role-model

role-model

role-model is an open protocol for capability-aware AI routing, plus a reference router that implements that protocol.

It gives a router a durable contract for describing what a request needs, what an endpoint can do, what policy allows, and why a final routing decision was made.

What role-model does

Every AI workload eventually faces the same question: which model should handle this request? The answer depends on task type, required capabilities, cost, latency, and whether the model is running locally or in the cloud. role-model makes that decision explicit, explainable, and portable.

At a high level, role-model separates AI routing into a few stable pieces:

1. Requests describe task type, required capabilities, modalities, tool needs, and constraints.
2. Endpoint identities and profiles describe concrete routable endpoints rather than abstract model names.
3. Routing policy applies hard denies, preferences, budgets, and tie-break rules.
4. Observability artifacts record the decision, trace, usage, and observed performance.

The reference router supports hybrid routing across three deployment shapes:

• Local / Local - route between local models (e.g., llama-swap peers) based on role, task, and capability
• Local / Cloud - route between a local model and a cloud provider based on cost, latency, and task difficulty
• Cloud / Cloud - route between cloud providers (e.g., OpenAI, DeepSeek, Moonshot) based on capability and economics

This means a single runtime can serve a quick chat request from a fast local model, route a complex coding task to a capable cloud model, and fall back to a cheaper cloud endpoint when the primary is degraded, all under one explainable routing contract.

Install the runtime

For end users, prefer the packaged standalone runtime over a source build.

macOS and Linux

curl -fsSL https://raw.githubusercontent.com/try-works/role-model/main/scripts/install.sh | sh

The installer downloads the latest GitHub Release archive, installs it under ~/.local/share/role-model/<version>/<target>/, and exposes a role-model launcher in ~/.local/bin.

Windows

irm https://raw.githubusercontent.com/try-works/role-model/main/scripts/install.ps1 | iex

The installer downloads the latest GitHub Release archive, installs it under %LOCALAPPDATA%\Programs\role-model\<version>\<target>\, and creates a role-model.cmd launcher.

Manual downloads

If you do not want to use installer scripts, download the matching archive from GitHub Releases.

Platform	Archive	Launch
Windows	role-model-win32-x64.zip	role-model.bat or role-model.exe
macOS x64	role-model-darwin-x64.tar.gz	role-model
macOS arm64	role-model-darwin-arm64.tar.gz	role-model
Linux x64	role-model-linux-x64.tar.gz	role-model

Test a release candidate

Stage builds are published separately as GitHub prereleases named stage-rc-<stage-sha>. They run as role-model-stage on http://127.0.0.1:3457 and use isolated stage state, so they can be tested beside the stable runtime on port 3456.

Download the candidate archive and SHA256SUMS.txt from its prerelease page, verify the checksum, extract it, and run the stage launcher. Prereleases are never selected by the normal installer. A candidate is promoted to main and a stable vMAJOR.MINOR.PATCH release only after a maintainer explicitly records that the exact package was installed and tested.

Update an installed runtime

Updates are currently manual. Stop the running runtime, back up its persistent state, and then re-run the installer or extract the newer release archive. Installer-based updates keep each application version in a versioned directory and repoint the launcher; they do not remove the persistent runtime state.

On Windows, production state is stored under %LOCALAPPDATA%\role-model-runtime. This includes the Message Graph and its encryption and scoped-digest keys under standalone-runtime\track-b\managed-keys. Runtime updates reuse these keys and do not rotate them, so the Message Graph remains readable after an update.

Do not delete, replace, or copy the Message Graph without both original key files. If either key is missing or invalid, the runtime fails closed instead of generating a replacement that would make existing graph data unreadable. After updating, start the new runtime against the same state directory and confirm the Message Graph opens before removing the old application version.

See Install the router for the complete update and backup guidance.

Installation for Pi

The pi-role-model package connects Pi to an externally running role-model runtime.

Start the role-model runtime first, then install the public Pi package:

pi install npm:@try-works/pi-role-model

For local checkout testing from this repository, install the package directly:

pi install ./packages/pi-role-model

Inside Pi, run:

/role-model setup
/role-model status
/role-model doctor
/role-model alias list
/role-model alias choose
/role-model alias use <alias>
/role-model requests
/role-model explain latest

Use those slash commands only from an interactive Pi session. pi -p "/role-model status" is unsupported because Pi print mode does not currently invoke extension commands.

By default the package connects to http://127.0.0.1:3456 and registers role-model as the role-model provider using /api/role-model/downstream/openai. Set ROLE_MODEL_ENDPOINT before starting Pi to use a different local runtime. Remote endpoints require explicit trusted allowRemote behavior, and runtimes that report authentication.required fail closed unless a future supported token source is configured. For local development installs and the full command reference, see packages/pi-role-model/README.md.

For explicit provider prompts, use the provider-relative role-model alias that Pi lists for provider role-model, for example:

pi --no-session --provider role-model --model baseline.remote-only -p "<prompt>"

baseline.remote-only is the canonical provider-relative form. role-model/<alias> is compatibility-only for Pi surfaces that explicitly require a qualified id. Raw HTTP curl calls to the runtime are debug-only fallback tools, not the primary supported Pi workflow.

Installation for DeepSeek Harness

The @try-works/dsh-role-model bundle connects the DeepSeek Harness to an externally running role-model runtime, with the same capability as the Pi package: every ordinary conversation request carries role_model.intent metadata, so the runtime routes on the role, task, capabilities, modalities and tool classes the request needs.

Build the package, then install the bundle into a harness profile with the plugin manager, passing the absolute package directory:

corepack pnpm --filter @try-works/dsh-role-model run build

plugin_manager { action: "install_bundle", target: "<repo>/packages/dsh-role-model" }

The runtime's aliases, models and endpoints then appear in the harness's main model selector under the role-model group, with the recommended alias first. Inside the harness, run:

/role-model status
/role-model doctor
/role-model alias list
/role-model alias recommended
/role-model alias use <alias>
/role-model requests
/role-model explain latest

By default the bundle connects to http://127.0.0.1:3456; set endpoint in its profile patch row to use another channel, for example the stage runtime on 3457. Remote endpoints require explicit allowRemote, and a runtime reporting authentication.required fails closed. The plugin never starts, stops, installs or updates the runtime, and never reads provider secrets.

A profile-launched harness also needs hostLlmModule pointing at the harness's @deepseek-ai/dsh-llm entry: the plugin must use the harness's own LlmAdapter and LlmError classes, because a second class identity would collapse every failure code to the unroutable UNKNOWN. See packages/dsh-role-model/README.md for the full configuration reference, the reason the route needs its own adapter, and the development and verification commands.

Develop from source

Prerequisites

• Node.js 24 (required for node:sqlite and SEA support)
• pnpm 10.x (via corepack enable)
• Go 1.24+ (for llama-swap vendor binary and Windows launcher)

corepack enable
corepack pnpm install

Smoke test

corepack pnpm run smoke

For a fuller walkthrough, see docs/public/quickstart.md.

Development build

Run the bridge and UI in development mode (separate processes):

# Terminal 1: bridge server
cd role-model-router/apps/runtime-host-bridge
corepack pnpm exec tsx scripts/start-for-qa.ts

# Terminal 2: UI dev server
cd role-model-router/apps/runtime-ui
corepack pnpm exec react-router dev --port 5173 --host 127.0.0.1

Then open http://127.0.0.1:5173 in your browser.

Production build (all platforms)

Build the UI and package the SEA runtime:

# Build UI static files
corepack pnpm --filter @role-model-router/runtime-ui run build

# Package the bridge as a single executable
corepack pnpm run runtime:package-sea

Output: role-model-router/dist/release/<platform-arch>/role-model-dev by default. Set ROLE_MODEL_BUILD_CHANNEL=production for role-model or ROLE_MODEL_BUILD_CHANNEL=stage for role-model-stage.

Windows desktop launcher

Build a complete Windows package with dedicated browser window:

# 1. Build UI
corepack pnpm --filter @role-model-router/runtime-ui run build

# 2. Package bridge SEA runtime
corepack pnpm run runtime:package-sea

# 3. Build Go launcher
cd role-model-router/apps/launcher
go build -o ../../dist/release/win32-x64/role-model-launcher.exe .

# 4. Bundle UI files
cp -r ../runtime-ui/build/client ../../dist/release/win32-x64/

Then double-click role-model-launcher.exe in dist/release/win32-x64/. It will:

• Start the bridge server on port 3456
• Open Microsoft Edge in app mode (dedicated window)
• Serve the UI directly from the bridge (no separate dev server)

Documentation

Read this	If you want
docs/public/README.md	the docs hub
docs/public/introduction.md	what role-model is and why it exists
docs/public/quickstart.md	a real end-to-end smoke run
docs/public/concepts/how-role-model-works.md	the system flow
docs/public/concepts/protocol-overview.md	the protocol surface
docs/public/concepts/routing-overview.md	how routing decisions happen
protocol/README.md	canonical schemas and fixtures
role-model-router/README.md	reference router packages and runtime apps
docs/protocol/routing-policy.md	routing policy reference
docs/protocol/taxonomy-v1.md	taxonomy V1 groups, roles, tasks, and Pi classification
docs/protocol/roles.md	role metadata reference
docs/protocol/tasks.md	task metadata reference