ollama-harness 0.1.0

Self-improving agentic research swarm with local LLM inference

ollama-harness

Local-first agentic research pipeline. An LLM talks to the web, your filesystem, a browser, and itself — search, synthesize, evaluate, revise, remember. No cloud API required.

pip install ollama-harness
oh                        # interactive REPL
oh research "RL from human feedback"
oh /lit-review "RAG reranking" save to review.md
oh /design https://stripe.com save to stripe-design.md

---

What it does

A single oh command drives an agentic loop:

1. Plan — identify what's known, what's missing, what queries to run
2. Research — multi-round web search with novelty gating and URL enrichment
3. Synthesize — produce a structured markdown document from the merged context
4. Evaluate — Wiggum scores output across 6 dimensions (relevance, completeness, depth, groundedness, specificity, structure)
5. Revise — if below threshold, the producer rewrites from evaluator feedback
6. Remember — compress the run, store in ChromaDB, inject relevant observations into future runs

Skills extend the loop with specialised agents: browser navigation, literature review, YouTube transcription, design-system extraction, and multi-file HTML page generation.

---

Install

pip install ollama-harness

Or from source with uv:

git clone https://github.com/upskilled-consulting/ollama-harness
cd ollama-harness
uv sync
uv sync --extra gpu   # CUDA torch
uv pip install -e .   # register the `oh` entry point

Prerequisites

Dependency	Purpose	Notes
Ollama	LLM inference (default)	ollama serve must be running
llama.cpp server	Alternative inference backend	Configure via HARNESS_ENDPOINTS
Node.js ≥ 18	Dashboard UI	start.py builds it automatically; manual: cd dashboard && npm install && npm run build
Playwright	Browser skills	playwright install chromium
whisper.cpp	Audio transcription	Build binary, place at whisper.cpp/

---

Quick start

# Interactive REPL
oh

# One-shot task (no quotes needed)
oh research the latest work on speculative decoding

# Literature review
oh /lit-review "LLM calibration and uncertainty" save to calibration-review.md

# Browser navigation
oh /browser https://arxiv.org "find the most cited paper on RLHF this year"

# Transcribe a YouTube video
oh /transcribe https://youtube.com/watch?v=...

# Extract a design system from a live site
oh /design https://example.com save to design.md

# Generate a themed HTML page from .md content files
oh /build-page design.md from content/ save to index.html

# Full design-extract + page-build in one command
oh /site https://example.com from content/ save to index.html

# Generate a themed .pptx deck from a PDF paper
oh /deck --design https://example.com --content paper.pdf --out slides.pptx

# Deck from a URL content source with an existing design system
oh /deck --design brand.md --content https://example.com/article --out deck.pptx

# Deck from a folder of .md files styled to match a live site
oh /deck --design https://example.com --content ~/notes/ --title "Q2 Review" --out deck.pptx

---

Skills reference

Command	Description
research <topic>	Multi-round web search + synthesis
summarize <url|path>	Fetch and compress a URL or local file
/lit-review <topic>	Fetch papers, annotate, synthesize into review
/annotate <url|path>	Annotate a paper or document (wiggum eval)
/browser <url> <goal>	LLM-guided web navigation + content extraction
/sitemap <url> [goal]	Crawl a domain, rank pages by goal
/design <url>	Extract design system tokens from a live URL
/build-page <design.md> from <dir/>	Generate themed HTML page from .md content files
/site <url> from <dir/>	Design extraction + page build in one command
/deck --design <url|md> --content <url|dir|pdf>	Generate a themed .pptx slide deck
/transcribe <url|path>	Transcribe YouTube video or local audio
/recall <topic>	Surface relevant observations from memory
/introspect	Generate a live capabilities doc from the skill registry
/orientation	Summarise project state + recent activity
/re-orient	Rebuild orientation cache from GitHub state
/suggest	Recommend next research tasks
/debug [filter]	Diagnose recent FAIL/ERROR runs
/email <contact> <goal>	Draft and send emails via Gmail
/sync-wiki	Sync lit-review corpus to GitHub wiki
/panel	Enable 3-persona wiggum review panel

Flags

Flag	Effect
--no-wiggum	Skip quality evaluation loop
--headed	Show browser window (browser/design tasks)
--keep-browser	Leave browser open after task
--reuse-browser	Reconnect to existing Chrome session

---

Configuration

Copy .env.example to .env and edit:

cp .env.example .env

Key variables:

# Model endpoints — llamacpp / vllm / openai-compatible
HARNESS_ENDPOINTS='{"qwen3-8b": {"url": "http://localhost:8082/v1", "model_id": "qwen3-8b", "backend": "llamacpp"}, "qwen3.6-35b": {"url": "http://localhost:8083/v1", "model_id": "Qwen3.6-35B-A3B-UD-IQ3_S.gguf", "backend": "llamacpp"}}'
HARNESS_PRODUCER_MODEL=qwen3.6-35b

# Pure Ollama (default — no HARNESS_ENDPOINTS needed)
# Just run: ollama pull qwen3:8b

# Semantic Scholar API key (optional — increases rate limit)
S2_API_KEY=your_key_here

# Gmail (for /email skill)
SENDER_NAME=Your Name
SENDER_EMAIL=[e-mail]

Multi-endpoint routing

HARNESS_ENDPOINTS maps a short tag to {url, model_id, backend}. Supported backends: llamacpp, vllm, openai. Models not listed fall through to Ollama. This lets you run a fast small model (8B) alongside a large one (35B) on separate ports and route to the right one per task.

---

Deck generation

/deck extracts a design system from any URL (or reads an existing .md design file), loads content from a URL, folder of .md files, or PDF (local or remote), and renders a fully themed .pptx using python-pptx.

oh /deck --design https://stripe.com --content research.pdf --out deck.pptx
oh /deck --design brand.md --content ~/notes/ --title "Q2 Review" --out deck.pptx
oh /deck --design https://notion.so --content https://example.com/paper.pdf

Content sources are auto-detected:

Source	Handling
https://... (web page)	Playwright scrape → structured markdown
https://....pdf	MarkItDown converts directly from URL
/path/to/file.pdf	MarkItDown converts local PDF
/path/to/folder/	All .md / .txt files in directory
/path/to/file.md	Single markdown file

Slide types are inferred from markdown structure: # → title slide, ## → section divider, bullet lists → content slides (auto-split at 6 bullets), > blockquote → callout, markdown tables → table slides.

---

Page generation

/build-page uses a three-pass decomposed strategy that handles any number of content files without context overflow:

1. Analysis — LLM reads title + abstract of every file, clusters by topic, assigns display roles (featured / card / compact)
2. Shell — generates HTML structure (nav, hero, cluster sections) with <!-- SECTION:filename.md --> placeholders
3. Sections — one LLM call per file, role-aware card HTML injected into the shell

Result: a complete, themed, clustered page regardless of how many files are in the content directory.

---

Dashboard

A React/TypeScript UI (Vite + Tanstack Query) provides live visibility into every run.

View	Description
Dashboard	KPI cards (total runs, pass rate, avg score, token spend) + recent activity feed
Runs	Master-detail split: compact run list on the left, full DAG inspector on the right. Click any run to see the pipeline graph, per-stage token counts, output preview, Wiggum scores with dimension bars, evaluator feedback, and an RLHF thumbs-up/down panel per node.
Submit	Fire a task directly from the browser; result appears live in Runs.
Analytics	Time-series charts for run volume, pass rate, and token usage.
Sessions	Group runs by session for multi-turn task tracking.
Artifacts	Browse output files written by runs.
Fine-tune	Training metrics (loss, accuracy curves) and RL dataset browser — preference pairs, reward feedback, GRPO rollouts, and DPO examples with Wiggum evaluator annotations.
MCP	Inspect registered MCP tool servers.

Two floating action buttons in the lower-right corner provide quick access without cluttering the sidebar:

• Terminal — a harness shell with cd navigation, command history (↑/↓), clear/help, and live run-status badges for any submitted task.
• Voice — the voice input panel for hands-free task submission.

---

Starting the full stack

python start.py          # starts inference servers, FastAPI, React dashboard

Or individually:

uvicorn harness.api.main:app --reload    # API server (port 8000)
cd dashboard && npm run dev              # React dashboard (port 5173)

---

whisper.cpp setup

The /transcribe skill uses the whisper.cpp binary for fast CPU/CUDA inference:

git clone https://github.com/ggerganov/whisper.cpp whisper.cpp
cd whisper.cpp
cmake -B build -DGGML_CUDA=ON
cmake --build build --config Release
./build/bin/whisper-cli --download-model base.en

Place the built directory at whisper.cpp/ in the repo root.

---

Development

uv sync --extra dev
pytest tests/
ruff check harness/

---

License

MIT