MAGI

MAGI is a 1995 EVA/MAGI-style multi-agent deliberation terminal. Three bio-computer nodes (MELCHIOR, BALTHASAR, CASPER) each run a distinct personality partition over an OpenAI-compatible API. They independently evaluate a question, cross-review each other's positions across multiple rounds, and resolve by quorum with a full replayable case trace.

Architecture

index.html + style.css          Terminal UI (fixed 1365x768 stage, viewport-scaled)
        |
     app.js                     Client orchestrator: state, rendering, deliberation dispatch
        |
  +-----------+------------------+
  |           |                  |
REPLAY      BYO KEY            LOCAL
  |           |                  |
fixtures.js  model-client.js    deliberation-server.js
             (browser fetch)    (Node.js proxy, port 3001)
                    |                  |
                    +--------+---------+
                             |
                     orchestrator.js     Multi-round convergence loop
                             |
                       protocol.js       Position/action parsing, judgment, convergence analysis
                             |
                       sound.js          Web Audio cue system (vote, accepted, rejected, error)

The three play modes:

• REPLAY: static fixture replay. No key, no backend.
• BYO KEY: browser-memory OpenAI-compatible mode. Bring a disposable low-limit API key.
• LOCAL: local proxy mode for development.

Quick Start

Run the static UI:

npm start

Open:

http://localhost:3000

Click 質詢/ASK+LOG in the bottom-right console to open the interrogation panel, final verdict, and decision log.

Play Modes

REPLAY

REPLAY is the default public demo path. It loads test/fixtures/magi-case-sample.json, runs a five-second MAGI deliberation animation, then shows the stored case result and trace log.

Use this when you want to see the ritual without an API key.

BYO KEY

BYO KEY calls an OpenAI-compatible chat-completions API directly from your browser.

Use only disposable, low-limit keys. The key is held in browser memory only. MAGI does not write it to localStorage, sessionStorage, or the repo.

Provider presets:

Preset	Base URL	Default model
DEEPSEEK	https://api.deepseek.com	deepseek-v4-pro
OPENAI	https://api.openai.com/v1	gpt-4o-mini
OPENROUTER	https://openrouter.ai/api/v1	openai/gpt-4o-mini
CUSTOM	blank	blank

Advanced per-node override JSON:

Leave this blank to run all three MAGI partitions through the same global API key, base URL, and model. Fill it only when you want MELCHIOR, BALTHASAR, and CASPER to use different OpenAI-compatible endpoints or models.

{
  "melchior": { "model": "gpt-4o-mini" },
  "balthasar": { "baseUrl": "https://api.deepseek.com", "model": "deepseek-v4-pro" },
  "casper": { "baseUrl": "https://openrouter.ai/api/v1", "model": "openai/gpt-4o-mini" }
}

Any missing per-node field falls back to the global BYO key/base URL/model.

LOCAL

LOCAL posts to a local proxy at http://localhost:3001/api/deliberate.

Start the proxy in another terminal:

node deliberation-server.js

Optional .env:

DEEPSEEK_API_KEY=sk-...
DEEPSEEK_MODEL=deepseek-v4-pro
OPENAI_COMPATIBLE_BASE_URL=https://api.deepseek.com

MAGI Nodes

The three nodes are not generic assistant personas. They are three incompatible loyalties split from Dr. Naoko Akagi's Personality Transplant OS:

• MELCHIOR-1: Naoko as scientist. Preserves mechanism, proof, instrumentation, and technical necessity.
• BALTHASAR-2: Naoko as mother. Preserves Ritsuko, the child, created life, and what asks to be saved.
• CASPER-3: Naoko as woman. Preserves private desire, attachment, humiliation, jealousy, and the man she cannot release.

Each node returns structured JSON. MAGI preserves dissent and resolves the final case by quorum.

Convergence Protocol

MAGI runs a bounded multi-round convergence loop (max 4 rounds, 12 model calls):

1. Round 1 (Independent): Each node receives the question alone and returns accept, reject, or deliberate with reasoning and confidence.
2. Round 2+ (Cross-Review): Each node reads the previous round's peer brief and returns hold, revise, or no_go with a critique.

Termination conditions

Condition	Result
Any node issues no_go	Immediate stop, verdict = no_go
Unanimous accept or reject	Immediate stop, verdict = accepted/rejected
Two consecutive cross-review rounds with identical positions and all hold	Stop as stable_hold
Position signature repeats (oscillation)	Stop as oscillation, verdict forced to deliberate
Round or model-call budget exhausted	Stop as budget_exhausted

Judgments

• Unanimous: All three nodes agree.
• Majority with dissent: Two agree, one dissents. Dissent is preserved in the trace.
• Deadlock: No majority. Verdict = deliberate.
• Partial: Some nodes errored but a majority still reached quorum.
• Failed: Too few usable responses.

Each completed case includes a termination block with the stop reason, final round count, and model-call budget usage. The ASK+LOG panel shows the verdict above the full round-by-round decision log.

Sound System

MAGI includes a Web Audio cue system (src/deliberation/sound.js) that plays distinct tones for different events:

Cue	Trigger
access	Panel open/close, mode switch
vote	Deliberation in progress (repeating)
accepted	Verdict = accepted
rejected	Verdict = rejected or no_go
error	Provider failure or model error

Sound is disabled by default. Click any interactive element to arm the audio context (browser autoplay policy). The SOUND:ARMED / SOUND:SAFE indicator in the status bar shows the current state.

Case Export / Import

Completed deliberation cases can be exported as JSON and re-imported later:

• Export: Click 搬出/EXPORT in the ASK+LOG panel. Downloads the full case file including all rounds, messages, judgment, and termination metadata.
• Import: Click 搬入/IMPORT and select a previously exported case JSON. The terminal replays the case result and trace log without re-running the model.

Case files follow the protocol defined in src/deliberation/protocol.js and can be validated with validateCase().

Development

Run tests:

npm test

Test suite covers:

• Deliberation protocol parsing (independent + cross-review messages)
• Convergence analysis (all termination conditions)
• Case validation and artifact building
• Provider status polling and routing posture
• Model client empty-response retry logic

Poll provider status and refresh static output:

npm run poll

Syntax checks used during development:

node --check app.js
node --check deliberation-server.js
node --check src\deliberation\model-client.js
node --check src\deliberation\orchestrator.js
node --check src\deliberation\protocol.js

dist/ is ignored by git, but is maintained locally as the static preview output.

Safety Notes

• Do not commit .env.
• Do not use production API keys in browser BYO mode.
• Some providers may block browser requests with CORS. Use LOCAL mode for those providers.
• BYO KEY is for public experimentation, not production secret storage.
• API keys are held in browser memory only. MAGI does not write them to localStorage, sessionStorage, or the repo.

Versioning

See CHANGELOG.md for release history. Current version: 1.0.0.

Technical Documentation

See TECHNICAL.md for protocol design, convergence algorithm, error handling, and architecture decisions.