# Typed Decision Router for Multica

This bundle creates a private **Typed Decision Router** agent on a Multica
instance. It includes an MCP bridge compatible with TypeSafe AI's Jev API
model, the pinned Python dependency, the agent policy, and an importable copy
of TypeSafe AI's open-source skill.

This is an independent community integration. It is not affiliated with,
sponsored by, endorsed by, or an official product of TypeSafe AI or Multica.
Product names are used only to identify compatibility. See [NOTICE](NOTICE).

The TypeSafe API key is deliberately excluded. The installer asks macOS
Keychain to collect it directly, so it does not enter this archive, Multica's
database, an agent environment, or shell history.

## Requirements

- macOS
- Python 3
- Multica CLI authenticated to the destination workspace
- An online Multica Codex runtime that supports `gpt-5.6-sol`
- Network access to `https://api.typesafe.ai`
- A valid TypeSafe API key

`uv` is used when available. Otherwise, the installer uses Python's built-in
virtual environment and `pip`.

## Install

Clone the public repository and run the installer:

```bash
git clone https://github.com/bhzdcz/multica-typed-decision-router.git
cd multica-typed-decision-router
python3 install.py
```

You can also download the `.zip` bundle from the GitHub release, extract it,
enter the extracted directory, and run:

```bash
python3 install.py
```

The installer chooses the first online Codex runtime. To select one explicitly:

```bash
python3 install.py --runtime-id RUNTIME_UUID
```

It then prompts for the API key through macOS Keychain, performs a live Jev API
self-test, imports the bundled `typesafe-ai` skill, registers the local `jev`
MCP server, creates or updates the agent, and assigns the skill and MCP server.

The installer is idempotent. Re-running it updates the MCP executable paths and
the matching agent configuration. An existing Keychain entry is retained unless
you pass `--replace-key`.

## Installed security model

- Keychain account: the logged-in macOS user
- Keychain service: `airoweb.multica.typesafe`
- Local integration: `~/.multica/integrations/jev`
- Agent permission: private
- Agent behavior: shadow mode with human escalation
- MCP behavior: typed Jev evaluation only; no external side effects

To rotate the credential, run:

```bash
python3 install.py --replace-key
```

## Bundle contents

- `install.py`: complete idempotent installer
- `agent-config.json`: portable agent definition and safety policy
- `mcp/server.py`: Keychain-backed Jev MCP bridge
- `requirements.txt`: pinned MCP dependency
- `skills/typesafe-ai.skill`: Multica-importable skill archive
- `skills/typesafe-ai/SKILL.md`: readable vendored skill source
- `manifest.json`: provenance and compatibility information
- `SHA256SUMS`: integrity hashes for the unpacked files
- `NOTICE`: third-party attribution and trademark statement

The vendored TypeSafe skill is MIT licensed and originates from
`https://github.com/typesafe-ai/skills/tree/main/skills/typesafe-ai`, pinned in
the manifest to the exact exported commit.

The Multica integration code in this repository is released under the MIT
License. TypeSafe's vendored skill retains its included upstream license.

## Service terms

The installer requires each user to supply their own API key and calls the API
through the documented endpoint. It does not expose the API as a standalone
service, distribute credentials, imitate the model, or remove proprietary
notices. Users remain responsible for their account's service agreement,
documentation, usage limits, acceptable-use rules, and submitted data.