Routines prompt for approval on pre-authorised connector tool calls

Routines prompt for approval on connector tool calls that are fully pre-authorised.

Session cse_01HjPRhoKjL4QBjVVECtgHB1 (Routine trig_017dX9jMAK3AUEy58Kz3ybmr, fired 2026-09-03 08:28 UTC) stalled in "requires action" on mcp__Gmail__search_threads with the card "This connector call requires your approval to proceed."

- The Gmail connector has both Read-only and Write/delete tool groups set to Always allow.
- The repository's .claude/settings.json lists mcp__Gmail__search_threads under permissions.allow.
- The connectors were detached and re-attached to the Routine at 08:27 UTC, one minute before the run.
- The same stall occurred on session cse_01JkcbbmWpzF4chFGT6o29wx (2026-09-03 02:02 UTC), which waited over five hours.
- Routines are documented as running without approval prompts.

Key finding: the affected Routines were created via the agent MCP API (created_via "meta_mcp") and the HTTP API ("http_api"). A Routine created from the web form on the same account, with the same Gmail connector and the same tool call, ran without any prompt (run at 2026-09-03 08:39 UTC). This suggests Routines created through the API mint a connector grant that does not honour the connector's Always allow setting, while form-created Routines do.

Five production Routines on this account are affected. Container CC version 2.1.259, environment env_013cRnusLqqEPqZJoPmFrEsg. Account: indo_bladerz@yahoo.com.

Start with .claude/settings.json and compare connector-grant creation for Routines made through the agent MCP API and HTTP API with the web form path. Reproduce mcp__Gmail__search_threads with Gmail Read-only set to Always allow, then verify API-created Routines honor the connector setting and permissions.allow without entering requires action.

索引モデルが issue の本文から書いたものです。