print mode (-p) aborts in-flight rlm.spawn() children instead of waiting for them

[hasan-emre-dinc]

Bug: -p (print mode) aborts in-flight rlm.spawn() children instead of waiting for them

Summary

A parent session that spawns child subagents via rlm.spawn() and ends its turn — the documented, expected pattern — has its process torn down the moment the parent's own turn goes idle, if invoked through -p (print mode). Any children still in flight get aborted before they can send their agent_message reply, so the parent never receives their reports.

Root cause

packages/coding-agent/src/modes/print-mode.ts calls:

const autonomousStatus = await connection.waitForHeadlessCompletion();

with no options. waitForHeadlessCompletion() (packages/coding-agent/src/modes/headless-completion.ts) takes a waitForRlmQuiescence?: boolean option specifically to choose between:

• session.waitForHeadlessIdle() — waits only for the parent's own turn/compaction to settle (no descendant awareness), or
• session.waitForRlmQuiescence() — "Wait for every admitted descendant run to publish its terminal parent message and for the resulting parent turns to drain" (agent-session.ts).

Since print mode never passes { waitForRlmQuiescence: true }, it always takes the first path. This is inconsistent with the rest of the codebase — acp-mode.ts's finalizePendingTerminal path explicitly passes { waitForRlmQuiescence: true } for exactly this reason, and daemon-mode.ts's RPC handler exposes it as a caller-controlled parameter. Print mode is the one entry point missing it.

Reproduction

prime-agent --provider <any> --model <any> -p "Using rlm.spawn, delegate three parallel independent subagents to review three files, one each. Wait for each subagent's agent_message reply, then combine all three reports into one summary."

Observed:

• The CLI prints only the "I've spawned them, waiting for replies" acknowledgment, then exits (exit code 0) — never the consolidated report.
• Inspecting ~/.prime/agent/session-artifacts/<parent>/sub-*/rlm-subagent.json and each child's own session .jsonl: all three children reach "status": "completed", but each was aborted (stopReason: "aborted", errorMessage: "Request was aborted") right after reading its assigned file and before ever calling agent_message.send.
• The parent's own session log records a [child-exited: no-reply] custom message (kind: "completed_without_reply") for at least one child.
• Child creation/completion timestamps confirm genuine parallel execution (all created within ~226ms of each other) — the parallelism itself works correctly; only the parent's premature exit is the problem.

Fix (verified locally)

- const autonomousStatus = await connection.waitForHeadlessCompletion();
+ const autonomousStatus = await connection.waitForHeadlessCompletion({ waitForRlmQuiescence: true });

Re-ran the exact repro after this change: all three children reach agent_message.send (confirmed in each child's own transcript), zero abort/no-reply notices, and the parent produces the actual consolidated report combining all three subagent findings.

npm run check (biome + tsgo + installer check + browser smoke) passes clean with the change.

Environment

• macOS (arm64), Node 26.5.0
• Source checkout, main branch
• Provider: custom OpenAI-compatible provider via models.json (not provider-specific — root cause is in the completion-wait logic, unrelated to any provider)

Happy to open a PR with this one-line fix + a regression test if a maintainer wants it — keeping this to a Discussion first per CONTRIBUTING.md.

[awhite0030]

Investigated this report and produced a candidate fix, validated on a fork of this repository.

Root cause, fix and validation from the working notes:

Root cause: In packages/coding-agent/src/modes/print-mode.ts, connection.waitForHeadlessCompletion() was called without { waitForRlmQuiescence: true }, which caused print mode to not wait for subagent replies.
Fix: Update the call to pass { waitForRlmQuiescence: true } in packages/coding-agent/src/modes/print-mode.ts.
Validation: Created a regression test test/suite/regressions/2270-print-mode-wait-rlm-quiescence.test.ts to assert that waitForHeadlessCompletion is called with this configuration. Ran test suite via npx tsx ../../node_modules/vitest/dist/cli.js --run test/suite/regressions/2270-print-mode-wait-rlm-quiescence.test.ts and successfully verified against full repository using npm run check.
Upstream discussion: #2270

Diff: awhite0030/prime-agent@main...fix-2270-print-mode-wait-rlm-5064527283883454858
Full pull request: awhite0030#76

The change passes npm run check and the focused regression tests against current main. If a fix for this is still wanted, I can open a pull request.

[trinode]

I just hit this issue now, running in -p mode

"I've dispatched all 5 workers and they're converting the integration test files in parallel. I'll consolidate their results as they report back. Ending my turn to let them work."

Then instantly terminates the subagents.

Took me a while to get to the conclusion it was prime-agent terminating itself (I was checking for out of memory or context space issues that could be causing errors)

I'd personally love this to be merged in as it's a breaking issue for my workflow.