Update Gemini API spec from Discovery, regenerate bindings, and add live (multimodal) integration tests

[o1lo01ol1o]

PR Description (Detailed)

Context / Motivation

This repo is an auto-generated low-level Haskell client for the Gemini (Generative Language) API. The
upstream OpenAPI spec link from the public docs was unavailable/404, and the repo’s existing spec had
fallen behind the latest API revision. We needed a repeatable way to:

• fetch the most current API description,
• convert it into an OpenAPI document that openapi-generator can consume, and
• regenerate the Haskell client reliably.

What Changed

1) Update the API spec source of truth

• Added the latest Google Discovery document to the repo: discovery-v1beta.json (added in commit 857cf98).
• Updated openapi.yaml to reflect the newer Discovery revision (commit 857cf98).

Why

• openapi-generator cannot consume Google Discovery format directly.
• The “official OpenAPI download URL” referenced via ai.google.dev was returning 404, so Discovery became
  the stable upstream artifact.

2) Add a reproducible Discovery → OpenAPI generation workflow

• Added flake input: flake.nix now pulls github:stackql/google-discovery-to-openapi (commit f8a9f56).
• Added Nix devshell tooling in nix/modules/flake/devshell.nix:
  • google-discovery-to-openapi: runs the upstream StackQL converter.
  • genai-generate-openapi: fetches https://generativelanguage.googleapis.com/$discovery/rest?
    version=v1beta and writes ./openapi.yaml for this repo.
  • Includes nodejs + python3 in the dev environment (commit f8a9f56), plus later additions (zlib, pkg-
    config) to build native deps required by Haskell packages.
  • Adds devShells.codegen to avoid pulling the full Haskell toolchain when you only need spec/codegen.

Why

• Keeps regeneration deterministic and “one command”.
• Avoids depending on ad-hoc manual spec downloads.
• Avoids disk-heavy nix develop paths for codegen-only flows.

3) Sanitize the converted OpenAPI so generators don’t choke

Inside genai-generate-openapi (embedded Node script in nix/modules/flake/devshell.nix), we sanitize schemas
to remove Discovery-only / invalid OpenAPI keywords and prevent codegen collisions.

Key fixes:

• Remove Discovery schema keywords like id, enumDescriptions, enumDeprecated.
• Convert Discovery type: any into an OpenAPI-friendly “free-form object”.
• Drop underscore-collision aliases (e.g. both _responseJsonSchema and responseJsonSchema) which otherwise
  cause duplicate record fields in Haskell.

Why

• openapi-generator validates the spec and rejects unknown fields.
• The Haskell generator normalizes names and can create duplicate record fields, which GHC rejects (your
  generationConfigResponseJsonSchema collision).

4) Regenerate the Haskell client from the updated spec

• Updated generator script to use the repo spec: run-generator.sh uses --input-spec openapi.yaml (commit
  f8a9f56).
• Regenerated outputs (commit 51c054a), including:
  • new lib/GenAI/Client/API/ApiDefault.hs
  • removed stale lib/GenAI/Client/API/Generativelanguage.hs
  • updated generated lib/GenAI/Client/Model.hs, Client.hs, etc.
• Updated .openapi-generator/* metadata files accordingly.

Why

• Ensures the generated Haskell bindings match the new API surface and revision.

5) Add live integration tests that hit the real API using generated client code

• Added tests/Live.hs (commit 240dade, plus follow-ups) and a Cabal test-suite entry test-suite live in
  haskell-google-genai-client.cabal.
• Live tests cover:
  • listModels (prints models when GEMINI_PRINT_MODELS=1)
  • generateContent (text-only)
  • multimodal image+text test that downloads a Haskell-logo image and asks the model to identify the
    language.

Live-test environment variables:

• GEMINI_API_KEY (required)
• GEMINI_MODEL (optional; defaults to a flash model)
• GEMINI_IMAGE_MODEL (optional; defaults to gemini-3-pro-image-preview)
• HASKELL_LOGO_URL (optional; defaults to GitHub avatar PNG)
• GEMINI_HTTP_TIMEOUT_SECONDS (optional; defaults to 120)
• GEMINI_PRINT_MODELS=1 to print available model names
• GEMINI_PRINT_MULTIMODAL_RESPONSE=1 to print multimodal response text parts

Why

• We want to validate that the generated library can successfully talk to the live API (not just compile).
• Multimodal specifically validates inlineData (base64 bytes) support and multipart content structure.

6) Improve developer ergonomics / documentation

• Updated README.md with:
  • spec refresh command using .#codegen
  • live test commands
  • model listing commands

Why

• Keeps the regeneration workflow discoverable and consistent for contributors.

How to Reproduce / Regenerate

Update spec and regenerate bindings

• nix develop .#codegen -c bash -c 'genai-generate-openapi && ./run-generator.sh'

Run live tests

• List models:
  GEMINI_API_KEY='...' GEMINI_PRINT_MODELS=1 cabal test live --test-options='--match listModels'
• Text gen test:
  GEMINI_API_KEY='...' GEMINI_MODEL='models/gemini-3-pro-preview' cabal test live --test-options='--match
  generateContent'
• Multimodal test:
  GEMINI_API_KEY='...' GEMINI_IMAGE_MODEL='models/gemini-3-pro-image-preview'
  GEMINI_HTTP_TIMEOUT_SECONDS=180 cabal test live --test-options='--match multimodal'

Notes / Known Tradeoffs

• The StackQL converter repo brings in Node dependencies; npm ci may report upstream vulnerabilities. This
  doesn’t ship in the Haskell library artifact, but it’s worth tracking if you want to vendor/pin/fork
  the converter.
• Live tests depend on external network + an API key and can be flaky due to model latency, rate limits, or
  transient image decoding errors. Timeouts are now configurable.