Investigations, field notes, and occasional opinions.

A re-indexing hash bug in llama-index-core re-embeds byte-identical content on every scheduled run. It has been shipping in defaults for thirteen months. A source-level inspection of the fsspec-maintained ecosystem shows the bug fires today on local filesystem, GCS, SFTP, SMB, and HDFS-via-pyarrow; the same bug sits dormant on S3, Azure Blob, Alibaba OSS, Google Drive, Dropbox, and ten others, masked by a stat-key mismatch that a single upstream commit would lift. Verified end to end against real OpenAI billing on local filesystem and S3, and source-verified across every fsspec backend maintained by the fsspec org.

AI coding agents burn tokens reconstructing codebase structure on every session: grep, read, grep again, piece together what a symbol graph already knows. I built CodeGraph, a local MCP server that parses a repo with tree-sitter and exposes a pre-computed call graph to Claude Code through six tools. I use it daily to keep Claude Code token bills sane. A headless benchmark on a 484-file FastAPI stack measured a 59% drop in tokens, a 60% drop in turns, and 82 seconds less wall time per investigation, with file-level recall held at 100%.

I built ChatGPT LightSession to fix a slowdown in my own long threads. Sixty thousand people have installed it since, and the number keeps climbing. The extension trims the conversation JSON on the way in, keeping long sessions responsive without touching a server.