Default subagents spawn with ~60k tokens of inherited context; workflow subagents receive the user's latest chat line as their governing instruction

Environment: Claude Code desktop app (Windows 11), model claude-fable-5-1, session with ~10 MCP connectors attached, Workflow and Agent tools.

Problem 1: spawn cost. A default subagent (Agent tool or Workflow agent() without agentType) starts with the parent's full context: system prompt, every tool schema including all attached MCP connectors, the skills listing, CLAUDE.md and hook output. Measured on a 396-character task ("read two files, return a type"): cache_creation_input_tokens: 60236 on the first call. A custom agent restricted to tools: Read, Glob, Grep on the same task: cache_creation_input_tokens: 5332. A file reader does not need Gmail, CRM, calendar, PDF or browser tool schemas. Two 26-agent workflows in one session cost about 3.2M tokens for one usable result, almost all of it spawn context.

Ask: make the lean profile the default for agent()/Agent when the task does not name tools, or at least strip MCP connector schemas and the skills listing from subagents unless requested; surface the per-spawn context size in the Workflow launch dialog.

Problem 2: relayed user message. Workflow subagents are handed the user's latest chat message verbatim as "the user's request", with precedence over the task text the orchestrating model wrote. In this session the latest line was "Hell no I'm not doing that" (the user declining to do a manual task themselves, and asking for agents to do it instead). 24 of 25 readers refused to open a file, the synthesis agent refused to write, and the run still cost the full spawn context per agent. The relayed line is frequently about something other than the subagent's task.

Ask: do not relay the raw latest chat line as an overriding instruction to subagents; relay the orchestrator's task as the instruction, or let the orchestrator state the user's intent explicitly.

Repro: any session with several MCP connectors; run Agent with a one-line read task and inspect cache_creation_input_tokens in the subagent transcript; then run a Workflow right after a user message that is a refusal about something else.