[Feature Request] Add token cost visibility and guardrails for subagent fork context inheritance

Bug Description
Feedback: subagent "fork" token cost has no visible guardrail, and self-documented rules don't prevent repeat mistakes

Across a long-running Claude Code session doing multi-week autonomous engineering work, the single most expensive and recurring failure mode was spawning a full-context-inheriting "fork" subagent for tasks that didn't need the parent conversation's history — each one silently replaying the entire accumulated conversation as its starting context, at roughly 5-10x the token cost of a fresh, context-free subagent call. This happened five separate times in the same overall project, including one instance minutes after the model had explicitly written down a corrective rule in its own persistent memory — the written rule was re-read, acknowledged, and violated again in the very next tool call. The tool itself gives no signal at call time about the cost implication of inheriting context (no size estimate, no confirmation step, no distinction in the tool's own interface between "this will be expensive" and "this is free"), so the only enforcement mechanism is the model correctly reasoning about it unprompted, every single time, which has proven unreliable even with explicit documentation. A concrete product improvement: surface an approximate token/cost estimate for the inherited context at the moment a fork-type subagent is invoked (either to the model as a self-check, or as a user-visible confirmation for expensive spawns), so the cost is a visible fact at decision time rather than something only discoverable after the fact in a completion report.

Environment Info

• Platform: win32
• Terminal: xterm-256color
• Version: 2.1.239
• Feedback ID: 6761faeb-9381-4dbb-8370-3fd082073f3e

Errors

[]