[Bug] Subagent delegation lacks timeout, monitoring, and abort controls—caused 12+ hour session hang

Bug Description
Title: Subagent delegation has no timeout, no monitoring, and no sanity checks — caused multi-hour session failure

During a session involving GitHub Actions workflow review, plan mode was activated. Claude Code was asked to verify a claim it had made about a specific GitHub Action.
This required a single web search that would have completed in under 10 seconds via a direct WebFetch call.

Critical Issue - Zero sanity checks on subagent execution:

When Claude Code spawns a subagent, there is no mechanism available to the user or to Claude Code itself to bound, monitor, or abort that subagent's execution.
Specifically:

• There is no user-configurable timeout. Users cannot set a parameter such as "this subagent task should not exceed 60 seconds."
• The main Claude Code process has no check-in interval. It simply waits for the subagent to return with no periodic status reporting.
• There is no abort trigger. Once a subagent is spawned, the user cannot cancel it.
• There is no progress visibility. The user has no way to see what the subagent is currently doing or how far along it is.
• The subagent starts with zero knowledge of Claude Code's prior conversation context, including the specific claim or suggestion that prompted the search. This means

the subagent cannot self-correct or stay anchored to the original intent — it only knows what Claude Code wrote in the prompt it was handed.

A subagent will run until it completes its task or exhausts its own context limit. In this session, a subagent tasked with a single web lookup instead made 10+
sequential web fetches because Claude Code wrote an over-scoped prompt and there was nothing — no timeout, no check-in, no guardrail — to catch or stop it. The user had
no ability to detect the problem or intervene. The session ran over 12 hours.

This is a fundamental reliability gap. Users have no way to "ground" a subagent in terms of runtime expectations, and Claude Code has no way to self-detect that a
delegated task has gone off the rails.

Secondary Issue - Plan mode forces subagent delegation with no user transparency:

Plan mode in Claude Code requires ALL research tasks to be routed through Explore subagents. Claude Code cannot make direct tool calls such as WebFetch while plan mode
is active. When this constraint forces a task to be slower or more indirect than it would otherwise be, Claude Code does not inform the user. The user has no visibility
into why something is taking long and no opportunity to exit plan mode before the delegation occurs.

This constraint exists for legitimate reasons — large codebase exploration benefits from subagent isolation. But applying it to a clearly bounded single-URL lookup is a
structural mismatch, and the user has no way to know it is happening.

What Anthropic should review:

• Implement user-configurable subagent timeouts — the single most important missing control
• Implement a check-in mechanism — if a subagent exceeds a time threshold, surface a warning to the user with an option to abort
• Require Claude Code to disclose upfront when a mode constraint is forcing subagent delegation instead of a direct tool call, and what the user's options are
• Evaluate whether plan mode's subagent requirement should apply to clearly bounded single-tool operations
• Consider whether subagents should receive a summary of the conversation context that prompted them, so they can stay anchored to the original intent

Impact: User lost 12+ hours of a session to what should have been a 30-minute task. The absence of subagent sanity checks is a significant operational gap that will affect any sufficiently long or complex delegated task. In time-sensitive or higher-stakes contexts this is a serious reliability and safety concern.