bug(subagents): --print/--json parent can settle while required children are still running [under investigation]

[tt-a1i]

Status

Wiring landed on research branch research/openpi-cursor-reliability-20260908 @ 7950570 and PR #515. PR is not merged. Isolated evidence only. Not a formal Benchmark.

Impact

In pi --print / --mode json, a parent that calls subagent_spawn and then ends the turn emits agent_settled and the process exits. Automatic result delivery only helps an interactive TUI. Required child results are lost.

Evidence — same E task+acceptor, cursor-grok-4.6-high-fast

Receipts still say source=a3edee28 because the harness hardcoded that SHA. Split is by spawn-result text + attempt window. Index: /tmp/openpi-reliability-20260908/results/e-high-fast-512-version-split.md.

Window	Product behavior	n	Correct
E-01..10	stock v0.8.1 spawn (“end your turn” / auto re-invoke). E-06..10 are still this window, not after-fix.	10	7
E-11..20	print-host spawn from 7950570 (“MUST call subagent_wait”). All 10 stdout traces show that text.	10	10

E-01 spawn result (trimmed): “end your turn; its result is delivered automatically and you are automatically re-invoked”. E-11 spawn result: “This host cannot re-invoke you after you end the turn. You MUST call subagent_wait…”

Isolated bun run check exit 0 (results/isolated-bun-check.log) and bun run test exit 0, 1470 pass (results/isolated-bun-test.log).

Fix (on the research branch only)

• buildSubagentSpawnResult({ canDeliverLater: false }) requires wait and does not say “end your turn”.
• On agent_settled, if the host cannot deliver later and direct children are still running, send up to two followUp + triggerTurn barriers.

Do not disable tools, loosen permissions, or lengthen timeouts.

Applicability

OpenPI v0.8.1 subagent spawn/delivery. Not Cursor native-tool rejection (PR #484). Related hung-after-success: #516.

[tt-a1i]

Isolated worktree now has a failing-repro-first helper plus wiring (not merged):

• extensions/subagents/src/print-host.ts — TUI-only canDeliverLater, pending follow-up text
• buildSubagentSpawnResult({ canDeliverLater: false }) requires subagent_wait and does not say “end your turn”
• --print/json/rpc parent agent_settled with running direct children sends up to two followUp + triggerTurn barriers

Unit tests: tests/extensions/subagents/print-host.test.ts and the new prompt contract. Same E task+acceptor has 10 before-samples on v0.8.1 stock spawn text (3 print-settle failures, 7 waits). After-sample revalidation is in progress. No merge/release.

[tt-a1i]

Before/after on the same E task+acceptor, cursor-grok-4.6-high-fast, isolated v0.8.1:

• Before (stock spawn text): 7/10 correct, 3/10 print-settle wrong result (attempts 01–10).
• After (print-host spawn text + bounded follow-up barrier): 10/10 correct (attempts 11–20). E-11 spawn result includes You MUST call subagent_wait and the model called Pi subagent_wait.

PR opened from research/openpi-cursor-reliability-20260908. Isolated bun run check + bun run test green. Not merged.