[Bug] Subagents (Agent tool) hit context limit without auto-compact ever triggering

Bug Description

Subagents spawned via the Agent tool (with run_in_background: true) hit "Context limit reached" and terminate without auto-compact ever triggering. The main session receives Context limit reached · /compact or /clear to continue from the subagent, and the subagent's work is lost.

This is distinct from #24591 (compaction fails with subagents) — in this case, compaction never activates at all within the subagent before it dies.

Steps to Reproduce

1. Launch 4-5 background subagents in parallel using the Agent tool with run_in_background: true
2. Each subagent performs research-heavy tasks (web searches, file reads, large text generation)
3. Wait for subagents to complete

Expected: Each subagent auto-compacts its own context when approaching limits and continues working.

Actual: Multiple subagents hit Context limit reached simultaneously and terminate. The main session displays:

⏺ Agent "task name" completed
  ⎿  Context limit reached · /compact or /clear to continue

In my case, 4 out of 5 background agents died this way in the same session.

Additional Context

• After this happens, the main session's context is also heavily consumed by the failed subagent results, pushing it near its own limit
• /compact on the main session also fails at this point ("Conversation too long"), creating a total deadlock
• The only recovery is /clear, which loses all session state
• This effectively makes parallel background agents unreliable for any non-trivial research tasks

Environment Info

• Claude Code version: 2.1.63
• Platform: macOS 26.2 (darwin, arm64)
• Model: claude-opus-4-6

Related Issues

• #24591 — Context compaction fails with multiple subagents (compaction fails; in our case it never triggers)
• #24976 — Context Limit Reached error should auto-compact and queue user message instead of failing
• #26317 — Context compaction fails with 'Conversation too long'
• #23047 — /compact fails with 'Conversation too long'