Best-in-class agent-harness configs, plugins, extensions, and local-model bootstrap — one terrarium-federated monorepo for making terminal agents and the local models behind them worth using.
Four harnesses, one skills standard, one endpoint pattern:
| Package | What | Pinned |
|---|---|---|
packages/harness-pi |
pi.dev (badlogic/pi-mono) — config, models.json local-endpoint presets, TS extensions |
@earendil-works/pi-coding-agent 0.80.3 |
packages/harness-opencode |
opencode — opencode.json presets, plugins, agents |
opencode-ai 1.17.13 |
packages/harness-hermes |
hermes-agent (Nous Research) — cli-config.yaml presets, tools/MCP wiring |
hermes-agent 0.18.0 |
packages/harness-omp |
oh-my-pi (can1357/oh-my-pi, hard fork of pi) — models.yml presets, headless RPC surface |
@oh-my-pi/pi-coding-agent 16.3.12 |
packages/skills |
Shared agentskills.io skills — written once, installed into all four harnesses (+ Claude Code) | — |
packages/shepherd |
shepherd runtime substrate — reversible traces, retained-output review, syscall-jailed agent runs (claude executor) | shepherd-ai 0.2.1 |
packages/local-models |
Hardware probe → tiered model matrix → serving configs (llama.cpp / vLLM / ollama, OpenAI-compatible) | see MODELS.md |
The convergence this repo exploits: all four harnesses consume the same Agent Skills standard, and all four speak to local models through OpenAI-compatible endpoints. Skills are authored once; a served model is wired into every harness with a config preset, not an integration.
./scripts/bootstrap.sh # proto toolchains + moon setup + lefthook
packages/local-models/probe.sh # what can this machine run?
# then follow the README of the harness you wantTerrarium-federated node (external; registered in terrarium docs/REGISTRY.md).
moon 2.2.6 workspace, proto-pinned toolchains, conventional signed commits,
release-please, affected-only CI + gitleaks. Install scripts never pipe remote
shells (curl | bash is banned here — download, pin, verify).