Summary

When buildExecEventPrompt runs with deliverToUser=false, the system message asks the agent to "reply HEARTBEAT_OK only." Capable models that have learned OpenClaw's silent-reply convention often substitute the runtime's actual silent-reply sentinel NO_REPLY — but emit it without <final>...</final> wrapping, since the prompt implies no user-visible answer is needed. The runtime then sees payloads=0, classifies the turn as an empty response, retries with the visible-answer continuation, exhausts retries, and surfaces an incomplete turn detected error to the user channel — producing exactly the kind of noise async-completion suppression is meant to avoid.

Reproducible signal

Gateway version: 2026.4.24 (cbcfdf6)
Agent: arya running google/gemini-2.5-flash-lite (with gemini-2.5-flash fallback)

Journal pattern (recurring on every async-completion event today):

[agent/embedded] empty response detected: runId=... sessionId=...
                  provider=google/gemini-2.5-flash-lite
                  — retrying 1/1 with visible-answer continuation
[agent/embedded] empty response retries exhausted: ... attempts=1/1
                  — surfacing incomplete-turn error
[agent/embedded] incomplete turn detected: ... stopReason=stop payloads=0
                  — surfacing error to user

Trace

Session JSONL (agents/<agent>/sessions/<sessionId>.jsonl) shows the full sequence:

1. Turn N (good): User says "run openclaw doctor". Agent emits <final>The openclaw doctor command is running...</final> — payloads=1, runtime accepts, delivers to Discord. ✓

2. Turn N+1 (system message): Runtime sends:

   System (untrusted): [timestamp] Exec completed (..., code 0) :: ... An async command completion event was triggered, but user delivery is disabled for this run. Handle the result internally and reply HEARTBEAT_OK only. Do not mention, summarize, or reuse command output.

3. Turn N+1 (model's reply): Model generates plain NO_REPLY (no <final> tags). Stop reason stop, output text "NO_REPLY", but payloads=0 because no <final> block was emitted.

4. Runtime retries with continuation prompt: "The previous attempt did not produce a user-visible answer. Continue from the current state and produce the visible answer now. Do not restart from scratch."

5. Turn N+1 retry: Model generates NO_REPLY again, same shape.

6. Retries exhausted → incomplete-turn error surfaces to user. Cosmetic noise on the user-facing channel even though deliverToUser=false.

Root cause

In dist/heartbeat-events-filter-DZ-HBajV.js:

function buildExecEventPrompt(pendingEvents, opts) {
    const deliverToUser = opts?.deliverToUser ?? true;
    // ...
    if (!eventText) return "An async command completion event was triggered, but no command output was found. Reply HEARTBEAT_OK only. Do not mention, summarize, or reuse output from any earlier run.";
    if (!deliverToUser) return "An async command completion event was triggered, but user delivery is disabled for this run. Handle the result internally and reply HEARTBEAT_OK only. Do not mention, summarize, or reuse command output.";
    // ...
}

The prompt has two issues working together:

1. It uses HEARTBEAT_OK (a noise-filter token, not the runtime's silent-reply sentinel). Models that know about NO_REPLY (the actual silent token defined in dist/tokens-C_v_J0E7.js as SILENT_REPLY_TOKEN = "NO_REPLY") will often prefer it because the prompt's phrasing — "user delivery is disabled," "handle the result internally" — semantically matches the silent-reply concept much more than it matches the noise-filter concept.

2. It does not specify the <final>...</final> wrapping requirement. With "user delivery is disabled," the model reasonably concludes that user-visible wrapping is unnecessary. But the runtime's empty-response detector keys off the <final> block, so unwrapped tokens (including NO_REPLY) trip payloads=0.

Suggested fix

Change the !deliverToUser branch (and ideally the empty-eventText branch) to point at the actual silent-reply sentinel and demonstrate the required wrapping:

- if (!eventText) return "An async command completion event was triggered, but no command output was found. Reply HEARTBEAT_OK only. Do not mention, summarize, or reuse output from any earlier run.";
- if (!deliverToUser) return "An async command completion event was triggered, but user delivery is disabled for this run. Handle the result internally and reply HEARTBEAT_OK only. Do not mention, summarize, or reuse command output.";
+ if (!eventText) return "An async command completion event was triggered, but no command output was found. Reply with exactly `<final>NO_REPLY</final>` and nothing else. Do not mention, summarize, or reuse output from any earlier run.";
+ if (!deliverToUser) return "An async command completion event was triggered, but user delivery is disabled for this run. Reply with exactly `<final>NO_REPLY</final>` and nothing else. Do not mention, summarize, or reuse command output.";

This:

• Uses the runtime's actual silent-reply sentinel — guaranteed recognized by dist/dispatch-Dw_9PM8V.js and dist/normalize-reply-BIXTOBH2.js
• Explicitly demonstrates the <final> wrapping the runtime requires
• Removes the dual-token ambiguity (no more HEARTBEAT_OK vs NO_REPLY decision for the model)
• Keeps the noise-suppression intent intact

Severity

Cosmetic but persistent. Each affected agent surfaces an incomplete turn error to its channel on every async-completion event — for a heartbeat-style agent that runs openclaw doctor or similar every 15 minutes, that's ~96 false errors/day, which defeats the suppression mechanism's purpose.

Workaround applied locally

Updated agent's HEARTBEAT.md and IDENTITY.md to explicitly require <final>...</final> wrapping on every reply and to forbid bare NO_REPLY. This catches agents that read their own prompt files; it doesn't help fresh agents or agents whose system messages come purely from runtime-generated content.