interactive
REPL with pipes, completion, and a session tree.
TRAAVIIS is the harness engineer. Give it a goal and it forges the harness for that job — routing each prompt to the right model tier, orchestrating sub-agents across the [&] Protocol stack, composing their results into one pipe, and stamping every action with a governance verdict. Pick a goal below and watch it build.
The orchestration above runs scripted, locally — no model is called from this page. In the CLI, traaviis routes each step to your configured provider and MODEL_TIER (local_small ▸ local_large ▸ cloud_frontier); the tools it drives never call models themselves.
$ npm install -g traaviis
· or npx traaviis · traaviis.com
What the orchestrator hands each sub-agent is a command line, and what
comes back is a structured value — not text. So results compose
with | exactly like a shell: generic combinators
(where, map,
each, count)
work with any command without re-parsing. Build a pipeline and watch it
stream — this is the real engine.
There is no plugin API beyond "register a command." Drop a
.mjs in ~/.traaviis/extensions/ and it joins
the same registry the built-ins use — instantly available to you, to a
pipe, and to the orchestrator's sub-agents. Declare a
capability and the built-ins it
needs, and it composes with everything else.
Register the example and run it:
// ~/.traaviis/extensions/deploy.mjs export default function (h) { h.command({ name: 'deploy', capability: 'ops.deploy', // resolvable by capability needs: ['build'], // composes with a built-in async run(h, [name], input) { await h.invoke('build', [name]); // reuse another command return { deployed: name }; // pipe-able output }, }); }
Sub-agents, providers, and governance backends are reached by capability, never by hard wiring. Pick a provider for each capability below and watch the harness re-resolve — and flag any unmet need. /plugins reports exactly this at runtime.
REPL with pipes, completion, and a session tree.
traaviis status — run once, exit.
traaviis --json products — structured for CI & agents.
traaviis --rpc — JSONL in, events out: drive it from an editor or another agent.
Before the orchestrator lets a sub-agent act, the action is piped through the box-and-box kernel, which returns a certificate: floor-then-gradient over an un-weakenable safety floor, feasible ▸ permitted ▸ best. The verdict is a value you can pipe, attach to the session tree, and gate CI on — so autonomy never outruns governance.
The engineer is small on purpose: a command registry, a pipe engine, capability manifests, a session tree, output modes — plus the orchestration loop that drives them. Everything else is an extension. Here's what's deliberately left as a seam, not baked in.
The orchestrator's fan-out is visible — every sub-agent prints in front of you.
The harness routes prompts; the commands it drives never call a model.
Governance is a verdict + certificate — see above.
Bridge to MCP with an extension when you actually want one.
Plans are session-tree nodes you export and replay.
Zero runtime deps, readable ESM. Read it. Fork it. Own it.