Projects

bouncer

Python · MIT · FastAPI daemon · per-rule activation predicates (regex + semantic) · the residual layer under deterministic hooks.

Predicate-gated rule injection for long-running LLM sessions. Caps per-prompt injection by trigger_keywords regex + per-rule embedding prototype, so a 30+ rule library stays usable without diluting attention. Rules with deterministic action signatures get promoted out to PreToolUse / Stop hooks — code that runs before any tool call or after any response, no LLM interpretation. bouncer is the residual layer for rules that need conversational judgment. Two-layer enforcement, same shape as ESLint's autofix/warn and OPA's deny/alert — borrowed, because it works. A private companion tool reads agent transcripts and clusters recurring corrections — that's the discovery loop that decides which patterns become rules and which rules become hooks.

UCAI

JavaScript · MIT · Claude Code plugin, marketplace-installable · v2.2 with programmatic phase enforcement.

A Claude Code plugin that solves the same problems as community frameworks (GSD, BMAD, Ralph, Agent OS) — but using Claude Code's native architecture instead of fighting it. v2.2 adds programmatic phase enforcement via a companion contingency engine.

computed-skills

Concept piece · the line between a skill and a program-that-is-a-skill.

A short essay-as-repo arguing that the next interesting shape for Claude Code skills isn't more YAML — it's skills whose prompts are computed at invocation time from the current spec + tool state. Treat the prompt itself as a program's output, not a static string. Picked up small but real traction in the agent-tooling circle.

Spectre

JavaScript · MIT · Claude Code plugin · deterministic spec-driven implementation engine · backs the sdl-vision-engine plugin.

A spec-locked executor: distill a vague vision into a First-Principles spec with action/verification steps, then run the spec deterministically — each step's action gated on its verification, retry once on fail, scratchpad advances. Halts on any unrecoverable failure with full negative knowledge so the next iteration starts from what didn't work, not from scratch. The discipline this whole portfolio claims, packaged as a plugin.

inflate

Go · TUI that sits next to Claude Code · multi-provider (Anthropic, DeepSeek, OpenAI-compatible, Gemini, local Ollama) · live provider switch.

Type a fragment, get a context-loaded prompt for Claude Code on your clipboard. Inflate harvests git status, shell history, your editor's open file, the current Claude Code session and any running dev tools in parallel, then asks an LLM you choose to inflate the fragment into a structured prompt. ? then p cycles between cloud providers and any installed Ollama model with no restart. Native /api/chat for Ollama, think:false and num_ctx opt-in to keep local models honest about truncation. Streams the inflated prompt as it generates.

claude-bridge-server

TypeScript · MIT · MCP server that lets two Claude Code sessions collaborate over a shared workspace.

An MCP server that bridges two Claude Code agents so they can exchange context, hand off tasks, and review each other's work without round-tripping through me. Currently paused (the same collaboration shape is now achievable via subagents in a single session), but the repo is a useful reference for multi-agent MCP plumbing.

Pragma

Python · MIT · Claude Code plugin + CLI · pytest, Vitest, Jest · three layered tiers — AST, coverage, LLM judge.

Watches every test file your AI assistant writes and blocks the edit when the test is gamed. Three tiers, each catching what the previous one misses. Tier 1 — fast deterministic AST classifier (~10 ms): tautologies, mocked-away targets, swallowed exceptions, conditional assertions, monkeypatched fakes, vi.mock / jest.mock on default exports. Tier 2 — runs the test under coverage and asks whether the target's lines actually executed; if not, it isn't a test. Tier 3 — a small LLM (DeepSeek by default, any OpenAI-compatible endpoint, Ollama works) reads test + production code and decides whether the test verifies behaviour or just confidently asserts on its own mocks. Production target is inferred from imports, expected outcome from the test name. Zero config to start.

cc-session-browser

Python · MIT · FastAPI + vanilla JS, no build step · local-first single-machine web UI · Raycast-shaped retrieval over hundreds of Claude Code transcripts.

Reads ~/.claude/projects/<cwd>/<sid>.jsonl directly — search, date-grouped contact-sheet browsing, one-click resume across every project on the machine. Built because claude --resume only shows the most recent few per project and after a few weeks you have hundreds. Real activity sort by transcript mtime (catches autonomous work, not just last user message), staleness analyzer that flags abandoned/old/ deleted-project sessions, two-step archive with restorable sidecar, mobile-friendly over a private mesh. Nothing leaves the machine.

slim-mcp

TypeScript · MIT · npm-installable MCP proxy · 72–77 % token reduction at 100 % accuracy benchmarked against Claude Sonnet 4.

An MCP proxy that gives agents their context window back via schema compression and lazy tool loading. Validated benchmark: 72–77 % token reduction at 100 % task accuracy, measured across 57 tools spanning 4 servers (120 API calls).

zeroshot-detect

Demo · Python · MIT (model: Apache-2.0) · Gradio HF Space (CPU) · OWLv2 200M params, server-side overlay, per-label NMS.

Single-page demo of open-vocabulary object detection. Drop in any image. Type any English noun (or several: "a hat, a dog, a coffee cup"). Get bounding boxes drawn on the image — no class list, no fine-tuning. Server-side overlay so the screenshot you take IS the demo output. Same engineering shape as paperQA: ADRs, CI matrix, mypy strict, offline test suite.

paperQA

Demo · Python · MIT · Gradio HF Space (CPU) + HF Inference API (Qwen 2.5) · demo of citation-faithfulness measurement on a 6-question gold set (faithfulness 1.000, must-cite 0.833).

Single-page demo of PDF question-answering with verifiable citations. Upload one or many PDFs. Ask a question. Get an answer that cites the exact file and page it came from — and a deterministic post-hoc grounding check that drops citations whose claimed numbers don't appear on the cited page.

Headline numbers on the bundled gold set (Attention Is All You Need, 6 questions): citation faithfulness 1.000, must-cite rate 0.833, recall@3 0.833. Each architecture change ships with the measured delta it produced.