Dependency Reviews
Working notes on every tool and library in use across the workspace. Updated whenever a dependency is discussed or used in a novel way.
Job applications
- job-ops — in progress — self-hosted job tracker with LLM scoring and resume tailoring; import pipeline works, RxResume integration incomplete (v4/v5 mode mismatch)
Orchestration & durable execution
- Hatchet — not adopted — correct durability model, wrong shape for single-machine workloads (needs Postgres + server)
- Icepick — not adopted — elegant agent/tool API on top of Hatchet, but 9 months stale and inherits full Hatchet infrastructure requirement
- AMD GAIA — not adopted — desktop app platform (GUI, voice, RAG, NPU); no
spawnSync()primitive, Python-only, wrong shape for headless batch pipelines
LLM infrastructure
- axe — recommended — local LLM agent runner; TOML agents, Ollama backend, sub-agent composition
- Ollama — recommended — local inference server; qwen3.5:27b for all agents; sequential-only on single GPU
- Claude — recommended — cloud LLM for agent work; strong at code; weak at fixing the class not the instance
- Agents — patterns for multi-agent orchestration
Financial research agents
- LangAlpha — not adopted (Ollama) — PTC financial research agent; market data layer is excellent, agent layer requires frontier LLM + Daytona sandbox
Podcast pipeline
- Docker — recommended — bespoke container per project; multi-stage build
- KittenTTS + espeak-ng + phonemizer — adequate — local TTS; functional but fragile path plumbing
- Cloudflare Pages, R2, D1, Workers — recommended — static hosting, object storage, edge DB, Workers; zero-ops once configured; many wrangler footguns
- Raindrop.io — recommended — bookmarking as podcast source queue; tag-before-publish is a bug waiting to happen
Runtime & tooling
- Bun — recommended — TypeScript runtime + package manager; replaces Node/npm/tsc/jest
- Zod — recommended — LLM output validation; schema must be kept in sync with agent TOML constraints manually
- OpenTelemetry — recommended — tracing; only reliable record of what happened in pipeline failures
- NanoClaw — agent platform running this workspace
- Claw Platforms — comparative review of NanoClaw alternatives (SwarmClaw, Moltis, nanobot, AlexClaw, PicoClaw)
Content extraction
- Defuddle — recommended — article extractor + Markdown output, Node/Bun compatible
- html-to-markdown — HTML→Markdown converter (Go); needs a separate extractor upstream
- Pydoll — headless Chromium automation (Python); for JS-gated paywalls only
Site generation
- mdsite — static site generator used for reviews site
After-action reports
- Podcast pipeline — 2026-04-04/05 — 6 incidents across 3 container launches