[BUG] Cloud routine hangs on a permission prompt for the built-in Artifact tool, contradicting "routines run autonomously"

A weekday routine created at claude.ai/code/routines invokes the first-party `/morning` skill, which renders a brief and publishes it as a private artifact. Every run parks in `REQUIRES_ACTION` on a pending permission prompt — `update the artifact "Morning Brief"` — with nobody awake to answer, so the run dies there and the closing summary never arrives either.

This contradicts the documentation. Routines are described as having "no permission-mode picker and no approval prompts during a run", and the scheduling comparison table lists Cloud routines as "Permission prompts: No (runs autonomously)."

The routine's stored config carries `allowed_tools: [Bash, Read, Write, Edit, Glob, Grep, WebFetch, WebSearch]` — no `Artifact`. A routine created through the MCP `create_trigger` path instead stores `config: {}` with no tool list at all, so the restriction appears to follow the creation surface.

There is no user-facing workaround: routines expose neither a permission-mode picker nor a tool picker; project `permissions.allow` doesn't gate `Artifact` (tested); `permissions.defaultMode: auto` is documented not to take effect from project or local settings; and environments have no permission-mode variable, with setup scripts skipped whenever a cached environment exists.

**Why it matters:** the shipped `/morning` skill's own description offers to "set it up as a recurring weekday task", and its output is an artifact. So a first-party skill's advertised scheduled use case can't work on cloud routines.

**Expected:** either routines include `Artifact` in their tool set, or a run fails loudly instead of parking indefinitely on a prompt no one can answer.

Possibly related: #88731 (Artifact tool absent in `sdk-cli` entrypoint sessions), #61196 (closed; the MCP-connector analog on remote routines).

Start at the cloud routine creation path at claude.ai/code/routines and compare it with the MCP create_trigger path; inspect the stored tool configuration and reproduce the REQUIRES_ACTION state for the Artifact prompt. Done means a routine using the first-party /morning skill completes without an unanswered prompt, or fails loudly with a clear result.

由索引模型根据 Issue 内容生成。