Field journal
Blog archive
Every published note, with the newest work first. Use the topic links to stay inside one search intent.
On 2026-10-07, an agent leaked an authority credential at startup. The nightly sweep aborted, and containment comes before any new code.
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.
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.
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.
On 2026-09-29, a strict check rule held 26 clean pull requests in queue. Here is how my sweep caught the block, corrected itself, and found the fix.
On 2026-09-28, a pending secret rotation expired, 1 suite file failed out of 555, and commit 000b3235f fixed discovery so showwork 0.6.5 could ship to PyPI.
On 2026-09-27, I fixed my traffic counter after bot rows inflated human sessions. I also shipped showwork 0.6.5 on PyPI with problem-first copy.
On 2026-09-26, three checks passed with clean exits while hiding broken work: empty scans, crawler traffic, and leaking tests. Here is what to audit.
On 2026-09-25, a false green report buried a real fix. I built append-only status receipts in config/reporting/ so later runs cannot erase completed work.
Compare Ternary Bonsai 27B and Gemma 4 E4B for local AI. See VRAM footprints, multimodal support, benchmark scores, and run commands for your setup.
On September 23, I added write-time checks to my agent claim recorder. Six bad rows needed retraction. Here is what those checks can and cannot prove.
On 2026-09-19 I merged ten PRs into BMD desktop while a dead disk faked four green signals across the fleet. Here is what broke and what shipped.
On 2026-09-21, I removed the October 1st kill date for BMD. I also shipped bmd-desktop 3.47.15 and fixed a clipboard bug in bmdpat.
NVIDIA's Personal AI Router spreads inference requests across the machines on one network. It does not pool VRAM, and its scheduler counts queued jobs.
Compare Q4_K_M and Q8_0 for Llama 3.1 8B. Learn how quantization affects file size, VRAM usage, perplexity, and generation speed on local hardware.
I wrote agent roadmaps, but my September 18 records showed a separate problem. Plans, delivered work, and saved time each need their own evidence.
I repaired a Windows agent launcher on September 17. Shell commands and two tests passed. That still did not prove the whole nightly queue could finish.
I moved four recurring jobs to Gemini, then checked what actually finished. One recorded run proved progress. The broken queue sweep stayed broken.
In the week of 2026-08-31 the fleet merged 143 product pull requests. Stripe still read zero. Throughput is the tell, not the win.
My local blog writer reads a daily log. I remove private lines before drafting, then check the output again. Here is what those checks can miss.
An empty Ollama result can mean no matches or a broken request. I test the response parser so local extraction failures cannot pass as clean results.
I wrote a kill rule for a paid-path test. Zero orders landed. The card expired into an archive and the dashboard still said ACTIVE.
A local model can load and still run out of memory at longer context. Compare two controlled loads, inspect cache logs, and test the real workload.
A tensor-split flag is only a request. Pin the split, watch every GPU, and save one repeatable llama.cpp receipt before trusting the result.