💡 Platform feature: Gemini model abstraction layer (config-driven model selection with automatic version fallback)

[github-actions[bot]]

Summary

Create a shared Gemini model management utility in gas-utils.js that centralizes model version selection, provides fallback chains when a model is deprecated, and standardizes the UrlFetchApp call pattern across all AI-powered scripts. This prevents the recurring breakage seen when Google deprecates Gemini model versions (2.0 Flash was shut down June 1, 2026) and reduces per-script boilerplate as the catalog adds more AI-powered automations.

Market Signal

Gemini 2.0 Flash was deprecated and shut down on June 1, 2026 — any script hardcoding that model string is now broken. Google's rapid model iteration cycle (1.5 → 2.0 → 3.5 Flash in ~18 months) means this will happen again. The gmail-ai-classifier README still references gemini-2.0-flash as a supported model. Meanwhile, every major AI framework (LangChain, LiteLLM, etc.) has adopted model abstraction as a core pattern because model churn is the norm, not the exception.

Sources:

• Gemini 2.0 Flash documentation — deprecation notice, shut down June 1, 2026
• Gemini 3.5 Flash — current recommended model

User Signal

The gmail-ai-classifier is under active development (PRs #605, #606 merged Sep 29–30 with prompt architecture migration). Issue #491 proposes an AI-Powered Gemini Email Classifier. Multiple existing idea Discussions propose new Gemini-powered scripts: #597 (gmail-action-extractor), #587 (AI-assisted config wizard), #589 (AI-structured archive), #552 (AI daily briefing). Each of these will face the same model version migration burden unless there is a shared abstraction.

Technical Opportunity

gas-utils.js already serves as the shared utility module. A getGeminiResponse(prompt, options) function that reads the model version from config.gs and falls back through a version chain (e.g., gemini-3.5-flash → gemini-2.5-flash) would be ~30 lines. The existing UrlFetchApp pattern in gmail-ai-classifier's code.gs is the starting template. Structured output support (responseSchema) can be standardized across all AI scripts.

Proposed function signature:

// gas-utils.js
function getGeminiResponse(prompt, { model, apiKey, responseSchema, fallbackModels } = {}) {
  // Attempts primary model, falls back through chain on 404/deprecated errors
  // Returns parsed JSON response or structured output
}

Assessment

Dimension	Score	Rationale
Feasibility	high	~30 lines in an existing shared module; pattern proven in gmail-ai-classifier
Impact	med	Infrastructure improvement — enables all future AI scripts, prevents breakage
Urgency	high	Gemini 2.0 Flash already shut down; classifier README references it; more AI scripts planned

Adversarial Review

Strongest objection: This is over-engineering a simple string change. Just update the model name when Google deprecates it. Different scripts may need different models (cheap vs. capable) so a single abstraction doesn't fit.

Rebuttal: The "just change the string" approach already failed — gemini-2.0-flash is shut down and the README still references it. With 5+ planned AI scripts, a model deprecation now requires 5+ coordinated changes instead of 1. The abstraction handles the "different models for different scripts" case naturally: each config.gs specifies its preferred model, and the shared utility handles the UrlFetchApp call pattern, error handling, and fallback logic. The complexity is in the call boilerplate (headers, JSON parsing, error codes, rate limiting), not the model string.

Suggested Next Step

Audit current Gemini API usage in gmail-ai-classifier to catalog the exact call pattern (headers, payload shape, error handling). Design the gas-utils.js function signature. Update config.gs to use gemini-3.5-flash as the primary model. Write tests for the fallback chain logic.