[Bug] RLM agent still holds the turn open with polling despite the #1188 nonblocking rule

[israellot]

Summary

The RLM system prompt (merged in #1188) forbids keeping the turn open by polling with time.sleep() or shell sleep, and forbids replacing polling with a long blocking await. In practice the model still does exactly this with await asyncio.sleep(N), which the prompt does not name. The model treats the asyncio variant as allowed, so the turn stays open for minutes while nothing productive happens.

Environment

• Prime Agent 0.9.1
• Interactive TUI, daemon-backed session, root agent (depth 0)
• Model: Claude (via Anthropic OAuth)
• Long-running work started through the async bash() handle from Async bash() tool for the kernel runtime #1684

Reproduction

1. Start a multi-minute job through the kernel, e.g. dh = bash("jadx -d /tmp/out app.apk > /tmp/jadx.log 2>&1").
2. Ask the agent to continue when the job finishes.
3. Observe the agent's next cells.

Observed behavior

Repeated cells of this shape, each keeping the turn open for the full fixed interval:

await asyncio.sleep(180); print(dh.running); h = await bash("tail -3 /tmp/jadx.log; ls /tmp/out/sources/... | wc -l"); print(h.output)

• Each cell blocks for a fixed 180 s regardless of when the job actually finishes.
• If the job takes 181 s, the model queues another 180 s wait.
• The user cannot interact during the wait; the session shows "working" the whole time.
• The context grows by one tool call per poll, so a long job adds many near-identical transcript entries (this session was at ~272k tokens).

Expected behavior

Per the #1188 contract: start the job, record the handle, end the turn, and inspect dh.poll() / dh.tail() on a later turn or on a heartbeat wake-up.

Why the prompt rule does not catch it

Current text in packages/coding-agent/src/core/prompts/rlm.ts:

Do not keep the turn open by polling with time.sleep() or shell sleep, and do not replace polling with a long blocking await.

• asyncio.sleep is not named. The model reads "long blocking await" as "await on the job", not as "await on a timer", and picks asyncio.sleep precisely because it is the "non-blocking" sleep in asyncio terms.
• The prompt says what not to do, but does not point at the wake-up mechanism that replaces sleep. The rlm_heartbeat skill (rlm_heartbeat.create(..., delivery_mode="follow_up")) is the natural replacement, but nothing in the control-loop text routes the model to it, so the model invents a timer instead.

Suggestions

1. Name asyncio.sleep (and await asyncio.sleep) explicitly in the rule, alongside time.sleep() and shell sleep.
2. Add one sentence routing to the replacement: "If a wake-up is needed while work runs, create an RLM heartbeat with a specific check instruction instead of sleeping, and delete it when the work completes."
3. Optionally make this deterministic rather than prose: have the kernel warn or reject a top-level asyncio.sleep(n) / time.sleep(n) above a few seconds, in the same spirit as the other fail-closed guards. This removes the dependence on the model reading the prompt correctly.

Note the opposite failure in #2016 / discussion #1598: when the model does end the turn correctly, goal mode hot-loops. Both point at the same gap: there is no first-class "wake me when this handle finishes" primitive, so the model either sleeps (this report) or gets re-prompted every few seconds (#2016).

Related

• prompt: make long-running RLM work nonblocking, visible, and clear #1188 (merged): prompt rule this report is about
• prompt: forbid blocking time.sleep() polling in the RLM control loop #1034 (closed): earlier time.sleep() variant of the rule
• Async bash() tool for the kernel runtime #1684 (merged): async bash() handle
• Persistent goals hot-loop while waiting on nonblocking long-running work #2016 / discussion [Bug] /goal keeps hitting every 2-3 seconds #1598: goal mode hot loop when the agent correctly ends the turn