[FEATURE] Routines with connectors get permanently blocked on tool-call approval, even in "Auto mode"

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

I created several Routines (via the create_trigger/update_trigger MCP API, exposed through Claude Code) that use MCP connectors (Gmail, Google Drive, Zapier, Google Calendar) and are meant to run fully unattended on a cron schedule.
Every Routine's Permissions tab shows: "Claude created this routine, so it runs in Auto mode — connector calls are checked by a classifier."
In practice, this classifier blocks execution: firing these Routines (both via scheduled cron and via fire_trigger) repeatedly leaves the session in SESSION_STATUS_REQUIRES_ACTION, waiting on a pending_action for ordinary read calls — e.g. mcp__Gmail__search_threads, mcp__Google-Drive__search_files, mcp__Zapier__inspect_zapier_actions — with no one available to approve it, since the Routine runs unattended.
I tried adding permission entries to the repository's .claude/settings.json (both connector-UUID-qualified and friendly-name-qualified tool patterns, e.g. mcp__Gmail__*), but this had no effect — the same tool call still triggers the same pending approval. There does not appear to be any UI toggle to grant a Routine full auto-approval for its connector tools.
Question: Is there a supported way to configure a Routine so its connector tool calls (or a specific allow-listed subset) are always auto-approved during unattended runs, without a human clicking approve each time? If this is an intentional safety boundary rather than a bug, documentation on the "Auto mode" classifier behavior would help — right now it's unclear what triggers a review versus an auto-approval.

Proposed Solution

?

Alternative Solutions

?

Priority

Critical - Blocking my work

Feature Category

CLI commands and flags

Use Case Example

?

Additional Context

?