Two related friction points came up discussing Claude Code setup for
Serokell work:
settings.json's permissions.allow list could pre-approve obviously
safe, frequently-used commands (e.g. Bash(git status:*),
mcp__youtrack__get_issue) so they stop prompting every time. But there
are a lot of these, they vary per engineer/project, and some run behind a
proxy (e.g. rtk, which itself prompts to approve rtk git *) — a
moving target, not a one-time fix.- For MCP integrations configured in the web UI, it's worth checking what
actions are auto-approved by default and whether that's the right
default in either direction.
This needs someone to actually dig into current defaults and propose a
baseline allowlist — or a way to derive one from real usage (Claude Code
ships a fewer-permission-prompts-style skill that scans transcripts for
this) — rather than hand-picking commands ad hoc. Scoped out separately
since it's its own investigation, not a quick doc fix.
Two related friction points came up discussing Claude Code setup for
Serokell work:
settings.json'spermissions.allowlist could pre-approve obviouslysafe, frequently-used commands (e.g.
Bash(git status:*),mcp__youtrack__get_issue) so they stop prompting every time. But thereare a lot of these, they vary per engineer/project, and some run behind a
proxy (e.g.
rtk, which itself prompts to approvertk git *) — amoving target, not a one-time fix.
actions are auto-approved by default and whether that's the right
default in either direction.
This needs someone to actually dig into current defaults and propose a
baseline allowlist — or a way to derive one from real usage (Claude Code
ships a
fewer-permission-prompts-style skill that scans transcripts forthis) — rather than hand-picking commands ad hoc. Scoped out separately
since it's its own investigation, not a quick doc fix.