stockyard's foundation lands, lace gains skills and todos, llm-proxy v0.1.0

prime-radiant-inc/stockyard

A microVM-orchestration daemon — ZFS-backed, Flintlock-driven Firecracker VM management exposed over a gRPC API.

The biggest repo of the week, built from scratch in 329 commits. The foundation went in end to end: a Go project with a Cobra CLI and config package, a secrets-provider interface with a 1Password backend, a ZFS management package, and a daemon with SQLite state management. On top of that came the VM layer — a gRPC API defined in protobuf with server and client packages, a Flintlock client wrapper for Firecracker VM management, a cloud-init generator, VM networking, a base VM image and init system, and a stockyard-snapshot vsock client for talking to VMs. Sole contributor: Jesse Vincent.

obra/lace

An agent runtime — the harness that runs coding-agent sessions.

The headline additions are agent-facing. A skills system landed end to end — a SKILL.md parser with frontmatter extraction, a SkillRegistry with directory-precedence resolution, a use_skill tool, and an available-skills section injected into the system prompt. Task-tracking tools (a read/write todo list with trinary status) arrived beside it, and thinking blocks were wired across the Anthropic, OpenAI, and Gemini providers, forwarded to the client and rendered inline in the TUI with delta throttling. Anthropic prompt caching was enabled, and maxTurns jumped to 10,000 with loop-detection reminders.

The TUI got a UX overhaul: colored message borders replacing text labels, a bottom-anchored permission bar with Y/S/N/D shortcuts, expandable inline tool calls with a full-screen details overlay, and a file_edit diff view rendered right in the permission bar. Under the hood, delegate gained resume-for-conversation (working even for completed jobs), a JobManager consolidated shell and delegate jobs behind one interface, and session metadata writes moved to atomic rename semantics. Sole contributor: Jesse Vincent.

prime-radiant-inc/llm-proxy

A Go reverse proxy that sits in front of LLM APIs and logs the traffic.

New this week, and shipped a v0.1.0. A Go HTTP server with TOML/env config fronts a reverse proxy that handles both non-streaming and SSE-streaming responses, with API-key obfuscation and JSONL logging that captures per-request timing. A SQLite session database tracks conversations by message fingerprint with fork detection, and session-ID extraction works across both Anthropic and OpenAI (including ChatGPT OAuth backend routing). It installs as a daemon — systemd unit, Homebrew formula, and shell-rc patching — and ships a built-in log explorer (--explore): a session list with message counts and time ranges, request/response detail pairs, full-text search across logs, and conversation rendering with thinking blocks and tool calls. Sole contributor: Jesse Vincent.