@felan-ai/ext-fusion

Runs the same explicit prompt against 2–8 authenticated, session-allowed chat models in parallel, automatically compares their answers, and lets the user review each answer before explicitly requesting a synthesis. It is provider neutral and uses the configured Core Models service; OpenRouter models work through their normal configured provider. It does not call OpenRouter's beta Fusion server tool.

Configuration and model selection

The default entry uses pi.ui forms, model-list views, progress and paged review documents. It does not require a TUI, client-specific hint or injected Fusion host. /fusion-models searches names and provider/model references, pages through the catalogue, toggles participants and selects the comparison model. Saving is explicit; cancelling leaves the previous lineup unchanged.

Configured extensionConfig.fusion values provide the initial lineup. Picker choices are saved in the current session branch as felan.fusion-lineup.v1 and restored when reopening commands or reloading the extension. Saving also awaits pi.updateConfig({ participants, fusionModel }) to persist global settings through the host-owned writer. If that writer is unavailable or rejects the update, the session copy is retained, an explicit error is shown, and the picker remains open for retry or cancellation. No global save is silently assumed.

{
  "extensionConfig": {
    "fusion": {
      "participants": ["openrouter/anthropic/claude-sonnet-4", "openai/gpt-5"],
      "fusionModel": "openrouter/anthropic/claude-sonnet-4",
      "concurrency": 4,
      "timeoutSeconds": 120,
      "maxOutputChars": 12000,
      "thinking": "off"
    }
  }
}

participants defaults to []; a run never silently chooses models. Configure 2–8 distinct provider/model-id references. The picker displays only authenticated models in the active session's allowed scope. The fusion model must also be authenticated and in that scope. Nested IDs such as openrouter/anthropic/claude-sonnet-4 are supported.

Commands and review

• /fusion <prompt> asks every participant the same prompt in a fresh, tool-free model request, then asks the selected fusion model for a comparison of agreements, disagreements, unique contributions, and uncertainty.
• /fusion reopens the latest unfinished review on the active session branch. After synthesis, it asks for a new prompt and starts a new run.
• /fusion-models searches and updates the participant and comparison model lineup; canceling leaves the saved lineup unchanged.
• Review full originals and comparison through document paging. Synthesis is offered only after comparison succeeds, with optional instructions and an explicit available provider/model override. Separate actions retry failed participants, comparison or synthesis. Closing/reopening does not repeat requests. A separate progress element can cancel an in-flight stage even while its review action is awaiting completion.

With N participants, comparison uses N+1 requests. Choosing Fuse makes one additional model request; retries make the requests they explicitly repeat. The confirmation form discloses this count before each new comparison run. Review documents show provider-reported usage and estimated cost when supplied, plus latency; these estimates are not provider invoices and missing pricing is not free usage.

Fusion sends only the /fusion prompt and bounded source answers—not the session transcript, project files, or the active model's tools. The prompt and bounded answers/comparison are saved as a custom entry in the current session branch for review and recovery. Model outputs and the comparison are untrusted evidence; agreement is not verification. The fused result is added to the transcript only after explicit user action. Provider cancellation is best effort and a request already received by a provider may still incur charges.

The comparison stage never writes files or automatically fuses. Execution remains with the extension and configured Models provider, not the UI client. Cancellation and shutdown join outstanding provider calls, including providers that return after their abort signal fires. Late results cannot publish a successful review or transcript message.

The catalogue limit is 10000 models, with 40 visible rows per page. References must fit 300 UTF-8 bytes. Prompts are limited to 20000 characters; form input also obeys the protocol's 32-KiB field limit. Review documents use 24000-byte pages; the existing 240000-byte saved-review limit remains. Unsupported data is rejected rather than silently omitted. Model-output bounding still follows maxOutputChars and is explicitly disclosed.

Host API

The package root's default export is the service-backed execution entry. It declares optional Core MODELS_SERVICE (felan.models@1). Without a provider, both commands remain registered but warn that Models is unavailable; no inference runs. UI unavailability or a cancelled/deferred confirmation does not approve a request. The root entry retains FUSION_CONFIG metadata for configuration discovery.

import { MODELS_SERVICE } from '@felan-ai/agent-core';

services.set(MODELS_SERVICE, modelsService);

Inference uses ModelsService.getAvailable and completeSimple, with fresh availability checks before dispatch. There is no ambient credential discovery or direct HTTP client. The pure runFusion, compareFusionAnswers and synthesizeFusion APIs remain unchanged.

Hosts supply CreateAgentCoreSessionOptions.updateExtensionConfig(packageName, patch) for global settings persistence. Core binds the package identity and validates writable declared fields; the host owns policy and durable storage. The extension's resolved pi.config snapshot changes on reload, not during the persistence callback.

The named createFusionExtension(host?) factory accepts an explicit FusionHost. Configured hosts should use the service-backed default for generic UI across clients. The package exports no TUI entrypoint.

pnpm --filter @felan-ai/ext-fusion build
pnpm --filter @felan-ai/ext-fusion type-check
pnpm --filter @felan-ai/ext-fusion test

The new npm package requires a maintainer's manual 0.0.0 bootstrap and trusted-publisher setup before its intended CI release; see docs/maintainers/releasing.md. Local verification does not publish the package.