Building thoughtful software with AI.Notes from the systems behind the work.
Answers from my AI ordering arm ran back through my rule-based planner, so the planner's caps and gates erased nearly every disagreement; I pulled it out into its own pipeline and gave the model the sales history it had been missing.
Read moreReading 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.
Reading chat history out of an AI code editor, I hit a timing number that is either a real date or a count since the app opened, with nothing to say which.
A tool server for an AI assistant was committed whose queries all filter by organization, but a failed token falls through to an environment default that stays on unless one exact mode string is set, and none of the deploy recipes set it.
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.
The local coding agent got tools to save its attention cache to system RAM and restore it by name. Rereading the diff, restore only flips a sticky switch, every reply recopies a frozen cache, a specialist adapter can answer on the manager's cache, and nothing tests that recall works.
A checkpoint gate allowed to stop fine-tuning runs turns out, on a read-back of the diff, to pass on crashed metrics, missing data and a zero-on-error helper, and its reasoning probe mostly grades answer length.
Pasted screenshots now get pulled out of agent transcripts and into file storage before they hit my analytics database, but a single oversized image sends every picture in the session back inline and leaves the uploaded ones orphaned.
A fallback let my local models use tools without native tool calling, and a shared helper roots every path, but the registry carries no rules, so several tools build their own paths and skip the guard entirely.
A fake fix button became a real pull-request pipeline, and then I found the guardrails it generates fall back to a one-word placeholder, while the template tests cheer for that placeholder and fail a real detector.
Coding-agent runs moved into the background, with their logs streamed to object storage under a comment promising crash safety. Reading the diff back, the upload only becomes a file when my code closes it, so a killed run leaves no log and a row stuck on running.
The repo agent had been leaving its cloud sandbox running after every job. The commit that turned the delete back on also let the agent skip every permission prompt, and what holds up in it is narrow staging and a finally block. Precedence between two modes, unchecked steps and killed runs are still loose.
A committed run log shows my guidance-writing agent producing reports about its own work while the pipeline logged nothing changed. Reading the diff, most of the decisions that matter take the agent's word: model-picked confidence, an unreadable answer scored as perfect, and file paths scraped from logs instead of the diff.
The first join between agent hook events and the transcript needed a tag inside the conversation. The fingerprint matcher that replaced it the same day leaves the prompt alone, but a start and its end now fight over one record.
A checkpoint added an MCP endpoint to my observability tool that validates the request envelope, then runs any registered tool with whatever repository ID the caller typed. Canned data is the only reason nothing real has leaked.
The local-model edit pipeline has a solid shape: snapshots, bounded retrieval, hash-checked diffs, staging, quarantine and an audit log. Reading the diff back, the config full of safety rules is never loaded, and apply runs any post-step command a proposal carries.
Rereading the commits that gave the little chat tool an online mode showed that the key check decides nothing: a failed key still gets saved over a good one, the online switch only asks whether any key exists, and the periodic recheck can only ever say yes.