Experiment: can Bash be trimmed from allowed-tools in pre-fetch-only skills?

[chrispezza]

skill-audit flagged that several skills grant Bash while their bodies never shell out — but pre-fetched !... context execution may itself require the grant, so trimming blindly could break every skill's context loading.

The experiment (needs any logged-in claude CLI; the desktop app session cannot run claude -p):

1. Create a scratch project with .claude/skills/probe-withbash/SKILL.md and .claude/skills/probe-nobash/SKILL.md, identical except allowed-tools: Bash, Read vs allowed-tools: Read. Body of each:

---
name: probe-nobash
description: Internal test probe. Use only when explicitly invoked.
allowed-tools: Read
---
Pre-fetched: !`echo PROBE_NOBASH_EXECUTED`

Reply with exactly one line: the pre-fetched value shown above if it is the word-output of the command (starts with PROBE_), or the text RAW-COMMAND-VISIBLE if you see an unexecuted command/backticks instead.

2. From that directory: claude -p "/probe-withbash" then claude -p "/probe-nobash".
3. If BOTH print their PROBE_*_EXECUTED marker, the Bash grant is not required for pre-fetch and can be trimmed from pre-fetch-only skills (both pr-descriptions, component-scaffold). If the nobash probe shows the raw command, the grant is required — close this issue documenting that and keep the grants.

Tracked in docs/ROADMAP.md open decisions.