SF Bay Area • Git Page • All icons from iconics
Important
A pre-main() constructor was silently stripping LD_ / DYLD_ from Codex tool subprocesses, forcing 11–300× slowdowns through CUDA/MKL fallbacks across every supported OS. OpenAI's specialized debugging team investigated for weeks and walked away without a root cause — the diagnostics couldn't see code that ran before they loaded. I traced the regression to a single commit, built the reproduction harness, and shipped the upstream fix in rust-v0.80.0 with attribution. It was the main blocker to Codex spawning and controlling effective subagents.
Proof: Issue #8945 | PR #8951 | Release notes (rust-v0.80.0)
Full Investigation Details
In October 2025, OpenAI assembled a specialized debugging team to investigate mysterious slowdowns affecting Codex. After a week of intensive investigation: nothing.
The name fits. pre_main_hardening() executed before main() — before profilers attached, before logging initialized, before any instrumentation could see it run. It stripped LD_LIBRARY_PATH and DYLD_LIBRARY_PATH from the process environment, then control returned to main() and the constructor was gone. Users could see the variables set in their shell. Inside codex exec, they were empty.
Within 3 days of their announcement, I had the introducing commit (PR #4521) and a working hypothesis. I sent both to @tibo_openai.
Identification is not proof. The next 2 months were repro harnesses, benchmarks across CUDA / Conda / MKL / HPC stacks, and cross-referencing 15+ scattered user reports until the pattern was undeniable.
| Date | Event |
|---|---|
| Sept 30, 2025 | PR #4521 merges, enabling pre_main_hardening() in release builds |
| Oct 1, 2025 | rust-v0.43.0 ships (first affected release) |
| Oct 6, 2025 | First “painfully slow” regression reports |
| Oct 1–29, 2025 | Spike in env/PATH inheritance issues across platforms |
| Oct 29, 2025 | Emergency PATH fix lands (did not catch root cause) |
| Late Oct 2025 | OpenAI’s specialized team investigates, declares there is no root cause, identifies issue as user behavior change |
| Jan 9, 2026 | My fix merged, credited in release notes |
| Platform | Issues | Failure Mode |
|---|---|---|
| macOS | #6012, #5679, #5339, #6243, #6218 | DYLD_* stripping breaking dynamic linking |
| Linux/WSL2 | #4843, #3891, #6200, #5837, #6263 | LD_LIBRARY_PATH stripping → silent CUDA/MKL degradation |
Compiled evidence packages:
Platform-specific failure modes- Reproduction steps with quantifiable performance regressions (11–300×) and benchmarks
Pattern analysis- Cross-referenced 15+ scattered user reports over 3 months, traced process environment inheritance through
fork/execboundaries
Comprehensive Technical Analysis
Investigation Methodology
The bug was designed to be invisible:
- Pre-main execution
- Used
#[ctor::ctor]to run beforemain(), before any logging or instrumentation - Silent stripping
- No warnings, no errors — just missing environment variables
- Distributed symptoms
- Appeared as unrelated issues across different platforms and configurations
- User attribution
- Everyone assumed they misconfigured something (shell looked fine)
- Wrong search space
- Team was debugging post-
mainapplication code
[!NOTE] Standard debugging tools cannot see pre-main execution. Profilers start at
main(). Log hooks are not initialized yet. The code executes, modifies the environment, and vanishes.
OpenAI confirmed and merged the fix within 24 hours, explicitly crediting the investigation in v0.80.0 release notes:
"Codex CLI subprocesses again inherit env vars like LD_LIBRARY_PATH/DYLD_LIBRARY_PATH to avoid runtime issues. As explained in #8945, failure to pass along these environment variables to subprocesses that expect them (notably GPU-related ones), was causing 10×+ performance regressions! Special thanks to @johnzfitch for the detailed investigation and write-up in #8945."
Restored:
| GPU acceleration | Internal ML/AI dev teams |
| CUDA/PyTorch | ML researchers |
| MKL/NumPy | Scientific computing users |
| Conda environments | Cross-platform compatibility |
| Enterprise drivers | Database connectivity |
When the tools are blind, the system lies. Everyone else has stopped looking.
- claude-cowork-linux⭐410
- The Linux port of Claude Desktop's Cowork mode. Bubblewrap sandbox in place of a VM; the ASAR is unpacked from the host before any sandboxed code runs. Highest-adoption project in the portfolio.
- llmx
- Local-first codebase indexer. BM25 + mdbr-leaf-ir neural embeddings (Burn) fused via Reciprocal Rank Fusion; deterministic chunking; runs in-browser via WebGPU/WASM. Live at llm.cat.
- dota
- Post-quantum secrets manager. v7 TC-HKEM hybrid (ML-KEM-768 + X25519); Argon2id master key; AES-256-GCM encrypted JSON vault. Terminal UI.
- claude-wiki⭐20
- Comprehensive Markdown documentation mirror for Anthropic's Claude, featuring 2000+ articles on APIs, SDKs, agents, and integrations.
- pyghidra-lite⭐34
- Token-efficient MCP server for Ghidra, enabling analysis of ELF, Mach-O, and PE binaries with Swift, Objective-C, and Hermes support.
- raley-bot
- Automated grocery shopping assistant leveraging web API for intelligent product selection, price tracking, and coupon clipping via CLI and MCP.
- indepacer
- A Python CLI tool for querying PACER, downloading federal court dockets and documents, and managing case data efficiently.
- claude-warden⭐60
- Security hooks for Claude Code: blocks SSRF probes, caps subagent spawn budgets, compresses MCP outputs, and exports every tool call to OTEL traces.
- claude-cowork-linux⭐410
- Run the official Claude Desktop app's Cowork mode natively on Linux. Bubblewrap sandbox in place of a VM; the ASAR is unpacked from the host before any sandboxed code runs.
- specHO
- LLM watermark detection via phonetic/semantic analysis (The Echo Rule) — live demo at definitelynot.ai
- codex-patcher
- Automated Rust code patching tool leveraging tree-sitter for syntax-aware modifications and reliable LLM-generated updates.
- htmx-docs
- Curated HTMX documentation in Markdown, including API references, Big Sky repos, and relevant RFCs, organized for easy access.
- filearchy
- Filearchy is a Wayland file manager forked from cosmic-files, enhancing workflows with custom MIME icons, extended archive support, and terminal integration.
- nautilus-plus
- Enhanced Nautilus file manager with sub-millisecond search, large animated thumbnail support, and crash prevention features.
- indepacer
- CLI tool for querying PACER, enabling case searches, docket downloads, and document retrieval from federal court records.
Self-hosting bare metal infrastructure (NixOS) with post-quantum cryptography, authoritative DNS, and containerized services.
- Cosmic Code Cleaner @ definitelynot.ai
- LLM paste sanitizer with vectorhit algorithm — fix curly quotes, invisible Unicode, confusable punctuation, dedent blocks
- LLMX Ingestor @ llm.cat
- WebAssembly codebase indexer — private, deterministic chunking and BM25 search for large folders
- LINTENIUM FIELD @ internetuniverse.org
- Terminal-based ARG experience — interactive mystery with audio visualizations
- Observatory @ look.definitelynot.ai
- WebGPU deepfake detection running 4 ML models in browser
dota — Post-Quantum Secrets Manager
Defense of the Artifacts: A secrets manager engineered for cryptographic longevity. While current encryption remains secure, "harvest now, decrypt later" attacks mean secrets stored today may be vulnerable to quantum computers within their lifetime. dota addresses this with hybrid post-quantum encryption that provides security against both classical and quantum adversaries.
| Layer | Implementation | Why It Matters |
|---|---|---|
| Key Encapsulation | ML-KEM-768 + X25519 hybrid | NIST-standardized lattice crypto with classical fallback — if either is broken, the other protects |
| Key Derivation | Argon2id (memory-hard) | Resists GPU/ASIC brute-force; tunable time/memory parameters |
| Storage | SQLCipher (AES-256-CBC) | Encrypted at rest with authenticated pages; survives partial file corruption |
| Hardware Auth | HMAC-SHA1 challenge-response | YubiKey/SoloKey required for unlock — no master password alone can decrypt |
The TUI (Ratatui) provides vim-style navigation, fuzzy search across entries, secure clipboard integration with auto-clear, and TOTP generation for 2FA codes.
Stack: Rust • pqcrypto (ML-KEM) • x25519-dalek • argon2 • SQLCipher • Ratatui
llmx — Codebase Indexer for Local Agents
Live Demo:llm.cat (WebAssembly — runs entirely in browser, no upload)
Local-first codebase indexing with real neural embeddings (mdbr-leaf-ir) running via WebGPU. No server, no API calls, no data leaving your machine. Hybrid search combines BM25 keyword ranking with vector similarity using RRF for best-of-both-worlds retrieval.
llmx index ~/projects/myapp # Build trigram + BM25 index
llmx search "authentication middleware" --limit 20
llmx export --format md --max-tokens 8000 # Context-window-aware export
llmx serve --port 8080 # Local HTTP API for agents| Capability | Implementation |
|---|---|
| Neural Embeddings | mdbr-leaf-ir vectors with WebGPU acceleration — ~50ms inference, same quality as server-side |
| Hybrid Search | BM25 + vector similarity fused via RRF — handles exact matches and semantic similarity |
| Smart Chunking | Deterministic by file type: functions, headings, JSON keys — same input always yields identical chunks |
| Semantic Exports | Hierarchical outline format (llm.md) with function names and heading breadcrumbs for selective retrieval |
- Proof
- 7,147 files → 31 MB index → 1,625 tokens
- Stack
- Rust • tantivy • tree-sitter • WASM • WebGPU
claude-warden — Security Hooks for Claude Code
A defense-in-depth hook system for Claude Code that addresses token efficiency, security boundaries, and observability. Born from months of production use identifying failure modes in LLM coding agents.
The Problem: Claude Code's default behavior can burn tokens on verbose command output, leak internal network topology via SSRF, spawn unbounded subagents, and produce unobservable execution traces.
| Hook | Threat Model | Mitigation |
|---|---|---|
| quiet-overrides | Token exhaustion from npm install, cargo build, git log | Injects -q/--silent/--quiet flags; caps output at configurable byte limit |
| ssrf-protection | Agent fetching http://169.254.169.254 (cloud metadata) or internal services | Blocks RFC1918/link-local ranges; allowlist for legitimate internal APIs |
| mcp-compression | MCP tool outputs flooding context window | gzip + base64 for large payloads; configurable threshold |
| subagent-budget | Recursive agent spawning exhausting API quota | Per-session spawn limits; depth tracking; cost estimation |
| otel-tracing | Black-box execution; no audit trail | Exports spans to Grafana/Loki with tool calls, durations, token counts |
# Example: warden blocks verbose npm and injects quiet flag
$ claude "install dependencies"# [warden] Intercepted: npm install → npm install --silent# [warden] Output capped at 4096 bytes (was 847KB)Stack: Shell • jq • OpenTelemetry • Prometheus • Grafana/Loki
- claude-wiki ⭐20 — Comprehensive Anthropic documentation wiki — 749+ docs across 24 categories
- observatory — WebGPU deepfake detection with 4 ML models — live: look.definitelynot.ai
- specHO — LLM watermark detection via phonetic/semantic analysis — live: definitelynot.ai
- burn-plugin — Claude Code plugin for the Burn deep learning framework
- raley-bot — Automated grocery assistant with F5 bot detection evasion, unit pricing across bizarre measurements, automatic coupon clipping, and MCP server for Claude Desktop
Primary server: Dedicated bare-metal NixOS host (details available on request)
| Security | Post-quantum SSH • Rosenpass VPN • nftables firewall |
|---|---|
| DNS | Unbound resolver with DNSSEC • ad/tracker blocking |
| Services | FreshRSS • Caddy (HTTPS/HTTP/3) • cPanel/WHM • Podman containers |
| Network | Local 10Gbps • Authoritative BIND9 with RFC 2136 ACME |
