B5 Fusion for pi

Multi-model aggregation ensemble for pi coding agent. Every turn runs all enabled candidates in parallel; a single aggregator fuses their drafts into one best answer.

Not a router. Not a tier gate. All enabled candidates run on every turn. The aggregator sees every draft (reasoning + text) and synthesizes.

       ┌── candidate 1 (deepseek-v4-flash, thinking: minimal) ──┐
       ├── candidate 2 (MiniMax-M3,        thinking: low)     ──┤
turn ──├── candidate 3 (deepseek-v4-pro,   thinking: high)    ──├─→  aggregator
       ├── candidate 4 (qwen3.7-plus,      thinking: medium)  ──┤   (deepseek-v4-flash)
       └── candidate 5 (GLM-5.1,           thinking: medium)  ──┘

Inspired by Anthropic's Claude Fable 5 research direction (mid-tier models fused at modest budget can match one frontier model).

---

Highlights

• True aggregation — every candidate runs every turn, no role/tier routing
• Self-contained config — each candidate provides its own baseUrl + apiKey + model
• OpenAI-compatible — works with OpenRouter, vLLM, LM Studio, Anthropic-via-proxy, custom gateways
• Tool calling — the aggregator can call all of pi's tools (read, bash, edit, MCP, …)
• 7 languages — UI is fully translated: zh-CN, en-US, de-DE, ru-RU, ko-KR, ja-JP, sa-Deva (Sanskrit/IAS)
• Smart abort — once the quorum of candidates succeeds, slow ones are aborted to save latency
• Reasoning forwarding — <THINKING> blocks from each candidate flow to the aggregator

Install

Requires pi ≥ v0.80 and Node ≥ 20.

From npm (when published):

pi install npm:b5-fusion

From this repo (clone + local install):

git clone https://github.com/<your-org>/b5-fusion-pi
pi install ./b5-fusion-pi

Or try without installing (ephemeral, current run only):

pi -e ./b5-fusion-pi

The extension will register a virtual provider. After install, switch to it with /model b5-ensemble/b5-fusion.

Looking for the user-facing config sample? See config.example.json. The runtime settings file lives at ~/.pi/b5-fusion-settings.json.

---

Quick start

1. Install (one of the methods above).

2. Edit config with at least one candidate and one aggregator:

   // ~/.pi/b5-fusion-settings.json
   {
     "version": 6,
     "enabled": true,
     "language": "auto",          // or "zh-CN" / "de-DE" / …
     "candidates": [
       { "label": "fast",    "enabled": true,
         "baseUrl": "https://api.openrouter.ai/v1", "apiKey": "$OPENROUTER_API_KEY",
         "model": "deepseek/deepseek-chat", "thinking": "minimal" },
       { "label": "code",    "enabled": true,
         "baseUrl": "https://api.openrouter.ai/v1", "apiKey": "$OPENROUTER_API_KEY",
         "model": "anthropic/claude-3.5-sonnet",   "thinking": "low" },
       { "label": "reason",  "enabled": true,
         "baseUrl": "https://api.openrouter.ai/v1", "apiKey": "$OPENROUTER_API_KEY",
         "model": "deepseek/deepseek-reasoner",    "thinking": "high" }
     ],
     "aggregator": {
       "baseUrl": "https://api.openrouter.ai/v1",
       "apiKey": "$OPENROUTER_API_KEY",
       "model":  "anthropic/claude-3.5-sonnet",
       "thinking": "low"
     },
     "minSuccessfulProposers": 2,
     "quorumGraceSeconds": 10,
     "proposerTimeoutSeconds": 1200,
     "aggregatorTimeoutSeconds": 1800,
     "draftMaxChars": 8000,
     "shuffleCandidates": true
   }
   

   $OPENROUTER_API_KEY is read from the environment — or paste the key directly.

3. Reload pi (or press /reload).

4. Pick the model with /model b5-ensemble/b5-fusion.

5. Chat — the status bar shows live candidates, then the panel shows results.

---

Commands

Command	Description
/b5-status	Show current lineup + aggregator + UI language
/b5-reload	Reload config from disk
/b5-lang [locale]	Show / switch UI language. auto clears the override

Examples:

/b5-lang                  # list supported locales
/b5-lang zh-CN            # switch to Simplified Chinese
/b5-lang auto             # follow the OS locale again

Supported locales: zh-CN 简体中文 · en-US English · de-DE Deutsch · ru-RU Русский · ko-KR 한국어 · ja-JP 日本語 · sa-Deva संस्कृतम्.

---

How it works

User prompt
   │
   ▼  (every turn)
┌────────── proposer fan-out ──────────┐
│ candidate 1 ─────────► draft         │
│ candidate 2 ─────────► draft         │
│ ...                                   │
│ candidate N ─────────► draft + 🧠    │
└────────── (when quorum met) ──────────┘
   │
   ▼
┌── aggregator (single model, tool-capable) ──┐
│  sees: original prompt + every draft         │
│       + every thinking/reasoning block       │
│  produces: one fused answer (may call tools) │
└──────────────────────────────────────────────┘
   │
   ▼
User

Latency optimisation: as soon as minSuccessfulProposers drafts arrive, B5 Fusion waits quorumGraceSeconds (default 10s) then aborts the still-running slow candidates. The aggregator starts immediately, so wall time ≈ slowest fast candidate, not slowest candidate.

Quorum failure: if fewer than minSuccessfulProposers succeed, the aggregator is told to run anyway but warned that draft references are reduced.

---

Config reference

See docs/configuration.md for the full schema, default values, and per-field semantics.

Highlights:

Field	Default	Meaning
enabled	false	Master switch
language	"auto"	"auto" follows OS locale; otherwise a supported locale code
candidates[]	[]	Enabled candidates run on every turn
aggregator	(empty)	Single fusion model — should support tool calling
minSuccessfulProposers	2	Required quorum for fusion to proceed
quorumGraceSeconds	30	Wait this long after quorum, then abort slow candidates
proposerTimeoutSeconds	300	Hard per-request timeout for proposers
aggregatorTimeoutSeconds	480	Hard per-request timeout for aggregator
draftMaxChars	8000	Per-candidate truncation sent to aggregator
shuffleCandidates	true	Randomize order to remove position bias

API keys can be inline strings or $ENV_VAR references resolved from process.env.

---

Architecture

extensions/b5-fusion/
├── index.ts                # Extension entry: provider + commands + prompt hook
├── config.ts               # Settings types, persistence, validation
├── ensemble-engine.ts      # Proposer fan-out, quorum, aggregator streaming
├── openai-client.ts        # OpenAI-compatible HTTP + SSE streaming
├── aggregator-prompt.ts    # Multilingual user-message builder
├── b5-fusion-prompt.ts     # Multilingual system prompt dispatcher
├── i18n.ts                 # Translation core (detect + override + fallback)
└── locales/
    ├── zh-CN.ts            # 简体中文
    ├── en-US.ts            # English (reference)
    ├── de-DE.ts            # Deutsch
    ├── ru-RU.ts            # Русский
    ├── ko-KR.ts            # 한국어
    ├── ja-JP.ts            # 日本語
    └── sa-Deva.ts          # संस्कृतम् (IAST transliteration)

See docs/architecture.md for the per-module contract.

---

Limitations

• Each enabled: true candidate must be reachable. There is no in-built retry across endpoints.
• The aggregator decides whether to use tools. Proposers are reference-draft only and never get tool definitions.
• Reasoning budget. Different vendors expose reasoning differently. The extension transparently forwards reasoning_content (DeepSeek-style) and <THINKING> blocks; other vendor-specific formats will land as plain text.
• The interface is an OpenAI-compatible /v1/chat/completions endpoint. Anthropic-Native / Gemini-Native is not auto-translated; route through OpenRouter (or any compatible proxy) to use those models.

---

Development

# Install bun (used for the locale smoke test)
curl -fsSL https://bun.sh/install | bash

# Run all checks
bun run test

# Validate key coverage across locales
bun run test:locales

# Validate the npm package shape
npm run test:package

---

License

MIT © 2026 B5 Fusion Contributors.