[Bug] Context limit reached during parallel subagent execution prevents result aggregation

Bug Description
I unexpectedly burned through 4% of the (larger) max plan with a single prompt, which the parent agent was not able to synthesize because the subagents' responses could not be aggregated by the agent (anonymized excerpt below):

  ⏺ Task Output(non-blocking) ad488fa
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) abd3530
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) aeeeaa1
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) a64efda
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) aad6d02
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) ad8a2f3
    ⎿  Read output (ctrl+o to expand)

  ⏺ Task Output(non-blocking) a42b912
    ⎿  Read output (ctrl+o to expand)
    ⎿  Context limit reached · /compact or /clear to continue

  ✻ Brewed for 1m 13s

  ⏺ Agent "Assess project-alpha build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-beta build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-gamma build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-delta build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-epsilon build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-zeta build" completed
    ⎿  Context limit reached · /compact or /clear to continue

  ⏺ Agent "Assess project-eta build" completed
    ⎿  Context limit reached · /compact or /clear to continue

The session setting was set to auto-compaction. My lesson learned was to be more explicit about telling the agent to instruct subagents to write their reports to a file. I'm reporting this incident because the product could be more robust to handle this scenario when auto-compaction is configured.

The prompt causing this was (anonymized):

systematically assess the current state of maven build for all of these projects. Use parallel subagents. You can use up to 1000 tool calls. Provide a template for the subagents to populate so we have a consistent format for the reports. Some projects are multi-module. Expect some maven dependencies to fail to resolve. The subagents should also make note of what they learn about JRE versions.

Environment Info

• Platform: darwin
• Terminal: ghostty
• Version: 2.1.19
• Feedback ID: 862620be-1bd0-48d9-9684-a9c1f5df646f