Engineering notes
Failures, tests, mechanisms, and fixes
Detailed accounts of production behavior: the symptom, negative results, the test that isolated the cause, and the configuration or code that fixed it.
- How routing a LangGraph RAG chat with JEV got me 98% accuracyJEV-first routing scored 0.981 against 0.957 on 322 labeled messages. How JEV routes, reranks evidence, and grades answers in a LangGraph RAG chat.
- What I learned fine-tuning a small model: mistakes you can avoidWhat I learned directing coding agents through TabCrew fine-tuning: the task I set was too narrow, how I corrected it, and what the evaluation showed.
- Fix Claude Code’s “not supported in WSL” Chrome integration on a Windows hostClaude Code in WSL refused Chrome integration in its settings panel. The browser tools use an account-keyed WebSocket bridge, not the local native-host socket, so a WSL session drives the Windows Chrome with a startup flag and one pairing step. This note traces the relay that was never used.
- Fix “No Codex IAB backends were discovered” on WindowsCodex reported “No Codex IAB backends were discovered” while its in-app browser worked. A runtime comparison traced the failure to a legacy plugin.
- Fix Codex browser startup errors on WSL and WindowsA Codex browser task failed on WSL with Exec format error, broken task paths, and Windows error 267. The repairs that restored Chrome access.
- Fix Codex’s missing browser service on WSL and WindowsChrome connected, but Codex could not load browser-service.mjs. How a Windows/WSL marketplace conflict blocked plugin installation and how it was repaired.
- Building a strict stateless MCP 2026-07-28 server for OpenRouterHow an OpenRouter server moved to strict MCP 2026-07-28 request-response semantics, a shared CLI operations layer, bounded API calls, and explicit credential scopes.
- Building a Verifiable Exchange: Signed Orders, Replayable Execution, and On-Chain AnchorsHow a toy Rust exchange grew into a study of signed orders, deterministic replay, independent validation, Merkle proofs, on-chain commitments, and what those mechanisms still cannot prove.
- Why WhatsApp Desktop Kept Hanging: A WebView2 GPU Failure on Windows 11WhatsApp Desktop hung on every startup. Cache, browser profile, package integrity, and overlays were each ruled out; the cause was WebView2's GPU rendering path. One environment variable fixed it.
- Building a LangGraph portfolio agent that has to show its workHow the robot on th3nolo.com answers from approved public sources, checks its citations, and says when the evidence runs out.
- Fix Codex’s “sandboxCwd is not a local file URI” on WSLA verified repair for Codex’s “sandboxCwd is not a local file URI” error on WSL, followed by CODEX_CLI_PATH and Windows plugin import fixes.
- Serving GLM-5.2-504B on RTX PRO 6000: the vLLM sparse attention fix on SM120A missing vLLM layer pattern caused repetitive GLM-5.2 output on SM120. This note explains the fix and the configuration that restored inference with 240K-token context.
How routing a LangGraph RAG chat with JEV got me 98% accuracy
LangGraphJEVOpenRouterEvaluationNext.jsRAG
JEV-first routing scored 0.981 against 0.957 on 322 labeled messages. How JEV routes, reranks evidence, and grades answers in a LangGraph RAG chat.
On this page
- 01 The new homepage
- 02 Why the chat needed a better classifier
- 03 What JEV is
- 04 How the LangGraph agent works
- 05 One question through the graph
- 06 Evidence the answer must see
- 07 JEV reorders the evidence
- 08 Testing with the real JEV
- 09 Filling the contact form from the chat
- 10 The CV easter egg
- 11 What it costs
- 12 The email stopped, and no check told me
- 13 Sources and credits