Blog

Every fix to my incident loop after release was found and verified by reading code. When the demo finally ran end to end, it worked, and that changed what I think it's for.

#mlops-incident-commander · part 4/4

The eval harness is the only signal for tuning the incident loop, which only works if ground truth stays hidden. A month after release I found postmortem memory handing it back.

#mlops-incident-commander · part 3/4

Detection in my incident loop has no thresholds at all, only a model's judgement. Remediation ignores the model entirely. Why I split the trust that way, and where it leaks.

#mlops-incident-commander · part 2/4

A LangGraph loop that watches a live classifier, opens incidents, and proposes fixes. Most of the repo is fake on purpose, and I wrote down which parts.

#mlops-incident-commander · part 1/4

On release morning the Linux append path finally ran in CI and passed. Then both packages went to PyPI, and the question the project started with is still open.

#presidio-compliance-stack · part 5/5

Six months with Claude Code, from usage paranoia and a code tsunami to a short addiction, skills, Opus, and seven lessons along the way.

#journal

The NRIC recognizer accepted a range of place-of-birth codes on a recall argument. Reading all 294 rows of the official table showed nine of them don't exist.

#presidio-compliance-stack · part 4/5

Measuring my own audit store turned up a concurrency claim that didn't hold, and a case where a record names a recognizer that found nothing while the chain still verifies.

#presidio-compliance-stack · part 3/5

Building the v1 plans for a Malaysian PII pack and its audit store surfaced four places the plans were wrong — the worst was an audit log that silently lost records on Windows.

#presidio-compliance-stack · part 2/5

Before building a Malaysian PII pack for Presidio, I made the audit layer argue for its place in both directions. It survived, on a narrower reason than the one I started with.

#presidio-compliance-stack · part 1/5

Packaging the gate for PyPI found four real bugs and produced a release workflow that has never run — committed as publish.yml.disabled, because a package on PyPI is a claim I can't make yet.

#pr-compliance-gate · part 6/6

The investigator was broken by an event loop, a missing environment variable, and a structured-output method that quietly stopped guaranteeing anything — none of which a 100%-covered suite could see.

#pr-compliance-gate · part 5/6

Every path the investigator resolved was checked against its sandbox. The sandbox itself was located by an unvalidated id from the POST body — so the guard was fine and the ground under it moved.

#pr-compliance-gate · part 4/6

The investigator is the only part of the gate that goes off and reads a repository by itself, so most of its design is a list of things it cannot do — including fail.

#pr-compliance-gate · part 3/6

Routing in the gate is a pure-Python rule ladder, not an LLM call — but the list reducer underneath it meant an investigation could never actually resolve anything.

#pr-compliance-gate · part 2/6

A portfolio piece, plainly — a LangGraph and Agent SDK pipeline that reviews PRs against policy, built around two cost gates that keep most traffic away from the expensive models.

#pr-compliance-gate · part 1/6

An echo tool grew a callback_url and lint had nothing to say about it — which is why the cassette gate has a second step that only asks whether the server's surface moved.

#mcp-cassette · part 6/6

A cassette is a verbatim transcript of what a third-party server said, committed to your repository and fed to a model — so lint reads it for injection smells, and its rule packs are TOML on purpose.

#mcp-cassette · part 5/6

Adding remote transport meant framing SSE myself to keep the dependency list short — and then a recording fixture hung because the server was talking on a stream nobody was listening to.

#mcp-cassette · part 4/6

The clean POSIX shutdown path had never run anywhere. CI ran it on four machines at once and found two bugs the blunt Windows path had already solved.

#mcp-cassette · part 3/6

Three things about stdio nearly stopped the recording proxy from working at all — and shutting it down cleanly on Windows ended in os._exit(130) after I gave up on the correct design.

#mcp-cassette · part 2/6

I wanted to start a project with MCP and found there was no easy way to test the interaction without calling the agent over and over — so I built the vcrpy equivalent for it.

#mcp-cassette · part 1/6

The smallest tool in my Claude Code monorepo — a wrapper that auto-names every unnamed session with its folder and a timestamp, so the session history is actually browsable.

#claude-code-command-center · part 10/10

file-sync keeps my Claude config in step across two folders, newer wins either way — until a raw copy clobbered a machine-specific statusLine.command and taught me why settings.json needs a merge, not a copy.

#claude-code-command-center · part 9/10

Claude sessions are disposable, so I harvest their lessons into a weekly master file on a schedule — plus a hand-run skill I built for a separate credit pot Anthropic announced but never shipped.

#claude-code-command-center · part 8/10

I'd copied the same CLAUDE.md and configs across a dozen repos and let them drift — Vantage is a single HTML file that surveys every repo, diffs any two files, and copies the good version across.

#claude-code-command-center · part 7/10

Planning got cheap enough to do from my phone, so plans piled up across ten repos — docket is the single board I run them all from, one at a time or batched by project.

#claude-code-command-center · part 6/10

Two skills, same name, no way to tell which one Claude Code fires — so I built a browser that lists every skill, agent, and hook on the machine, shows which one wins, and lets me read what each actually does.

#claude-code-command-center · part 5/10

I already had a usage dashboard; I built the terminal version out of pure "can I" curiosity — and it's what finally justified extracting the shared parsing into a library.

#claude-code-command-center · part 4/10

The Claude Code status line shows the moment; I wanted the story — so I built a dashboard that breaks token usage down by project, tool, and model, with a rate-limit trajectory that tells me when I'll hit the wall.

#claude-code-command-center · part 3/10

My Claude Code status line started as a context gauge back in the 200k days, grew rate limits and a cost figure — and why I trust two of those three and squint at the third.

#claude-code-command-center · part 2/10

Why I pulled a dozen standalone Claude Code tools into one monorepo called C4 — and why that would have been a maintenance trap before Claude Code could read the whole thing.

#claude-code-command-center · part 1/10

I built figureflow to answer two questions — can you wrap React Flow in Python, and can an LLM author interactive diagrams with it. Both got answered. The second answer is why I archived it.

#figureflow

The generic recipe underneath my Claude Code wrapper — a WebSocket-to-PTY bridge into a container, written down with the five things that actually cost me time.

#claude-code-html-wrapper · part 3/3

A self-hosted web front end for Claude Code — one Docker container per user, sessions you can resume from a browser tab, and no cloud in the loop.

#claude-code-html-wrapper · part 1/3

My Claude Code wrapper has no database, no Docker SDK, and no permission prompts — each absence was a decision made in the spec, and each one is why the server fits in seven files.

#claude-code-html-wrapper · part 2/3

The redesign looked AI-generated and I couldn't say why — so I asked the model that designed it to critique it. It listed its own tells, and iteration 19 became a subtraction pass.

#claude-code-plugin-toggler · part 6/6

Reskinning my Claude Code plugin manager around teal and terracotta — a palette that stays related to Claude without looking like an Anthropic product, and a VSCode panel that does the same trick against the editor.

#claude-code-plugin-toggler · part 5/6

Going from a two-tier local/global plugin model to three scopes — Local, Project, User — because the missing one lets a repo carry its own opinions to whoever works in it.

#claude-code-plugin-toggler · part 4/6

Every line of my Claude Code plugin manager was written by Claude Code. What I actually did all those days — specs, sequencing, and being the first user — and where the model out-debugged me.

#claude-code-plugin-toggler · part 3/6

I promised to write up the path problems once I understood them. It took three releases, a 29-line band-aid, and Claude Code finding the one character I couldn't see.

#claude-code-plugin-toggler · part 2/6

Setting up yet another project with my usual Claude Code plugins, I got annoyed enough to build a toggler — a browser UI and a VSCode panel over one JSON file.

#claude-code-plugin-toggler · part 1/6

Resetting the blog. Expect short, occasional notes on building AI coding agents and the tooling around them.