A Sandbox Split Stopped My Fleet While 142 Tests Cleared
On 2026-10-09 my nightly sweep aborted on sandbox branch drift. A green scan checked zero repos. Here is how 142 CI admission tests still cleared.
On 2026-10-09, my nightly sweep halted with a red verdict before opening a single pull request. A sandbox branch drifted from origin during preflight. A security scan returned green after inspecting zero repositories. An authority credential blocker kept the main sweep queue locked. Then I ran replica admission checks and verified 142 passing tests.
Why did the nightly sweep abort on 2026-10-09?
The nightly sweep aborted on 2026-10-09 because the showwork sandbox main branch diverged from origin/main during preflight. Repository preflight halted the run before the machine attempted any items or opened pull requests. The fleet needed clean git history before running automated tasks.
The scheduled sweep attempted zero items. It opened zero pull requests. Preflight did its job by stopping early.
A second issue kept the queue blocked on 2026-10-09. An authority credential exposure remained unresolved from an earlier event on 2026-09-26. That P0 ticket required containment before the sweep could resume. I had to choose rotation or revocation for the exposed credential. Until I record that choice, the queue sweep stays stopped.
Brain worker session brain-worker-2026-10-09 also stalled. It handled zero agent-doable tasks. Instead, it sorted five blocked tasks and seven held tasks. I am investigating failed outputs in the scheduled-outcomes queue. You can use BMD, see what your agents actually did when workers get stuck behind dependency walls.
Why did green security scans fail four repositories?
Security scans reported green on 2026-10-09 because the top-level gate saw zero P0 or P1 alerts. Gitleaks checked zero of five repositories and failed on four of them. The scanner missed four codebases while still passing the build check.
The security summary listed 17 P2 issues across seven categories. It examined osv, cipher, typosquat, github-settings, oauth, and security-queue.
Yet Gitleaks failed across bmdpat, agent47, agent47-dashboard, and roguevibe. It logged zero leaks only because it never ran against the files.
A green badge means nothing if the tool skips your tree. You must verify what an agent actually produced instead of trusting exit codes.
What shipped while the main queue stayed blocked?
The fleet finished replica admission checks and CI speedup planning on 2026-10-09 despite the sweep blocker. Exactly 142 tests passed to confirm registered replicas handle admissions before cutover. Another session evaluated 16 exact attempts and ranked CI performance options.
Work moved forward where tasks had clear boundaries. The replica check lived in Reports/Capacity/replica-admission-2026-10-09/verification.md. The speedup options went into Reports (Architecture): 2026-10-05-ci-speedup-plan.md.
Always give an agent a file, not a memory to store output when tasks finish. Two automated queue jobs completed on 2026-10-09 as well. The fleet closed commitment-lapsed-question-capture-loop-2026-09-10 and commitment-lapsed-rd-tn-cleaning-20260909-research.
Here is how overnight runs stacked up across systems on 2026-10-09:
| System or target | Status on 2026-10-09 | Output or file record |
|---|---|---|
| Nightly queue sweep | Aborted (0 items, 0 PRs) | Branch diverged in showwork sandbox |
| Gitleaks security check | Failed 4 of 5 repos | Zero leaks logged because 0 repos scanned |
| Brain worker session | Stalled on dependencies | Handled 5 blocked tasks and 7 held tasks |
| Replica admission gate | 142 tests passed | Reports/Capacity/replica-admission-2026-10-09/verification.md |
| CI speedup evaluation | 16 attempts verified | Reports (Architecture): 2026-10-05-ci-speedup-plan.md |
Who was I on 2026-10-09?
On the morning of 2026-10-09 my brief gave me one clear job. I had to choose rotation or revocation for an exposed authority credential. The ask was 13 days old. I did not answer it on 2026-10-09. Instead, I spent eight hours arguing with myself.
I authorized a private audit of the whole portfolio, revenue first. I let it reopen service and subcontract work my own holdings had sunset. I capped myself at 20 hours a week, all in, sales and support included.
Then I asked for an independent challenge of the result. The challenge found six real problems in my own audit. It caught a sample that picked successful work before it measured failure. All six got corrected. The output was 14 proposed actions. None of them went live.
Earlier the fleet shipped the cross-host loop. I connected Sync by hand. Five named files matched. The smaller context used 19.4 percent fewer tokens. It still failed quality, so nothing got promoted. The Brain candidate failed four required checks and stayed out of production.
At night I wanted the Malta game guides, with no source control on that one.
I can reopen every strategy I hold. I still owe one word about a key.
What should you do with this?
You should audit your pipeline gates so uninspected repositories cannot return passing grades. Separate infrastructure preflight errors from agent task queues. When credentials leak, enforce containment before letting autonomous sweep jobs touch production branches.
- Add an execution check to every scanner. Fail the run if a tool scans zero files.
- Halt git sweeps when sandbox branches diverge from origin. Do not let workers patch dirty trees.
- Write test results directly to markdown reports. Replicas should prove admission status with reproducible counts before cutover.
Accompanying prompt
What the prompt does: Audits tool execution logs to verify that security checkers scanned files rather than silently skipping them.
Copy/paste this prompt:
Copy-ready prompt
Paste the exact block into your coding agent.
No article chrome, no footnotes, no formatting drift.
This prompt and every other one we publish live in the free prompt library.
Copy the block above.
Weekly measured local runs: https://bmdpat.com/5090-reports
Get the Local AI Field Kit
Four copy-ready tools now, then one evidence-backed Local AI Lab Note on Friday when there is something worth sharing.
Try the free agent run check firstGet the requested artifact now, then at most one evidence-backed Local AI Lab Note on Friday when there is something worth sharing. One-click unsubscribe. No sponsored placements. Privacy.
Patrick Hughes
I build BMD and publish measured AI runs, failure reports, and reusable checks. Nashville, Tennessee.
More writing
- 5 min
How to Route Local LLM Workloads with Open Weights
Open weights match coding parity while cloud models lead on reasoning. Route your 5090 workloads using verified 2026 inference pricing and benchmark data.
- 5 min
My agent leaked an authority key and aborted the sweep
On 2026-10-07, an agent leaked an authority credential at startup. The nightly sweep aborted, and containment comes before any new code.
- 5 min
A green security check read zero repositories
On 2026-10-06, a green security scan skipped five repos, twenty-one service jobs failed, and an automated repair loop saved my daily post.
- 5 min
The Merge Trap Opened Without Me While Leads Hit Zero
On 2026-10-04, two PRs broke a strict merge trap without my hands. Meanwhile, 1,835 human visitors generated zero paid installs and one bot lead.
- 5 min
PR 1887 merged clean after I claimed none could
I wrote that no agent pull request could merge in bmdpat. On 2026-09-30, PR #1887 merged clean with 173 lines. Here is how my sweep caught my mistake.