⚠️ DEPRECATED & CONSOLIDATED

Warning

This repository has been consolidated into the unified agent-skills repository. All active development, updates, and maintenance have moved to https://github.com/rodritimecoder/agent-skills. Please update your harness mounts, remotes, and dependencies accordingly.

---

workstation-clis

High-performance local CLI utilities, autonomous agent execution engines, and multimodal inference pipelines designed for dual-GPU AI workstations.

---

🏛️ The Architecture: How The Whole Process Works

This repository houses the runtime tools and execution engines powering an autonomous agent workstation. Rather than relying on cloud APIs, bloated context windows, or corporate framework ceremony, the system is engineered around four structural pillars:

flowchart TD
    subgraph L1["1. Constitutional Glue (GEMINI.md / AGENTS.md)"]
        G1["Upstream-First Authority"]
        G2["Mandatory Grounded Verification Protocol<br/>(Launch → Doctor → Drive → Evidence → Cleanup)"]
        G3["No-Comments Discipline & TypeScript Invariant"]
    end

    subgraph L2["2. Execution & Review Engine (review-hub)"]
        D1["review-hub review ponytail (Minimal Diff & YAGNI)"]
        D2["review-hub review poteto (State Matrix & Domain Modeling)"]
        D3["review-hub review thermonuclear (1000-Line Ceiling & Structural Judo)"]
        D4["review-hub review verification ([REQ-XX] Traceability Maps)"]
        D5["review-hub review unslop (31 AI Tell Filters)"]
    end

    subgraph L3["3. Workspace Pipeline (Interpretable Context Methodology - ICM)"]
        I1["Root AGENTS.md (Routing & Negative Scope Target &lt;60 lines)"]
        I2["Stage CONTEXT.md (Explicit Input / Process / Output Contracts)"]
        I3["Filesystem as State Machine (COMPLETE vs STALE via Timestamps)"]
    end

    subgraph L4["4. Grounded Memory & Continuity (graph-memory)"]
        M1["Local BGE-Small-en-v1.5 384d Embeddings (:8082)"]
        M2["SQLite FTS5 BM25 Keyword Search"]
        M3["Reciprocal Rank Fusion (K=60)"]
        M4["Verbatim Exact-Quote LLM Distillation (:8000 Qwen3.6-35B)"]
    end

    subgraph HW["Local Hardware Substrate (Dual AMD Radeon RX 9060 XT - 32GB VRAM)"]
        P1[":8000 llama-swap / :4000 LiteLLM"]
        P2[":8080 Qwen3.6-35B ROCMFP4"]
        P3[":8081 audio.cpp Nemotron ASR"]
        P4[":8082 BGE-Small-v1.5 Embeddings"]
        P5[":8085 LightOnOCR-3-1B Flagship"]
        P6[":8083 vllm-d3-mini Decision"]
    end

    HW --> L1
    L1 --> L2
    L2 --> L3
    L3 --> L4

Loading

---

🔬 The Grounded Verification Protocol ("Prove It Works")

The primary machine invariant across all CLI tasks is the 5-step Grounded Verification Loop:

1. Launch: Start the real local service, daemon, or CLI environment from a clean state.
2. Doctor: Run a read-only health check ("is this instance worth driving?") before attempting user flows.
3. Drive: Exercise the real user journey programmatically (via PTY, curl, or containerized CDP browser).
4. Evidence: Capture observable proof (exit codes, terminal transcripts, response bodies, screenshots) that survives process teardown.
5. Cleanup: Terminate child processes cleanly without stranding ports or orphan background workers.

Upgraded Engineering Disciplines

• Pre-Implementation State Transition Matrix: Before writing multi-step async or UI logic, agents must tabulate Current State × Event → Next State + Rollback (enforced via poteto and typescript kitchens).
• Numbered Requirement Traceability ([REQ-XX]): Every acceptance criterion carries a [REQ-XX] tag linked directly to verification feature maps (features/<feature>.md) and PR planning blocks in check-plan.mjs.

---

⚡ Local Hardware & Inference Substrate

All models run 100% locally on dual AMD Radeon RX 9060 XT GPUs via ROCm and Vulkan backends:

Port	Model / Service	Role & Performance
:8000	llama-swap	Model hot-swapper with dynamic concurrency sets ($\le$ 32GB resident)
:4000	LiteLLM Proxy	OpenAI-compatible proxy with dynamic VRAM balancing and failover
:8080	Qwen3.6-35B-A3B Uncensored Genesis	Primary Workstation LLM (22 GB ROCMFP4, dual-GPU -ts 45/55)
:8081	audio.cpp (Nemotron 3.5 + Kokoro)	Streaming speech-to-text (40 languages) & fast text-to-speech
:8082	bge-small-en-v1.5 (F16 GGUF)	Dedicated 384-dimensional symmetric vector embeddings
:8083	vllm-d3-mini (vision)	The decision engine: /v1/systemone bool/choice/score probabilities + image gating
:8084	CrispEmbed Suite	Ultra-compact scene-text detection & line recognition (DBNet + PARSeq)
:8085	LightOnOCR-3-1B	#1 Flagship Video & Slide OCR (~16k t/s prefill, ~330 t/s gen)
:8088	DeskForge-Qwen3.5-4B	Visual GUI grounding & computer-use coordinate extraction

---

📦 Curated CLI Catalog

This suite contains 15 focused engineering, sandbox, and inference CLIs:

Tool Directory	Identifier	Core Capabilities
worktree-harness	worktree-harness / swarm-dispatch	Autonomous multi-agent Git worktree orchestration with swarm-dispatch 1-command parallel compiler, pre-write section claims, deterministic test quality gates, and automated merge conflict arbitration.
openrouter-free	openrouter-free	Multimodal & subagent CLI with compound bash batching, verified free models (Nemotron-3 Ultra/Lightning, Gemini Flash Lite, Mercury Decide), calibrated decisions, and TTS speech synthesis.
review-hub	review-hub	Unified code review and architecture audit gateway: Ponytail (minimal diff & YAGNI), pstack (23 playbooks), Poteto Mode, Thermonuclear, Unslop, TypeScript, No-Comments, and Grounded Verification.
github-architect	github-architect / git-sync	Autonomous fleet synchronizer (sync-all, status-all), disciplined 6-stage branch sync engine (git-sync), multi-machine bootstrapper, and pre-push secret/blob security guard.
subagent-spaces	subagent-spaces	CPU-pinned rootless Podman sandboxes partitioning 16 CPU threads into dedicated worker slots with zero host blast radius.
(retired) omnicli	—	Retired 2026-10-20; archived at archived/omnicli/. Each local endpoint is now owned by a skill: local-llm, ocr, jev-decision, audiocpp.
graph-memory	graph-memory	Local hybrid RAG (BM25 + BGE-small 384d) with exact-quote LLM distillation, prompt recall hooks, and session digests.
litellm-vram-switcher	vram-rig-ctl	Multi-GPU VRAM switcher, llama-swap orchestrator (:8000), and LiteLLM failover manager (:4000).
tailscale-ops	tailscale-ops	Tailscale mesh networking, homelab node discovery, HTTPS serving, and peer file transfer.
hybrid-browser	hybrid-browser	Tiered browser automation agent using Podman Chromium, CDP, and DeskForge-Qwen3.5-4B visual grounding.
video-miner	video-miner	6-stage ICM video mining orchestrator combining Nemotron ASR, vllm-d3-mini visual gating, and LightOnOCR code extraction.
agentic-engineering	agentic	Autonomous 7-step engineering pipeline: target-local Kanban (.kanban/), Event Listener, Fresh-Context Judge, Splitter, and multi-model sandboxes.
markitdown-cli	markitdown-cli	Universal document converter to Markdown (PDF, DOCX, XLSX, PPTX, CSV, HTML, Audio).

---

🚀 Tool Highlights

1. review-hub — On-Demand Code Review & Architecture Audit Hub

Enforce elite engineering rubrics on any codebase without bloating prompt context:

# View available review suites
review-hub list

# Run specific audit or review pass
review-hub review ponytail [target]           # Minimal diff & YAGNI audit
review-hub review poteto [target]             # Domain modeling & State Transition Matrix
review-hub review thermonuclear [target]      # 1,000-line ceiling & structural judo
review-hub review verification [target]       # Grounded Verification harness & [REQ-XX] map
review-hub review unslop [target]             # 31-rule AI tell scanner

# Run all core review suites sequentially
review-hub all [target]

2. graph-memory — Local Hybrid RAG & Exact-Quote Distillation

Zero-dependency local memory system combining keyword and semantic meaning:

# Search notes or docs using hybrid reciprocal rank fusion (K=60)
python3 graph-memory/rag/search.py "conductance matrix in nodal analysis"

# Rebuild vector index via local bge-small endpoint (:8082)
python3 graph-memory/rag/build_index.py --corpus "/path/to/notes"

# Distill notes into verified question-quote triplets via Qwen3.6-35B (:8000)
python3 graph-memory/rag/distil.py "/path/to/notes" --limit 10

3. Local inference endpoints — one skill each

There is no gateway CLI. Each port is owned by a skill; read the skill's SKILL.md and run its script:

# LLM coder on :8080 (local-llm) — see skills repo hardware/local-llm/
python3 hardware/local-llm/scripts/llm_client.py "Write an async rate limiter" --stream

# OCR PDFs, slides and screenshots with LightOnOCR-3-1B on :8085
ocr lecture.pdf --pages 1-5 -o notes.md

# Calibrated decision gating with vllm-d3-mini on :8083 (add --image for vision)
jev bool --context "..." --question "Is this safe to run?"

# Speech-to-text with Nemotron-3.5 on :8081
audiocpp asr transcribe meeting.mp3 --out meeting.md

4. hybrid-browser — Vision-Grounded Browser Agent

Automates complex web interactions using Chrome DevTools Protocol (CDP) inside rootless Podman containers, with DeskForge-Qwen3.5-4B visual fallback for Canvas/SVG elements:

python3 hybrid-browser/scripts/hybrid_browser.py "https://news.ycombinator.com" --action "read_headlines"

---

🔗 Related Repositories

• academic-agent-skills: Academic pipeline repository containing Canvas LMS synchronization (canvas-cli), cloud Zoom recording extraction (zoom-recorder), Cornell active-recall notes (cornell-notes), and Open Knowledge Format bundles (okf-architect).
• my-agent-skills: Canonical upstream private repository synchronizing 25+ agent skills across local harnesses (sync-skills).