:
Then, unless `--internal` or `--ci` is set, present a flow-continuity choice:
[Re-evaluate to confirm quality (recommended) / Compare the two versions / Done]
All messages below follow the session locale.
Evaluate agent skill quality. Find the weakest link. Fix it. Prove it worked.
Repo: Evol-ai/SkillCompass
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…
- **Recommended model: Claude Opus 4.6** (`claude-opus-4-6`). Directed improvement requires understanding complex rubric feedback and generating precise,…
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…
**Locale**: All templates in this spec are written in English. Detect the user's language from the session and translate user-facing text at display time per…