Files
oh-my-openagent/src/auth/antigravity/thinking.ts
YeonGyu-Kim d444e62b20 feat(antigravity-auth): add request/response transformation with tools and thinking
🤖 GENERATED WITH ASSISTANCE OF OpenCode
2025-12-13 00:35:34 +09:00

572 lines
16 KiB
TypeScript

/**
* Antigravity Thinking Block Handler (Gemini only)
*
* Handles extraction and transformation of thinking/reasoning blocks
* from Gemini responses. Thinking blocks contain the model's internal
* reasoning process, available in `-high` model variants.
*
* Key responsibilities:
* - Extract thinking blocks from Gemini response format
* - Detect thinking-capable model variants (`-high` suffix)
* - Format thinking blocks for OpenAI-compatible output
*
* Note: This is Gemini-only. Claude models are NOT handled by Antigravity.
*/
/**
* Represents a single thinking/reasoning block extracted from Gemini response
*/
export interface ThinkingBlock {
/** The thinking/reasoning text content */
text: string
/** Optional signature for signed thinking blocks (required for multi-turn) */
signature?: string
/** Index of the thinking block in sequence */
index?: number
}
/**
* Raw part structure from Gemini response candidates
*/
export interface GeminiPart {
/** Text content of the part */
text?: string
/** Whether this part is a thinking/reasoning block */
thought?: boolean
/** Signature for signed thinking blocks */
thoughtSignature?: string
/** Type field for Anthropic-style format */
type?: string
/** Signature field for Anthropic-style format */
signature?: string
}
/**
* Gemini response candidate structure
*/
export interface GeminiCandidate {
/** Content containing parts */
content?: {
/** Role of the content (e.g., "model", "assistant") */
role?: string
/** Array of content parts */
parts?: GeminiPart[]
}
/** Index of the candidate */
index?: number
}
/**
* Gemini response structure for thinking block extraction
*/
export interface GeminiResponse {
/** Response ID */
id?: string
/** Array of response candidates */
candidates?: GeminiCandidate[]
/** Direct content (some responses use this instead of candidates) */
content?: Array<{
type?: string
text?: string
signature?: string
}>
/** Model used for response */
model?: string
}
/**
* Result of thinking block extraction
*/
export interface ThinkingExtractionResult {
/** Extracted thinking blocks */
thinkingBlocks: ThinkingBlock[]
/** Combined thinking text for convenience */
combinedThinking: string
/** Whether any thinking blocks were found */
hasThinking: boolean
}
/**
* Default thinking budget in tokens for thinking-enabled models
*/
export const DEFAULT_THINKING_BUDGET = 16000
/**
* Check if a model variant should include thinking blocks
*
* Returns true for model variants with `-high` suffix, which have
* extended thinking capability enabled.
*
* Examples:
* - `gemini-3-pro-high` → true
* - `gemini-2.5-pro-high` → true
* - `gemini-3-pro-preview` → false
* - `gemini-2.5-pro` → false
*
* @param model - Model identifier string
* @returns True if model should include thinking blocks
*/
export function shouldIncludeThinking(model: string): boolean {
if (!model || typeof model !== "string") {
return false
}
const lowerModel = model.toLowerCase()
// Check for -high suffix (primary indicator of thinking capability)
if (lowerModel.endsWith("-high")) {
return true
}
// Also check for explicit thinking in model name
if (lowerModel.includes("thinking")) {
return true
}
return false
}
/**
* Check if a model is thinking-capable (broader check)
*
* This is a broader check than shouldIncludeThinking - it detects models
* that have thinking capability, even if not explicitly requesting thinking output.
*
* @param model - Model identifier string
* @returns True if model supports thinking/reasoning
*/
export function isThinkingCapableModel(model: string): boolean {
if (!model || typeof model !== "string") {
return false
}
const lowerModel = model.toLowerCase()
return (
lowerModel.includes("thinking") ||
lowerModel.includes("gemini-3") ||
lowerModel.endsWith("-high")
)
}
/**
* Check if a part is a thinking/reasoning block
*
* Detects both Gemini-style (thought: true) and Anthropic-style
* (type: "thinking" or type: "reasoning") formats.
*
* @param part - Content part to check
* @returns True if part is a thinking block
*/
function isThinkingPart(part: GeminiPart): boolean {
// Gemini-style: thought flag
if (part.thought === true) {
return true
}
// Anthropic-style: type field
if (part.type === "thinking" || part.type === "reasoning") {
return true
}
return false
}
/**
* Check if a thinking part has a valid signature
*
* Signatures are required for multi-turn conversations with Claude models.
* Gemini uses `thoughtSignature`, Anthropic uses `signature`.
*
* @param part - Thinking part to check
* @returns True if part has valid signature
*/
function hasValidSignature(part: GeminiPart): boolean {
// Gemini-style signature
if (part.thought === true && part.thoughtSignature) {
return true
}
// Anthropic-style signature
if ((part.type === "thinking" || part.type === "reasoning") && part.signature) {
return true
}
return false
}
/**
* Extract thinking blocks from a Gemini response
*
* Parses the response structure to identify and extract all thinking/reasoning
* content. Supports both Gemini-style (thought: true) and Anthropic-style
* (type: "thinking") formats.
*
* @param response - Gemini response object
* @returns Extraction result with thinking blocks and metadata
*/
export function extractThinkingBlocks(response: GeminiResponse): ThinkingExtractionResult {
const thinkingBlocks: ThinkingBlock[] = []
// Handle candidates array (standard Gemini format)
if (response.candidates && Array.isArray(response.candidates)) {
for (const candidate of response.candidates) {
const parts = candidate.content?.parts
if (!parts || !Array.isArray(parts)) {
continue
}
for (let i = 0; i < parts.length; i++) {
const part = parts[i]
if (!part || typeof part !== "object") {
continue
}
if (isThinkingPart(part)) {
const block: ThinkingBlock = {
text: part.text || "",
index: thinkingBlocks.length,
}
// Extract signature if present
if (part.thought === true && part.thoughtSignature) {
block.signature = part.thoughtSignature
} else if (part.signature) {
block.signature = part.signature
}
thinkingBlocks.push(block)
}
}
}
}
// Handle direct content array (Anthropic-style response)
if (response.content && Array.isArray(response.content)) {
for (let i = 0; i < response.content.length; i++) {
const item = response.content[i]
if (!item || typeof item !== "object") {
continue
}
if (item.type === "thinking" || item.type === "reasoning") {
thinkingBlocks.push({
text: item.text || "",
signature: item.signature,
index: thinkingBlocks.length,
})
}
}
}
// Combine all thinking text
const combinedThinking = thinkingBlocks.map((b) => b.text).join("\n\n")
return {
thinkingBlocks,
combinedThinking,
hasThinking: thinkingBlocks.length > 0,
}
}
/**
* Format thinking blocks for OpenAI-compatible output
*
* Converts Gemini thinking block format to OpenAI's expected structure.
* OpenAI expects thinking content as special message blocks or annotations.
*
* Output format:
* ```
* [
* { type: "reasoning", text: "thinking content...", signature?: "..." },
* ...
* ]
* ```
*
* @param thinking - Array of thinking blocks to format
* @returns OpenAI-compatible formatted array
*/
export function formatThinkingForOpenAI(
thinking: ThinkingBlock[],
): Array<{ type: "reasoning"; text: string; signature?: string }> {
if (!thinking || !Array.isArray(thinking) || thinking.length === 0) {
return []
}
return thinking.map((block) => {
const formatted: { type: "reasoning"; text: string; signature?: string } = {
type: "reasoning",
text: block.text || "",
}
if (block.signature) {
formatted.signature = block.signature
}
return formatted
})
}
/**
* Transform thinking parts in a candidate to OpenAI format
*
* Modifies candidate content parts to use OpenAI-style reasoning format
* while preserving the rest of the response structure.
*
* @param candidate - Gemini candidate to transform
* @returns Transformed candidate with reasoning-formatted thinking
*/
export function transformCandidateThinking(candidate: GeminiCandidate): GeminiCandidate {
if (!candidate || typeof candidate !== "object") {
return candidate
}
const content = candidate.content
if (!content || typeof content !== "object" || !Array.isArray(content.parts)) {
return candidate
}
const thinkingTexts: string[] = []
const transformedParts = content.parts.map((part) => {
if (part && typeof part === "object" && part.thought === true) {
thinkingTexts.push(part.text || "")
// Transform to reasoning format
return {
...part,
type: "reasoning" as const,
thought: undefined, // Remove Gemini-specific field
}
}
return part
})
const result: GeminiCandidate & { reasoning_content?: string } = {
...candidate,
content: { ...content, parts: transformedParts },
}
// Add combined reasoning content for convenience
if (thinkingTexts.length > 0) {
result.reasoning_content = thinkingTexts.join("\n\n")
}
return result
}
/**
* Transform Anthropic-style thinking blocks to reasoning format
*
* Converts `type: "thinking"` blocks to `type: "reasoning"` for consistency.
*
* @param content - Array of content blocks
* @returns Transformed content array
*/
export function transformAnthropicThinking(
content: Array<{ type?: string; text?: string; signature?: string }>,
): Array<{ type?: string; text?: string; signature?: string }> {
if (!content || !Array.isArray(content)) {
return content
}
return content.map((block) => {
if (block && typeof block === "object" && block.type === "thinking") {
return {
type: "reasoning",
text: block.text || "",
...(block.signature ? { signature: block.signature } : {}),
}
}
return block
})
}
/**
* Filter out unsigned thinking blocks
*
* Claude API requires signed thinking blocks for multi-turn conversations.
* This function removes thinking blocks without valid signatures.
*
* @param parts - Array of content parts
* @returns Filtered array without unsigned thinking blocks
*/
export function filterUnsignedThinkingBlocks(parts: GeminiPart[]): GeminiPart[] {
if (!parts || !Array.isArray(parts)) {
return parts
}
return parts.filter((part) => {
if (!part || typeof part !== "object") {
return true
}
// If it's a thinking part, only keep it if signed
if (isThinkingPart(part)) {
return hasValidSignature(part)
}
// Keep all non-thinking parts
return true
})
}
/**
* Transform entire response thinking parts
*
* Main transformation function that handles both Gemini-style and
* Anthropic-style thinking blocks in a response.
*
* @param response - Response object to transform
* @returns Transformed response with standardized reasoning format
*/
export function transformResponseThinking(response: GeminiResponse): GeminiResponse {
if (!response || typeof response !== "object") {
return response
}
const result: GeminiResponse = { ...response }
// Transform candidates (Gemini-style)
if (Array.isArray(result.candidates)) {
result.candidates = result.candidates.map(transformCandidateThinking)
}
// Transform direct content (Anthropic-style)
if (Array.isArray(result.content)) {
result.content = transformAnthropicThinking(result.content)
}
return result
}
/**
* Thinking configuration for requests
*/
export interface ThinkingConfig {
/** Token budget for thinking/reasoning */
thinkingBudget?: number
/** Whether to include thoughts in response */
includeThoughts?: boolean
}
/**
* Normalize thinking configuration
*
* Ensures thinkingConfig is valid: includeThoughts only allowed when budget > 0.
*
* @param config - Raw thinking configuration
* @returns Normalized configuration or undefined
*/
export function normalizeThinkingConfig(config: unknown): ThinkingConfig | undefined {
if (!config || typeof config !== "object") {
return undefined
}
const record = config as Record<string, unknown>
const budgetRaw = record.thinkingBudget ?? record.thinking_budget
const includeRaw = record.includeThoughts ?? record.include_thoughts
const thinkingBudget =
typeof budgetRaw === "number" && Number.isFinite(budgetRaw) ? budgetRaw : undefined
const includeThoughts = typeof includeRaw === "boolean" ? includeRaw : undefined
const enableThinking = thinkingBudget !== undefined && thinkingBudget > 0
const finalInclude = enableThinking ? (includeThoughts ?? false) : false
// Return undefined if no meaningful config
if (
!enableThinking &&
finalInclude === false &&
thinkingBudget === undefined &&
includeThoughts === undefined
) {
return undefined
}
const normalized: ThinkingConfig = {}
if (thinkingBudget !== undefined) {
normalized.thinkingBudget = thinkingBudget
}
if (finalInclude !== undefined) {
normalized.includeThoughts = finalInclude
}
return normalized
}
/**
* Extract thinking configuration from request payload
*
* Supports both Gemini-style thinkingConfig and Anthropic-style thinking options.
*
* @param requestPayload - Request body
* @param generationConfig - Generation config from request
* @param extraBody - Extra body options
* @returns Extracted thinking configuration or undefined
*/
export function extractThinkingConfig(
requestPayload: Record<string, unknown>,
generationConfig?: Record<string, unknown>,
extraBody?: Record<string, unknown>,
): ThinkingConfig | undefined {
// Check for explicit thinkingConfig
const thinkingConfig =
generationConfig?.thinkingConfig ?? extraBody?.thinkingConfig ?? requestPayload.thinkingConfig
if (thinkingConfig && typeof thinkingConfig === "object") {
const config = thinkingConfig as Record<string, unknown>
return {
includeThoughts: Boolean(config.includeThoughts),
thinkingBudget:
typeof config.thinkingBudget === "number" ? config.thinkingBudget : DEFAULT_THINKING_BUDGET,
}
}
// Convert Anthropic-style "thinking" option: { type: "enabled", budgetTokens: N }
const anthropicThinking = extraBody?.thinking ?? requestPayload.thinking
if (anthropicThinking && typeof anthropicThinking === "object") {
const thinking = anthropicThinking as Record<string, unknown>
if (thinking.type === "enabled" || thinking.budgetTokens) {
return {
includeThoughts: true,
thinkingBudget:
typeof thinking.budgetTokens === "number"
? thinking.budgetTokens
: DEFAULT_THINKING_BUDGET,
}
}
}
return undefined
}
/**
* Resolve final thinking configuration based on model and context
*
* Handles special cases like Claude models requiring signed thinking blocks
* for multi-turn conversations.
*
* @param userConfig - User-provided thinking configuration
* @param isThinkingModel - Whether model supports thinking
* @param isClaudeModel - Whether model is Claude (not used in Antigravity, but kept for compatibility)
* @param hasAssistantHistory - Whether conversation has assistant history
* @returns Final thinking configuration
*/
export function resolveThinkingConfig(
userConfig: ThinkingConfig | undefined,
isThinkingModel: boolean,
isClaudeModel: boolean,
hasAssistantHistory: boolean,
): ThinkingConfig | undefined {
// Claude models with history need signed thinking blocks
// Since we can't guarantee signatures, disable thinking
if (isClaudeModel && hasAssistantHistory) {
return { includeThoughts: false, thinkingBudget: 0 }
}
// Enable thinking by default for thinking-capable models
if (isThinkingModel && !userConfig) {
return { includeThoughts: true, thinkingBudget: DEFAULT_THINKING_BUDGET }
}
return userConfig
}