CODE THAT CONJURES ITSELF, ORNITH 1.0.
A desktop coding-agent harness (Next.js + Electron, React + TypeScript) for the self-scaffolding ornith:9b model on local Ollama — each session binds its own pluggable Python agent harness that plans tasks and streams native tool-calls: file edits, reads, shell runs. Like Wanda's constructs: the scaffolding appears exactly when the work demands it.
A HARNESS FOR A MODEL THAT BUILDS ITS OWN
Coding agents usually get one hard-coded harness and like it. Ornith 1.0 inverts that: built for the self-scaffolding ornith:9b model running on local Ollama, every session binds its own pluggable Python agent harness — the scaffolding is part of the experiment, not a constant.
The model plans tasks and streams native tool-calls — file edits, reads, shell runs — while the desktop app makes every step visible: an Inspector / Thinking / Critiques workspace surfaces the agent's ReAct reasoning and an L2–L7 critic ledger for every response. Chaos, fully observed.
"The best harness is the one the session conjures for itself."
- FILE SNAPSHOT
- Model — self-scaffolding ornith:9b, local Ollama
- Harnesses — pluggable Python, per session
- Tool-calls — native: edits, reads, shell runs
- Critics — L2–L7 ledger on every response
- Desktop — Next.js + Electron, React + TS
- Tested — full Playwright e2e suite
FROM EMPTY SESSION TO RUNNING AGENT
Bind a Harness
Each session picks from the harness registry (load / unload) — a pluggable Python agent harness that defines how this agent will work.
Plan the Task
The self-scaffolding ornith:9b model plans the work — running entirely on local Ollama, no cloud dependency.
Stream Native Tool-Calls
File edits, reads, and shell runs stream live as native tool-calls — you watch the code change as the agent reasons.
Face the Critics
An L2–L7 critic ledger attaches to every response — layered verdicts on the agent's output before you trust it.
Inspect Everything
The Inspector / Thinking / Critiques workspace lays out ReAct reasoning, tool activity, and critic verdicts side by side — with session and folder organisation keeping long projects sane.
CHAOS MAGIC, FULLY OBSERVED
Self-Scaffolding Model
A model that constructs its own working scaffolding — the harness binds per session instead of being welded to the app.
Harness Registry
Swap agent harnesses like spell books — each session binds the Python harness it needs, and the registry manages the collection.
Native Tool-Calls
No brittle text-parsing — the model emits native tool-calls that stream into the workspace as they execute.
Inspector Workspace
Reasoning, tool activity, and a seven-layer critic ledger per response — the whole session is glass-walled.
Desktop-Grade App
Session and folder organisation for real projects — a coding agent that lives where the code lives.
Playwright E2E Suite
A full end-to-end test suite covers the app — because a tool for judging agents should survive judgment itself.