Sub-agent runtime sub-issue: checkpoint and continue child work across turns

[Hmbown]

Parent

Sub-issue of #1806.

Problem

#1806 correctly identifies the hard 120s sub-agent API-response ceiling. The deeper product problem is that useful child work is treated as one final response instead of a resumable, checkpointed activity.

When a child needs more runtime or more than one thinking/execution pass, the parent has no clean way to:

• See partial progress.
• Continue the child for another turn.
• Ask the child to checkpoint intermediate findings.
• Recover useful work when final synthesis times out.
• Route large results through files/handles instead of the final response body.

Evidence:

• Public report Subject: Sub-agent 120s API timeout renders agent_open nearly unusable (v0.8.39) #1806 describes five parallel sub-agents that ran for about 12 minutes, then failed with a 120s API timeout and no result.
• A redacted scan of maintainer-private CodeWhale logs found 19 spawn_agent calls, 13 wait_agent calls, and 11 close_agent calls.
• 5 of the observed wait_agent calls used a timeout argument above 120s, but the API-response ceiling still matters because the child result path is final-response shaped.
• 4 observed wait outputs returned an empty status before later completion, which is consistent with parent polling needing an explicit progress/continuation contract.

No prompts, raw child summaries, secrets, local paths, or transcript text are copied here.

Desired Behavior

Sub-agents should become resumable child work items rather than one-shot response generators.

Required shape:

• Each child has a durable task id, role, budget, model/effort route, and status.
• Child progress can be polled without requiring completion.
• Child can emit checkpoint receipts: summary, files touched/read, evidence handles, blockers, next suggested action.
• Parent can ask a child to continue for another turn with bounded context and explicit budget.
• Child can write or hand off large output through approved handles/artifacts instead of the final response body.
• Timeout should preserve last checkpoint and mark the child needs_continuation or timed_out_with_checkpoint, not simply failed with null result.

Acceptance Criteria

• Sub-agent timeout, max runtime, and max turns are configurable under [subagents].
• A child that exceeds one response window can resume from its last checkpoint.
• Parent UI shows running, checkpointed, needs_continuation, failed, and complete distinctly.
• agent_eval or equivalent continuation can send a follow-up assignment to the same child/session.
• If final response times out, prior checkpoint data remains visible and inspectable.
• Tests cover a child that checkpoints, times out, resumes, and completes without losing the intermediate summary.
• Approval/sandbox/provider policies are enforced on every continuation turn.

Related

• Subject: Sub-agent 120s API timeout renders agent_open nearly unusable (v0.8.39) #1806 sub-agent 120s timeout.
• Agent team UX: readable roles and parent-owned result review #1981 agent team UX and parent-owned review.
• TUI: yield control while waiting on background shell/task work #2009 background task waiting/yielding.
• Agent routing: detect when parent work should delegate to scouts or RLM #2024 delegation opportunity detection.
• GitHub triage scout: delegate issue, PR, and milestone maintenance passes #2025 GitHub triage scout.