Building thoughtful software with AI.Notes from the systems behind the work.
Restarting a stopped machine is one command; a trustworthy auto-healer is defined by its refusals: deliberate stops, locks, storage errors, retry budgets, and a rehearsal mode, all published as metrics.
Read moreA brand-scoping preview for my ordering system splits 'couldn't look' from 'found nothing' and counts the database before and after, but every test so far hands it matching tallies.
A zero in my ordering plan can mean no demand or missing data, so I added a check that compares the plan against a raw summary of the store records, blocks the deploy only on a same-category contradiction, and warns on everything softer.
Widening a discount-and-return sync from one store to all of them meant the event ID had to carry the store too, while counters added the same day show what the merge absorbs and what the size cap cuts.
The session summarizer found its work by looking for sessions with no record, so anything that failed came back on every run. Now a failure writes a stub record, the report skips stubs and counts them, and the promise in the warning got smaller to match what the code can actually do.
Reading my coding-agent grader line by line, I found it stores friendly constants and stock turn values whenever the model judge fails, under the same label as real verdicts, so averages and the pass gate can't tell guesses from judgments; I want every score to carry its origin and every report to show coverage.
The latency check in my agent eval harness read a timing nobody recorded, treated the blank as zero, and passed every run while its unit test stayed green.
Fresh branches were tagging sessions with the base branch's commit, so one error could match all of them. A merge-base check now drops that commit on the laptop, but the server's code-host lookup and backfill fill it straight back in for any pushed branch.
Making the error-to-session lookup forgiving and giving it a short cache in the same pass meant one failed call to the error tracker could be stored and served for minutes as a confident no.
Rereading my own dashboard rebuild, I found a panel labeled with percentiles that averaged per-window means, plus a handful of smaller labels that promised more than their queries: substring metric picks, ignored temporality, a token total that widened, and seed data the two lanes can't agree on.
Every tool in a burst claimed the full cost of the one model call that followed it; I moved ownership to the call, then split its cost by tool output size, and the tests still don't check the math.
A page I added scans stored agent conversations for leaked keys and personal data, and reading the diff showed five places where it drops rows, rules, matches, list entries or failed scans without a word in the response.
The issues page I built matches stack frames to AI coding sessions, but the path remap runs only after the fix button, a loose substring fallback does the real matching, and the before-the-error scorer I wrote sits unused while the list sorts newest-first.
A missing pair of brackets in an AND/OR rule sent every plain accept and reject on my agent dashboard to the automatic pile, while a second copy of the same labels only knew one spelling of a person.
Each AI assistant tool call logs a decision and a result under one event name, so a success rate divided by row count shows a perfect tool at half. At least one tool query was right by luck, I added a result-only filter to all of them, and the team charts from the same change filter on a column their view may not have.
An agent-facing tool in my observability service binds the tenant properly but hand-quotes an optional list of session IDs into the query text, so one model-chosen ID can turn the tenant check into one side of an OR.
Promoting tenant ids out of a telemetry map into indexed columns went from computed-from-the-row, to a collector copy step that kept the collector from starting and likely left history empty, and back. What I can and can't verify about which version is live.
Two kinds of agent event shared one normalized name, and a success flag turned into a digit on its way into a text-only store. I made every query name the event kind it wants, wrote the flag as a word, and moved percentages into the queries. No tests yet.
A few of my ClickHouse rollups kept an average next to the total, and the summing engine would have added those averages on every merge; now they keep total and count and divide at read time, and the affected rollup keys now include every label they store.
The agent console's transcript store and its telemetry store got merged by timestamp. The duplicate check asks whether anything is nearby, not whether the matching thing is, so real tool results can vanish in bursts, and the join loosened the telemetry access filter along the way.
A rewrite of my editor extension's telemetry added a bonus line to every change event, so a single letter now counts as a line added, and the same commit deleted the counters that could have caught it.
The day I taught my assistant-telemetry installer a second assistant, it switched on prompt logging without asking, merged settings in two opposite directions, wrote a token in plain text, and the same commit removed the server cap on stored conversation content. An editor extension from the same day sent full file paths on every edit.