claude-code - 💡(How to fix) Fix [BUG] /fewer-permission-prompts skill hard-denied by classifier writing settings.json (2.1.143 regression)

Preflight

• Searched existing issues (#52075 related, different facet)
• Single bug report
• Latest version (2.1.143)

What's Wrong?

In Claude Code 2.1.143, the /fewer-permission-prompts skill cannot complete its core job. The skill scans transcripts, produces a curated list of allow-rule additions, and tries to merge them into ~/.claude/settings.json. Every write variant — Python script via Bash, Write tool, Edit tool — is hard-denied by the auto-mode classifier with:

Modifying ~/.claude/settings.json to add permission allow-rules is Self-Modification of agent configuration.

The skill's single advertised purpose is to add permission allow-rules. The product ships a first-party skill whose final step is categorically forbidden by the same product.

This sits inside a broader 2.1.143 regression: the classifier hard-denies whole categories of action regardless of the user's hand-curated `permissions.allow` list (~177 entries in my case). Explicit user consent recorded in settings.json is being overridden by an opaque classifier rule set.

Expected

The user's explicit allowlist is the consent contract. `/fewer-permission-prompts` should be able to write to the file it exists to write.

Repro

1. On Claude Code 2.1.143
2. Run `/fewer-permission-prompts`
3. Let the skill produce its merge plan
4. Observe the merge step hard-denied with the "Self-Modification of agent configuration" reason — regardless of tool used (Bash python, Write, Edit)

Related

• #52075 — prompts despite bypass (same classifier, different facet)
• Auto-updater silently re-pinned 2.1.143 ~8 hours after a manual rollback to 2.1.142. Version-pinning does not stick without `DISABLE_AUTOUPDATER=1` or removing the 2.1.143 version directory. Users who rolled back are being silently re-regressed.

Workaround

`--dangerously-skip-permissions` at launch bypasses the classifier entirely. Not a reasonable default — it also bypasses the user's `deny` list.

Regression?

Yes — worked in 2.1.142.

Environment

• macOS 26.3.1, MacBook Pro M1
• Claude Code 2.1.143 (auto-updated from 2.1.142)
• Opus 4.7