AI Providers

This document describes provider-specific behavior and options implemented in indusagi/src/ai/providers/*.

Environment Variables

These are read by getEnvApiKey in indusagi/src/ai/env-api-keys.ts.

• OPENAI_API_KEY for provider openai
• AZURE_OPENAI_API_KEY for provider azure-openai-responses
• GEMINI_API_KEY for provider google
• GROQ_API_KEY for provider groq
• CEREBRAS_API_KEY for provider cerebras
• XAI_API_KEY for provider xai
• OPENROUTER_API_KEY for provider openrouter
• AI_GATEWAY_API_KEY for provider vercel-ai-gateway
• ZAI_API_KEY for provider zai
• MISTRAL_API_KEY for provider mistral
• MINIMAX_API_KEY for provider minimax
• MINIMAX_CN_API_KEY for provider minimax-cn
• OPENCODE_API_KEY for provider opencode
• COPILOT_GITHUB_TOKEN or GH_TOKEN or GITHUB_TOKEN for provider github-copilot
• ANTHROPIC_OAUTH_TOKEN or ANTHROPIC_API_KEY for provider anthropic

Google Vertex uses Application Default Credentials and requires:

• GOOGLE_APPLICATION_CREDENTIALS or the default ADC file in ~/.config/gcloud/application_default_credentials.json
• GOOGLE_CLOUD_PROJECT or GCLOUD_PROJECT
• GOOGLE_CLOUD_LOCATION

Amazon Bedrock uses standard AWS auth sources:

• AWS_PROFILE or AWS_ACCESS_KEY_ID plus AWS_SECRET_ACCESS_KEY
• AWS_BEARER_TOKEN_BEDROCK
• AWS_CONTAINER_CREDENTIALS_RELATIVE_URI or AWS_CONTAINER_CREDENTIALS_FULL_URI
• AWS_WEB_IDENTITY_TOKEN_FILE

OpenAI Completions (`openai-completions`)

Implementation: providers/openai-completions.ts

Options:

• toolChoice?: "auto" | "none" | "required" | { type: "function"; function: { name: string } }
• reasoningEffort?: "minimal" | "low" | "medium" | "high" | "xhigh"

Highlights:

• Supports images via content parts when model allows input: ["text", "image"].
• Handles tool calls and tool results, including providers that require special tool call IDs.
• Supports provider compatibility overrides via model.compat.
• Adds OpenRouter Anthropic cache control to the last user or assistant text block.

Compatibility overrides for model.compat include:

• supportsStore, supportsDeveloperRole, supportsReasoningEffort
• supportsUsageInStreaming, maxTokensField
• requiresToolResultName, requiresAssistantAfterToolResult
• requiresThinkingAsText, requiresMistralToolIds
• thinkingFormat, openRouterRouting

OpenAI Responses (`openai-responses`)

Implementation: providers/openai-responses.ts and providers/openai-responses-shared.ts

Options:

• reasoningEffort?: "minimal" | "low" | "medium" | "high" | "xhigh"
• reasoningSummary?: "auto" | "detailed" | "concise" | null
• serviceTier?: "auto" | "flex" | "priority"

Highlights:

• Uses the Responses API and processResponsesStream for event handling.
• Applies service tier cost multipliers for flex and priority tiers.
• When reasoning is disabled on some GPT-5 models, it injects # Juice: 0 !important to disable reasoning.

Azure OpenAI Responses (`azure-openai-responses`)

Implementation: providers/azure-openai-responses.ts

Options:

• azureApiVersion?: string
• azureResourceName?: string
• azureBaseUrl?: string
• azureDeploymentName?: string
• reasoningEffort?: "minimal" | "low" | "medium" | "high" | "xhigh"
• reasoningSummary?: "auto" | "detailed" | "concise" | null

Environment variables:

• AZURE_OPENAI_BASE_URL
• AZURE_OPENAI_RESOURCE_NAME
• AZURE_OPENAI_API_VERSION
• AZURE_OPENAI_DEPLOYMENT_NAME_MAP (comma separated modelId=deploymentName pairs)

OpenAI Codex Responses (`openai-codex-responses`)

Implementation: providers/openai-codex-responses.ts

Options:

• reasoningEffort?: "none" | "minimal" | "low" | "medium" | "high" | "xhigh"
• reasoningSummary?: "auto" | "concise" | "detailed" | "off" | "on" | null
• textVerbosity?: "low" | "medium" | "high"

Highlights:

• Uses ChatGPT Codex endpoint with SSE parsing and retry logic.
• Requires a JWT access token, with account ID extracted from the token payload.
• Supports sessionId which maps to prompt_cache_key.

Anthropic Messages (`anthropic-messages`)

Implementation: providers/anthropic.ts

Options:

• thinkingEnabled?: boolean
• thinkingBudgetTokens?: number
• interleavedThinking?: boolean
• toolChoice?: "auto" | "any" | "none" | { type: "tool"; name: string }

Highlights:

• Supports both API keys and OAuth tokens.
• When using OAuth tokens, it mimics Claude Code headers and tool naming.
• Supports tool calls, tool results, and image inputs.

Amazon Bedrock (`bedrock-converse-stream`)

Implementation: providers/amazon-bedrock.ts

Options:

• region?: string
• profile?: string
• toolChoice?: "auto" | "any" | "none" | { type: "tool"; name: string }
• reasoning?: ThinkingLevel
• thinkingBudgets?: ThinkingBudgets
• interleavedThinking?: boolean

Highlights:

• Uses AWS SDK ConverseStreamCommand.
• Supports tool calls and tool results in a single user message (Bedrock requirement).
• Adds prompt caching blocks for supported Claude models.
• Handles model-specific support for reasoning signatures.

Google Generative AI (`google-generative-ai`)

Implementation: providers/google.ts and providers/google-shared.ts

Options:

• toolChoice?: "auto" | "none" | "any"
• thinking?: { enabled: boolean; budgetTokens?: number; level?: GoogleThinkingLevel }

Highlights:

• Supports text and image input.
• Uses thought signatures for reasoning and tool calls when available.
• Distinguishes Gemini 3 models that require thinkingLevel instead of thinkingBudget.

Google Vertex (`google-vertex`)

Implementation: providers/google-vertex.ts and providers/google-shared.ts

Options:

• project?: string
• location?: string
• toolChoice?: "auto" | "none" | "any"
• thinking?: { enabled: boolean; budgetTokens?: number; level?: GoogleThinkingLevel }

Highlights:

• Uses Vertex AI configuration with project and location.
• Uses the same message conversion and tool handling as google-generative-ai.

Google Gemini CLI / Antigravity (`google-gemini-cli` and `google-antigravity`)

Implementation: providers/google-gemini-cli.ts plus OAuth providers in utils/oauth.

Options:

• projectId?: string
• toolChoice?: "auto" | "none" | "any"
• thinking?: { enabled: boolean; budgetTokens?: number; level?: GoogleThinkingLevel }

Highlights:

• Uses Cloud Code Assist endpoint and requires OAuth credentials.
• Supports both Gemini CLI and Antigravity endpoints with retry logic.
• Injects system instructions for Antigravity and supports thought signatures.

OpenAI-Compatible Providers

The OpenAI Completions provider handles compatibility for multiple non-OpenAI backends. Providers include openrouter, xai, groq, cerebras, mistral, zai, and opencode. Compatibility is auto-detected from provider and baseUrl, or overridden via model.compat.