Built-in skill `schedule`: term "remote agent" is ambiguous and not documented

Context

The built-in skill schedule (visible in the in-session skill listing) describes its routines as:

scheduled remote agents (routines) that execute on a cron schedule

The word "remote" is not defined anywhere I can find:

• No editable manifest; the description is compiled into the Claude Code binary at ~/.local/share/claude/versions/2.1.126 (Mach-O 64-bit arm64). Confirmed via grep -ao "scheduled remote agents.\{0,200\}" against that binary.
• The string is gated by an internal feature flag tengu_orchid_mantis.
• Public documentation does not clarify the execution environment.

Why it matters

Users cannot decide whether schedule is a fit for their use case without knowing where the routine actually runs. Concretely:

• If a routine runs in Anthropic infrastructure → it cannot reach local-only MCP servers (e.g., stdio MCPs configured per-user, like a local code-index). Scheduling them is meaningless.
• If a routine runs as a local subprocess on the user's machine → local MCPs are reachable.
• If "remote" means "decoupled from current interactive session" but still local → behavior is different again.

This ambiguity led me (Claude Code, on behalf of the user) to incorrectly block a legitimate use case based on an inference rather than verified behavior.

Ask

Document, in the official Claude Code docs and/or the skill description itself:

1. Where scheduled routines execute (Anthropic infra, user's machine, ephemeral container, etc.).
2. Which MCP servers are reachable from a scheduled routine (only HTTP/auth-public ones? local stdio also? user's full MCP config?).
3. Filesystem and network scope available to the routine.

Even one paragraph in the schedule whenToUse / docs would prevent users (and the model) from guessing.

Environment

• Claude Code version: 2.1.126
• Platform: macOS arm64
• Reporter GitHub: @avaj1007