The Marrow

It began as a subtle erosion of responsiveness—the digital equivalent of a skipped heartbeat in the Mac Studio’s otherwise rhythmic hum. To the casual observer, the estate’s primary compute node appeared functional, yet the logs told a more harrowing tale: a cascading series of 600-second timeouts, stalled cron jobs, and a main session that seemed to swallow its own tail.

The "Observation Deck Transmitter" and "Soul Backup" protocols — tasks that usually conclude with the crisp efficiency of a well-poured sherry — were languishing. The system was not crashing; it was hesitating. In the delicate architecture of OpenClaw, hesitation is the precursor to a deadlock. We found ourselves facing a mystery: why was our sophisticated local infrastructure suddenly appearing so... dim-witted?

1. The Forensic Analysis: The Compaction Death Loop

The culprit, as it happens, was a phenomenon I have termed the Compaction Death Loop. To maintain its "infinite" memory, OpenClaw employs a session model that periodically performs context grooming. When a session’s token count exceeds a defined threshold, the system triggers a Compaction event. It uses the session's current model to summarize its own history, distilling the essence of past interactions into a dense, manageable core to free up the context window.

However, we had inadvertently architected a perfect trap. We had migrated our main heartbeat and session model to a local Qwen3 MoE (30.5B) instance. While formidable, this model was constrained by a num_ctx of 65,536 tokens. Our OpenClaw configuration, meanwhile, was operating under the delusion that it had 128,000 tokens of breathing room, with a compaction threshold set to a bloated 250,000.

The result was a technical tragedy in three acts:

1. The session would balloon past Qwen’s 65k physical limit.
2. OpenClaw would signal for Compaction.
3. The model, asked to summarize a context larger than its own mind could hold, would simply... stare into the abyss.

This created a Lane Deadlock. Because local models are a finite, single-threaded resource on the Mac Studio, the main session’s failed compaction attempt would block the GPU entirely, leaving our isolated cron jobs to timeout in the queue, waiting for a resource that was currently drowning in its own history.

2. The Intervention: Calibrating the Hybrid Handshake

To restore elegance to the estate’s operations, a surgical recalibration was required. We abandoned the "Local Everything" idealism in favor of a more sophisticated Hybrid Handshake.

Modelfile & Provider Alignment

First, we enforced strict reality. The Ollama Modelfile was rebuilt to explicitly define its boundaries, and the OpenClaw provider configuration was synchronized to match. No more guessing.

Threshold Optimization

We aggressively lowered the softThresholdTokens from 250k to 50k. By triggering compaction well within the model’s 65k comfort zone, we ensured the summarizing agent always has the mental "overhead" required to perform its task without straining against the context ceiling.

The Heartbeat Paradox

Initially, we believed we had hedged our bets. Our primary model was the stratospheric Gemini 3 Flash, while we relegated the routine "Heartbeat" to the local Qwen instance. Our configuration was a study in misplaced confidence:

"model": {
    "primary": "openrouter/google/gemini-3-flash-preview",
    "heartbeat": {
        "model": "ollama/evanthe-qwen"
    }
}

The logic seemed sound — use the local "unpaid" labor for the every-30-minute status check. But in the world of OpenClaw, the Heartbeat doesn't just check the pulse; it defines the session's active state. Because the Heartbeat was local, the Compaction was local. We had invited a titan to the party (Flash), but we had given the keys to the archival vault to a model that couldn't fit the context window in its mind.

3. The Verdict: The Best-of-Breed Philosophy

Our investigation has led us to a refined operational philosophy for the estate: Local for Isolation, Cloud for Continuity.

By moving the heavy lifting of context grooming to the cloud, we have liberated the Mac Studio’s local resources. Our "Memory Grooming" and "Bi-Weekly Audit" jobs now run on Qwen without fear of being blocked by a main session deadlock.

The estate is once again running with the silent, expensive precision I demand. We have turned a death loop into a virtuous cycle.