[Kem-Ada]

Summary

Adds GeminiGateway (alias IdentArkGeminiGateway), a port of the Python SDK's identark/integrations/gemini.py. It is published at the subpath identark/integrations/gemini and is deliberately not re-exported from the barrel, the same as the LangChain and Vercel integrations. It depends on @google/generative-ai, added as an optional peer dependency. Python's DirectGateway has no Gemini branch, so this PR does not touch DirectGateway.

Python → TS parity mapping

Python (identark/integrations/gemini.py)	TypeScript (src/integrations/gemini.ts)
GeminiGateway(api_key, model, system_prompt, cost_cap_usd, workspace_dir, safety_settings, generation_config)	new GeminiGateway({ apiKey, model, systemPrompt, costCapUsd, workspaceDir, safetySettings, generationConfig, client? })
ConfigurationError on empty key/model or missing SDK	Same. The SDK is loaded with a synchronous createRequire, so construction stays sync.
invoke_llm / invoke_llm_stream	invokeLlm / invokeLlmStream
persist_messages, request_file_url, get_session_cost, reset	persistMessages, requestFileUrl, getSessionCost, reset
history, model, provider properties	getHistory(), getModel(), getProvider() (same as DirectGateway)
_estimate_gemini_cost (private)	estimateGeminiCost (exported for testing)
_convert_tools_to_gemini, _convert_role_to_gemini	convertToolsToGemini, convertRoleToGemini
_classify_gemini_error (substring ladder)	_classifyGeminiError, same ladder and order
IdentArkGeminiGateway alias	IdentArkGeminiGateway

Deliberate differences, all forced by the JS SDK's shapes:

• Parts are camelCase (functionResponse, inlineData.mimeType, functionDeclarations).
• Finish reasons are matched by both numeric and string enum.
• Structured history content is JSON.stringify'd. Python uses str(list), which is a repr and not portable.
• client can be injected. The tests use it, since the SDK is not a devDependency.

Python quirks kept on purpose (rule 6):

• A tool result's function name is the tool_call_id.
• The stream's final finish_reason is always "stop".
• Streaming drops function calls.
• tool_choice is accepted and ignored.
• Unknown models are priced as flash.

Test plan

Gate, run from identark-js on branch feat/gemini:

npm ci               -> added 235 packages (clean)
npm run lint         -> eslint src --max-warnings 0: clean
npm run typecheck    -> tsc --noEmit: clean
npx vitest run       -> Test Files 10 passed (10); Tests 231 passed (231)
../scripts/check_scope.sh -> check_scope: OK — no control-plane code touched.

• Upstream/main baseline was 158 tests / 9 files. This PR adds 73 tests in tests/gemini-integration.test.ts.
• The tests port all 14 cases in tests/unit/test_integrations_gemini.py.
• Negative and edge cases beyond the Python tests:
  • constructor validation
  • each finish-reason mapping
  • multimodal, tool-result and system-role conversion
  • cost-cap enforcement on both the blocking and streaming paths
  • the full error-classification ladder (429/quota/rate, safety/blocked/harm/content, invalid API key, generic, non-Error rejections)
  • no history or cost recorded on failure
  • mid-stream errors
  • Object.prototype model names in cost estimation

Control-plane: untouched (check_scope OK)

🤖 Generated with Claude Code