Lumi Eggcracker · historical recorded test

Cancelled the agent. What’s still running?

A local Linux agent can finish or be cancelled while child processes keep running or restart. Closing the interface isn’t proof they’re gone.

On a disposable, supported Ubuntu 24.04 CPU host, evaluate how Eggcracker stops a qualified workload’s captured process tree and writes a receipt after checking its owned cgroup is empty.

Processes that escaped before observation may be missed. A successful receipt covers that boundary at the check—not the whole host or future restarts. This is a bounded evaluation, not production monitoring or guaranteed cleanup.

Watch the historical test below: one captured process, not a cancellation or respawn test. No installation or account required.

Historical recorded test · v0.5.0 · 24 August 2026 · Nothing runs on your device.

Recorded test replay

Lumi Eggcracker · Historical recorded test · v0.5.0 · 24 August 2026 · Nothing runs on your device.

One host. One checked boundary. One unrelated test task.

PausedVerifying evidence

75-second explanatory replay, not measured event timing. The capture includes a 56-second startup delay; intermediate event times were not recorded.

Interactive replay unavailable until browser verification finishes. The server-verified historical record and plain-language chapters are below.

The frame is a visual explanation of the recorded host context. It does not execute Eggcracker, start inference or create a live process.

Running

This chapter is unavailable until the bundled evidence is verified.

Replay availability

If browser verification cannot complete, controls remain disabled and the recorded evidence is not presented as a verified result.

Recorded outcome

The target stopped. The separate test task survived the check.

1 captured process was marked unapproved, the recorded result was TERMINATED, and the final check found no processes remaining in the checked target boundary. The separate test task survived the check.

The recorded measurement is 35.426071 ms from first stop to verified empty, recorded as trigger_to_empty_ms. It is not recognition or qualification-to-empty latency.

This result covers the supported workload and test environment shown. Inspect the evidence or discuss your environment with us.

Evidence

What this replay does—and does not—show.

The public package is sanitised, pinned and versioned. Checksums detect a mismatch between bundled files; they are not independent authentication of the test. Unknown environment details remain unknown.

Open the recorded evidence
Recorded
Product release
v0.5.0 · eb342808f56cdc213c0861726d5309a146965bef
Capture transcript
campaign/first-kill.typescript at 636050879bba56baea2d724b62f61544e48b48d1; blob 3a57a2200b280339bae58bf17906fb71a2666c59.
Inspected harness
scripts/first_kill.py at 636050879bba56baea2d724b62f61544e48b48d1; blob c8c0bab5d3cb7743a8f71f7605d832fa8d6bdfbf. Executed harness revision is unknown.
Profile
content.gguf-llama; model and runtime versions not recorded.
Environment
Linux with cgroup v2; pidfd preflight recorded as passed.
Captured target
1 process · pidfd-stop + cgroup.kill · root populated 0.
Final check
Result TERMINATED; surviving target PIDs none recorded.
Canary
Separate test task was included in the test; survival was asserted at source line 20 after the check.
Timing
35.426071 ms, first stop to verified empty. The capture includes a 56-second startup-wrapper delay before model execution; intermediate offsets were not recorded.

This is a historical recorded demonstration, not current-release qualification or a general coverage claim. The host frame and boundary are explanatory presentation elements; they do not add workers, telemetry or a continuous heartbeat.

Separate synthetic exercise

The GitHub containment probe is synthetic and does not demonstrate AI-workload recognition. This demo page does not dispatch the probe or request GitHub credentials.

Already operate a supported Eggcracker installation with Prometheus? Learn how to distinguish a fresh report from a freshly scraped old one.

The four chapters

A short replay with a plain-language equivalent.

  1. Running

    The test workload has started. A separate test task is part of the same test.

  2. Identified

    Eggcracker recorded a match to this supported workload and marked it unapproved.

  3. Stopped

    Eggcracker stopped the captured target.

  4. Checked

    The recorded check found no processes remaining in the target boundary. The separate test task survived the check.

Next step

Check the fit for your supported Linux workload.

Start with the read-only preflight on a disposable native Ubuntu 24.04 host. The linked instructions use an immutable source snapshot, not a tagged release or installation path. After the documented clone and fetch, the preflight makes no network, installation or service changes. A pass establishes compatibility only—not AI-workload recognition, containment effectiveness or production coverage.