Posts 264 entriesPage 3 of 27

Posts

Technical Deep-Dive

Forge Dogfood Ships a Grade-A Plane Plugin, JRig Loop Closes

First end-to-end /skill-creator --forge dogfood produced a Grade A 97/100 Plane plugin while the JRig-Verified provenance loop closed across schema, build pipeline, page target, homepage surface, and a new validator tier.

Read
Technical Deep-Dive

Guidewire MCP: v0.1.0 → v0.1.1 in 76 minutes

v0.1.0 shipped at 19:14 Mountain, surfaced an install-path defect at 20:30, patched in v0.1.1. The architectural insight: dual Postgres pools for tamper-resistant audit logs.

Read
Technical Deep-Dive

The Two Postgres Bugs the Tests Caught: A Real-DB Integration Test Case Study

A no-mocks testcontainers policy caught two production-fatal Postgres bugs in one test run — PG 15's schema USAGE removal and an asymmetric SELECT grant for a state-machine-driving sink.

Read
Technical Deep-Dive

Guidewire MCP v0.1.0: Carrier-Native Server Blueprint

How v0.1.0 of guidewire-mcp-for-claude shipped six foundation packages, five carrier-vocabulary tools, and a 30k-word blueprint in one day — read against the literature on module decomposition, audit-log integrity, and long-context attention degradation.

Read
Technical Deep-Dive

An Anti-Slop Framework Found Three Bugs Inside Itself on Day One

A 41-gate anti-slop framework shipped v0.1.0, then found three silent failure modes inside its own gates within hours of release.

Read
Technical Deep-Dive

Four production deploy gotchas: systemd, V8 JIT, tsc noEmit, Caddy

A mandy dashboard cutover surfaced four cross-layer deploy gotchas in fifteen minutes: tsc noEmit, systemd MemoryDenyWriteExecute vs V8 JIT, and two Caddy traps.

Read
Development Journey

Eight Deploy Iterations: Tailscale OIDC + Reusable Workflow

8 GitHub Actions iterations to land Tailscale OIDC + a cross-repo reusable workflow. P2 and P5 of VPS-as-the-Home, with the plan rewrite that held the line.

Read
Development Journey

Propagation Day: When the CLAUDE.md Spec Becomes the Migration Plan

Three CLAUDE.md spec entries hit critical mass on the same day across nine repos. The lesson is not the volume — it is that writing the spec first turned propagation from toil into a script run.

Read
AI Engineering

Forty-Four Minutes Before First Pitch: An LLM Fallback Chain and a Live Probability Gauge in One Session

Groq quota died 44 minutes before a live broadcast. The fix was a 4-tier LLM fallback chain ending in a facts-only renderer. Then we shipped JChad's Challenge in the same session.

Read
Architecture

The Rubric Sits On Top Of The Spec: A Schema Validator Postmortem

Postmortem on a schema validator debacle: when a stricter enterprise rubric is built on top of a permissive open spec, the rubric must be additive, never the spec's floor.

Read