Problem
--bare mode strips all auto-discovery, including OAuth and keychain authentication. This makes it unusable for tools that need clean, context-free API calls but rely on the CLI's built-in auth (e.g. team/company subscriptions where there is no explicit API key).
Use case
We maintain skill-validator, a CLI tool that scores Agent Skill files using an LLM-as-judge pattern. Each file is scored in isolation through a single prompt+response pair. We added a claude-cli provider that shells out to claude -p so users can score skills without managing API keys.
The scoring system is designed so that each call contains only a system prompt (the scoring rubric) and the file content. Local context like CLAUDE.md files, project memory, and rules act as hidden confounders that can bias scores and make them non-reproducible across environments.
--bare would solve this perfectly, except it also disables OAuth/keychain auth, which is the primary auth mechanism for users who don't have an explicit ANTHROPIC_API_KEY. This forces a choice between clean context isolation and working authentication.
Proposal
One of:
- A flag like
--no-context that skips CLAUDE.md, memory, rules, hooks, and plugins but preserves the full auth chain (OAuth, keychain, API key).
- A
--bare sub-option (e.g. --bare --auth) that re-enables auth discovery while keeping everything else stripped.
- Allow
--bare to read OAuth/keychain credentials while still skipping all other auto-discovery.
Current workaround
We use claude -p without --bare and document that scores from the claude-cli provider may be less consistent than API-based providers due to injected local context.
Problem
--baremode strips all auto-discovery, including OAuth and keychain authentication. This makes it unusable for tools that need clean, context-free API calls but rely on the CLI's built-in auth (e.g. team/company subscriptions where there is no explicit API key).Use case
We maintain skill-validator, a CLI tool that scores Agent Skill files using an LLM-as-judge pattern. Each file is scored in isolation through a single prompt+response pair. We added a
claude-cliprovider that shells out toclaude -pso users can score skills without managing API keys.The scoring system is designed so that each call contains only a system prompt (the scoring rubric) and the file content. Local context like CLAUDE.md files, project memory, and rules act as hidden confounders that can bias scores and make them non-reproducible across environments.
--barewould solve this perfectly, except it also disables OAuth/keychain auth, which is the primary auth mechanism for users who don't have an explicitANTHROPIC_API_KEY. This forces a choice between clean context isolation and working authentication.Proposal
One of:
--no-contextthat skips CLAUDE.md, memory, rules, hooks, and plugins but preserves the full auth chain (OAuth, keychain, API key).--baresub-option (e.g.--bare --auth) that re-enables auth discovery while keeping everything else stripped.--bareto read OAuth/keychain credentials while still skipping all other auto-discovery.Current workaround
We use
claude -pwithout--bareand document that scores from theclaude-cliprovider may be less consistent than API-based providers due to injected local context.