Add SwimMeet — multi-stage AI workflow client with extended providers, anonymization, and register pdf_form_filler in MCP

[Copilot]

Implements SwimMeet (apps/mae-client/), a standalone competitive AI document drafting client supporting multi-stage workflow orchestration across any combination of AI models. Also fixes a gap where pdf_form_filler — used by all three Forecast modules — was never exposed via MCP.

New: apps/mae-client/ — SwimMeet

Lightweight two-tier app (backend port 5003, frontend port 5174) for multi-model, multi-stage document drafting.

Backend (Express 5 + TypeScript):

• version-control.ts — in-memory document/version store (with stage metadata)
• routes/documents.ts — document CRUD REST API
• routes/generate.ts — proxies to Cyrano's /api/mae/write/multi for simple parallel generation
• routes/workflow.ts — POST /api/workflow/run + GET /api/workflow/presets for multi-stage orchestration
• workflow-engine.ts — core workflow executor (see Workflow Engine section)

Frontend (React 19 + Vite + Tailwind + Radix UI):

• DocumentEditor — split-panel: prompt + anonymize toggle + WorkflowSelector + ModelSelector left; WorkflowStagePanel + VersionPanel right
• WorkflowSelector — collapsible 2-col card grid for picking workflow type; synthesizer dropdown + expert personas textarea for workflows that need them
• WorkflowStagePanel — vertical timeline of completed workflow stages, each collapsible, with output cards and provider/persona badges
• ModelSelector — checkbox groups across OpenAI, Anthropic, Perplexity, Gemini, Groq, Cohere, DeepSeek, Local/Open, and a free-entry Custom Model section
• VersionPanel — version cards (with stage badges) + full-text dialog + Radix Tabs side-by-side diff

Workflow Engine

Six workflow types are supported, selectable per document run:

Workflow	Description
Head-to-Head (Parallel)	All models run the same prompt simultaneously for maximum diversity
Relay	Sequential chain — each model refines the previous model's output
Committee	All models draft in parallel → synthesizer combines the best elements
Critique Round	Parallel draft → models cross-critique each other's drafts (rotating) → synthesizer revises
EBOM Pipeline	Synthesizer writes brief/RFP → models draft → synthesizer composites → models critique → synthesizer final revision
Expert Panel	Each model adopts a named expert persona (Tax Attorney, Financial Advisor, etc.) → all run in parallel → synthesizer synthesizes views

Workflows requiring a synthesizer model show a dropdown (populated from the selected models). Expert Panel shows an editable textarea of personas. All stage outputs are saved as DocumentVersion entries with stageIndex/stageName/workflowType/persona metadata and appear in the VersionPanel for comparison.

Extended Model Provider Support

The ModelSelector and all API layers now support:

Group	Models	Notes
OpenAI	gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-3.5-turbo	Direct API
Anthropic	claude-3-5-sonnet, claude-3-opus, claude-3-haiku	Direct API
Perplexity	sonar, sonar-pro	Direct Perplexity API
Gemini	gemini-2.0-flash, gemini-1.5-pro, gemini-1.5-flash	Google provider
Groq	llama-3.3-70b, llama-3.1-8b, mixtral-8x7b	Routed via OpenRouter
Cohere	command-r+, command-r	Routed via OpenRouter
DeepSeek	deepseek-chat, deepseek-reasoner	⚠ Inline data-sovereignty caution (CN servers)
Local / Open	Mistral 7B free, Gemma 2 9B free	OpenRouter free tier
Custom	Any provider + model ID	Free-entry for Ollama, LM Studio, future providers

DeepSeek models display an inline warning when selected: "Data routed to DeepSeek servers (CN). Do not use with client-identifiable information."

Anonymization for Sensitive Documents

A "Anonymize for sensitive documents" toggle is added to the DocumentEditor. When enabled, prompts are passed through Cyrano's existing ClientAnonymizationService (token-based PII anonymization, MRPC 1.6 aligned) before being sent to any AI provider:

• Names, organizations, dates, amounts, emails, phones, case references → replaced with deterministic session-scoped tokens
• Tokens reversed locally after the AI responds — the provider never sees original identities
• Category 3 content (SSN, account numbers, date of birth) is blocked outright and never transmitted

Cyrano HTTP Bridge — MAE write endpoints

POST /api/mae/write        # single-model via aiService.call()
POST /api/mae/write/multi  # parallel multi-model, returns { versions[] }
GET  /api/mae/models       # expanded model catalogue

Both write endpoints accept anonymize?: boolean and forward it to aiService.call(). Provider enum expanded to include perplexity, google, deepseek, openrouter, groq, cohere. groq and cohere are resolved to the openrouter provider at runtime.

pdf_form_filler also added to the lazy-load tool cache.

MCP Server — pdf_form_filler registration (fix)

pdf_form_filler existed and was used internally by TaxForecastModule, ChildSupportForecastModule, and QDROForecastModule, but was absent from both ListToolsRequestSchema and the CallTool dispatch — making it invisible to any MCP client.

AI Orchestrator — provider normalization fix

normalizeProviderName() in both services/ai-orchestrator.ts and tools/ai-orchestrator.ts now maps openrouter, groq, and cohere (was throwing Unknown AI provider for these).

[!WARNING]

Original prompt

Implement completion of the standalone LexFiat Forecaster™ application and create a new standalone thin client for the MAE (Multi-Agent Engine) for document writing with multiple models and versions.

LexFiat Forecaster™ Completion

Complete the standalone forecaster application with PDF form filling, workflow integration, and branding implementation.

Current Status

• ✅ Standalone app exists in apps/forecaster/
• ✅ Frontend React UI functional
• ✅ Backend API with tax calculations
• ✅ HTTP bridge compatibility endpoints
• ❌ Missing PDF form filling for IRS/Michigan forms
• ❌ Workflow integration incomplete (tax_return_forecast, child_support_forecast, qdro_forecast)
• ❌ Branding implementation needs completion

Tasks

1. Implement PDF Form Filling

Files:

• Cyrano/src/tools/pdf-form-filler.ts (new or update)
• Cyrano/src/modules/forecast/formulas/tax-formulas.ts (update for form filling)

Requirements:

• Create pdf_form_filler MCP tool
• Implement fill_pdf_forms action in document_processor.ts
• Support IRS tax forms (1040, schedules) and Michigan state forms
• Integrate with existing tax calculations
• Add form validation and error handling

2. Complete Forecast Workflows

Files:

• Cyrano/src/engines/forecast/forecast-engine.ts (update)
• Workflow definitions for tax_return_forecast, child_support_forecast, qdro_forecast

Requirements:

• Register forecast workflows in MAE
• Implement child support calculations (Michigan FOC formula)
• Implement QDRO division scenarios
• Add workflow state management
• Integrate with PDF form filling

3. Implement Forecast Branding

Files:

• Cyrano/src/modules/forecast/branding.ts (new)
• apps/forecaster/backend/src/branding.ts (new)
• HTTP bridge endpoints (update)

Requirements:

• Apply mandatory branding strips/watermarks to all generated forecasts
• Support presentation modes: strip, watermark, none
• Implement risk acknowledgement for branding removal
• Auto-reset branding after 24 hours
• Ensure branding is enforced in both standalone and Cyrano-integrated modes

4. Register Forecast Tools

Files:

• Cyrano/src/mcp-server.ts (update for tool registration)
• Cyrano/src/http-bridge.ts (update endpoints)

Requirements:

• Register pdf_form_filler tool in MCP server
• Register forecast calculation tools
• Update HTTP bridge with PDF generation endpoints
• Add proper error handling and validation

MAE Thin Client (Multi-Version Document Writer)

Create a standalone thin client application for the MAE engine that allows document writing using multiple AI models and versions, similar to the EBOM spec creation process.

Concept

A lightweight web application that:

• Connects to Cyrano's MAE engine
• Allows users to write documents iteratively using different AI models
• Generates multiple versions of the same content
• Provides comparison and merging capabilities
• Maintains document history and version control

Implementation Tasks

1. Create MAE Thin Client App

Files:

• apps/mae-client/ (new directory)
• apps/mae-client/frontend/src/ (React app)
• apps/mae-client/frontend/package.json
• apps/mae-client/README.md

Requirements:

• React 19 + TypeScript + Vite setup
• Tailwind CSS + Radix UI for styling
• Document editor with rich text capabilities
• Model selection dropdown (OpenAI, Anthropic, etc.)
• Version history panel
• Comparison view for different versions

2. MAE Client Backend/API

Files:

• apps/mae-client/backend/src/ (Express API)
• apps/mae-client/backend/package.json

Requirements:

• Local API server for the client
• Integration with Cyrano's MAE engine via HTTP bridge
• Document storage (local files or database)
• Version management system
• Authentication (optional, local only)

3. Multi-Version Document Generation

Files:

• apps/mae-client/frontend/src/components/DocumentEditor.tsx
• apps/mae-client/frontend/src/components/VersionPanel.tsx
• apps/mae-client/frontend/src/components/ModelSelector.tsx

Requirements:

• Document editor with prompt input
• Generate multiple versions using different models
• Parallel generation for speed
• Version comparison UI (side-by-side diff)
• Merge functionality for combining versions
• Export options (PDF, DOCX, Markdown)

4. Integration with Cyrano MAE

Files:

• Cyrano/src/http-bridge.ts (add MAE client endpoints)
• apps/mae-client/frontend/src/lib/mae-api.ts

Requirements:

• HTTP endpoints for MAE document generation
• Support for multiple model selection
• Streaming responses for real-time generation
• Error handling and retry logic
• Rate limiting integration

5. Document Version Control

Files:

• apps/mae-client/backend/src/version-control.ts
• Database schema for document versions

Requirements:

• Store document versions with metadata (model used, timestamp, prompt)
• Version comparison algorithms
• Automatic saving of drafts
• Restore previou...

This pull request was created from Copilot chat.

---

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.

[MightyPrytanis]

@copilot I'd like users to have the ability to add more models in particular, I'd like them to have access to perplexity sonar, Grok, and Gemini, as well as a locally hosted models and the option to add others like cohere and DeepSeek, (although DeepSeek should come with cautions) also is there a way to use the token-based anonymizing system that we developed for Lex Fiat in the standalone for when it is being used for sensitive documents?

[MightyPrytanis]

@copilot I also would like the ability to use more than one workflow. For EBOM, we had an orchestrator design a RFP, then three models prepared proposed specs, then the orchestrator combined the anonymized specs into a draft, then the draft went to the the three competing models for critique/comment, then back to the orchestrating model for final revision and submission. There are infinite variations on this, as well as straight head to head competition, relays, a "committee review," a panel of experts, etc etc. This standalone should be flexible enough to adapt to the needs of the user in this respect.

Also: how do we feel about resurrecting the "SwimMeet" name for this standalone app?

[Copilot]

@copilot I'd like users to have the ability to add more models in particular, I'd like them to have access to perplexity sonar, Grok, and Gemini, as well as a locally hosted models a...

Done in commit 7d58efb.

Extended model support added:

• Perplexity Sonar (sonar, sonar-pro) — direct Perplexity API (already in Cyrano)
• Gemini (gemini-2.0-flash, gemini-1.5-pro, gemini-1.5-flash) — Google provider
• Groq (llama-3.3-70b, llama-3.1-8b, mixtral-8x7b) — routed via OpenRouter
• Cohere (command-r+, command-r) — routed via OpenRouter
• DeepSeek (deepseek-chat, deepseek-reasoner) — with a visible ⚠ caution flag: "Data routed to DeepSeek servers (CN). Do not use with client-identifiable information." The warning appears inline when you select a DeepSeek model.
• Local / Open (Mistral 7B free, Gemma 2 9B free) — via OpenRouter's free tier
• Custom model entry — enter any provider + model ID + optional label to add models not in the list (covers any future providers or locally-hosted Ollama/LM Studio endpoints)

Anonymization (Cyrano's ClientAnonymizationService)
A toggle labelled "Anonymize for sensitive documents" appears above the model selector. When enabled, the prompt passes through Cyrano's token-based PII system before leaving for any AI provider — names, organizations, dates, amounts, emails, phones, etc. are replaced with deterministic session-scoped tokens, then reversed locally on the response. Category 3 content (SSN, account numbers, date of birth) is blocked outright and never sent. The toggle shows a brief plain-language explanation when active.

[Copilot]

@copilot I also would like the ability to use more than one workflow. For EBOM, we had an orchestrator design a RFP, then three models prepared proposed specs, then the orchestrator combined the anon...

Done in commits 170d396 + 237b692. And yes — SwimMeet it is.

6 workflow types implemented:

Workflow	Description
Head-to-Head	All models run the same prompt simultaneously (existing behavior)
Relay	Sequential chain — each model refines the previous model's output
Committee	All models draft in parallel → synthesizer combines the best elements
Critique Round	Parallel draft → models cross-critique each other's drafts (rotating) → synthesizer incorporates feedback into final revision
EBOM Pipeline	Synthesizer writes a detailed brief/RFP → models draft from brief → synthesizer composites → models critique composite → synthesizer final revision
Expert Panel	Each model adopts a named expert persona (Tax Attorney, Financial Advisor, Family Law Mediator, Business Valuator, Estate Planner, Forensic Accountant — editable) → all run in parallel → synthesizer synthesizes expert views

UI: A collapsible WorkflowSelector panel appears above the model list with 6 workflow cards. Workflows that need a synthesis model (committee, critique, ebom, panel) show a synthesizer dropdown (populated from the currently selected models). The panel of experts shows an editable textarea of personas, one per line. Generate is disabled with a warning if the selected workflow needs a synthesizer but none is picked.

Stage results appear below the version list in a vertical timeline (each stage collapsible, last stage auto-expanded). All stage outputs are also saved as DocumentVersion entries with stageIndex/stageName metadata and appear in the main VersionPanel for comparison.

The architecture is fully extensible — new workflow types can be added in workflow-engine.ts + WorkflowSelector.tsx without touching any other files.