macfleet-dash v0.10.0: AI engines join the fleet view

Published Oct 29, 2026

1. READ Cached source reports The optional connector reads existing reports rather than prompting a model.
2. KEEP CONTEXT Identity and original time Explicit engine mappings and source timestamps keep each reading interpretable.
3. DISPLAY Engine and computer views Reported model, rates, slots and resource readings answer their own questions.
4. INSPECT Token windows and history Read selected time windows, gaps and retained values with their limits.

My fleet dashboard started with a practical question: which computers need attention? Local AI adds another: what is the inference engine reporting while those computers are doing the work?

macfleet-dash v0.10.0, released October 20, brings those views closer together. It adds an optional, read-only SparkDash connector, AI engine cards, token windows and longer resource histories. This write-up describes the public release. It does not certify the current state of every computer in my lab.

The computer and the engine answer different questions

A computer can be reachable while its inference service is unavailable. A model server can report a model without demonstrating that a useful agent task completed. The interface needs room for both observations.

The new engine cards can show the reported model and backend, decode and prefill rates, running and queued requests, and occupied versus total slots. They read cached SparkDash reports through the collector. The browser does not connect directly to a private inference endpoint.

An operator can explicitly associate an engine with an existing computer. That association is checked against its configured address; the integration does not create another computer or inflate fleet totals.

This is useful when deciding whether a delay coincides with a busy engine, a growing queue or constrained host resources. The readings help an investigation; the actual task still supplies the completion evidence.

Token windows need a time boundary

The release adds one-hour, six-hour and 24-hour engine token views, including totals by machine and model. Partial hourly UTC buckets are identified rather than presented as complete elapsed windows.

Engine traffic also stays separate from Hermes/client and cloud counters. Those measurements describe different parts of the system. Adding them together without reconciling their scope would produce an impressive-looking number with an unclear meaning.

Cached tokens remain within prompt totals. They are not another bucket to add on top of the same work.

History should show its gaps

Resource and AI history can cover one hour, six hours, 24 hours, seven days or 30 days. CPU and GPU average/peak charts use a fixed zero-to-100-percent scale and actual timestamp positions. Missing readings break the line.

The source’s original timestamp survives a failed read. A retained value is marked unverified; refreshing the dashboard does not make yesterday’s measurement new.

For reconnecting engines, the connector reuses SparkDash’s durable ledger and reconnect-aware rates. It does not calculate a competing rate from a cumulative counter that may have reset or resumed.

It remains optional

The existing Mac/Linux collector and probe protocol remain in place. The optional connector uses cached reports, so this addition does not introduce model prompts, benchmarks, settings changes or another engine-polling scheduler.

The same release includes the prepared coverage/repair views and aggregate usage improvements described in the change log. The public package contains application source and examples. Operator inventories, private routes, credentials and live snapshots stay with the installation.

Where to start

• Read the v0.10.0 release notes.
• Read the source and setup guide.
• Inspect the dated change log.
• See how the earlier current-model readout separates identity from health.

The release gives an operator more context in one place. Its usefulness still depends on the configured sources, their timestamps and the task being investigated. No new generation-speed, fleet-coverage or cost-saving result is claimed here.