luulam/9router
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

end

Browse Source
This commit is contained in:
luulam
2026-07-06 00:01:34 +07:00
parent 2729408ef3
commit e7470e955e
108 changed files with 4131 additions and 529 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
open-sse/translator/concerns/usage.js
View File

@@ -39,7 +39,16 @@ const USAGE_EXTRACTORS = {
},
kiro(raw) {
const input = n(raw.inputTokens), output = n(raw.outputTokens);
return { promptTokens: input, completionTokens: output, totalTokens: input + output };
// ponytail: Amazon Q (Kiro upstream) does not expose cache fields today,
// but pass through any cache_read/cache_creation/cached_tokens if the
// event shape grows them later so cost tracking keeps working without
// a second pass.
const cached = n(raw.cache_read_input_tokens) || n(raw.cachedTokens) || n(raw.cached_tokens);
const cacheCreation = n(raw.cache_creation_input_tokens);
const out = { promptTokens: input, completionTokens: output, totalTokens: input + output };
if (cached > 0) out.cachedTokens = cached;
if (cacheCreation > 0) out.cacheCreationTokens = cacheCreation;
return out;
},
ollama(raw) {
const input = n(raw.prompt_eval_count), output = n(raw.eval_count);

27
open-sse/translator/formats/claude.js
View File

@@ -150,6 +150,33 @@ export function normalizeClaudePassthrough(body, model = "") {
}
}
// 3. Drop thinking blocks whose signature is not Claude's (combo mixes models,
// so foreign signatures leak into history and Anthropic rejects them).
const thinkingEnabled = body.thinking?.type === "enabled";
if (Array.isArray(body.messages)) {
for (const msg of body.messages) {
if (msg.role !== ROLE.ASSISTANT || !Array.isArray(msg.content)) continue;
let hasToolUse = false;
let hasKeptThinking = false;
const kept = [];
for (const block of msg.content) {
if (block.type === CLAUDE_BLOCK.THINKING || block.type === CLAUDE_BLOCK.REDACTED_THINKING) {
if (isValidClaudeSignature(block.signature)) {
hasKeptThinking = true;
kept.push(block);
}
continue;
}
if (block.type === CLAUDE_BLOCK.TOOL_USE) hasToolUse = true;
kept.push(block);
}
msg.content = kept;
if (thinkingEnabled && !hasKeptThinking && hasToolUse) {
msg.content.unshift(buildThinkingPlaceholder("claude"));
}
}
}
return body;
}

4
open-sse/translator/formats/gemini.js
View File

@@ -12,6 +12,8 @@ export const UNSUPPORTED_SCHEMA_CONSTRAINTS = [
"default", "examples",
// JSON Schema meta keywords
"$schema", "$defs", "definitions", "const", "$ref", "$comment",
// Annotation keywords (rejected by Gemini/Antigravity - e.g. MCP tool schemas set these)
"deprecated", "readOnly", "writeOnly",
// Object validation keywords (not supported)
"additionalProperties", "propertyNames", "patternProperties", "enumDescriptions",
// Complex schema keywords (handled by flattenAnyOfOneOf/mergeAllOf)
@@ -19,7 +21,7 @@ export const UNSUPPORTED_SCHEMA_CONSTRAINTS = [
// Dependency keywords (not supported)
"dependencies", "dependentSchemas", "dependentRequired",
// Other unsupported keywords
"title", "optional", "if", "then", "else", "contentMediaType", "contentEncoding",
"title", "optional", "deprecated", "if", "then", "else", "contentMediaType", "contentEncoding",
// UI/Styling properties (from Cursor tools - NOT JSON Schema standard)
"cornerRadius", "fillColor", "fontFamily", "fontSize", "fontWeight",
"gap", "padding", "strokeColor", "strokeThickness", "textColor"

30
open-sse/translator/formats/openai.js
View File

@@ -6,42 +6,46 @@ export { VALID_OPENAI_CONTENT_TYPES, VALID_OPENAI_MESSAGE_TYPES };
// Filter messages to OpenAI standard format
// Remove: thinking, redacted_thinking, signature, and other non-OpenAI blocks
export function filterToOpenAIFormat(body) {
// opts.preserveCacheControl: keep cache_control on content blocks (e.g. for DashScope/alicode)
export function filterToOpenAIFormat(body, opts = {}) {
if (!body.messages || !Array.isArray(body.messages)) return body;
const keepCache = !!opts.preserveCacheControl;
function stripBlock(block) {
const { signature, cache_control, ...rest } = block;
return keepCache && cache_control ? { ...rest, cache_control } : rest;
}
body.messages = body.messages.map(msg => {
// Normalize developer role to system (many providers don't support developer)
if (msg.role === ROLE.DEVELOPER) msg = { ...msg, role: ROLE.SYSTEM };
// Keep tool messages as-is (OpenAI format)
if (msg.role === ROLE.TOOL) return msg;
// Keep assistant messages with tool_calls as-is
if (msg.role === ROLE.ASSISTANT && msg.tool_calls) return msg;
// Handle string content
if (typeof msg.content === "string") return msg;
// Handle array content
if (Array.isArray(msg.content)) {
const filteredContent = [];
for (const block of msg.content) {
// Skip thinking blocks
if (block.type === CLAUDE_BLOCK.THINKING || block.type === CLAUDE_BLOCK.REDACTED_THINKING) continue;
// Only keep valid OpenAI content types
if (VALID_OPENAI_CONTENT_TYPES.includes(block.type)) {
// Remove signature field if exists
const { signature, cache_control, ...cleanBlock } = block;
filteredContent.push(cleanBlock);
filteredContent.push(stripBlock(block));
} else if (block.type === CLAUDE_BLOCK.TOOL_USE) {
// Convert tool_use to tool_calls format (handled separately)
continue;
} else if (block.type === CLAUDE_BLOCK.TOOL_RESULT) {
// Keep tool_result but clean it
const { signature, cache_control, ...cleanBlock } = block;
filteredContent.push(cleanBlock);
filteredContent.push(stripBlock(block));
}
}

4
open-sse/translator/index.js
View File

@@ -109,7 +109,9 @@ export function translateRequest(sourceFormat, targetFormat, model, body, stream
// Always normalize to clean OpenAI format when target is OpenAI
// This handles hybrid requests (e.g., OpenAI messages + Claude tools)
if (targetFormat === FORMATS.OPENAI) {
result = filterToOpenAIFormat(result);
result = filterToOpenAIFormat(result, {
preserveCacheControl: !!PROVIDERS[provider]?.quirks?.preserveCacheControl,
});
}
// Final step: prepare request for Claude format endpoints

22
open-sse/translator/request/antigravity-to-openai.js
View File

@@ -138,12 +138,14 @@ function convertContent(content) {
// Text with thoughtSignature = regular text after thinking
if (part.thoughtSignature && part.text !== undefined) {
textParts.push({ type: OPENAI_BLOCK.TEXT, text: part.text });
if (part.text) {
textParts.push({ type: OPENAI_BLOCK.TEXT, text: part.text });
}
continue;
}
// Regular text
if (part.text !== undefined) {
if (part.text !== undefined && part.text !== "") {
textParts.push({ type: OPENAI_BLOCK.TEXT, text: part.text });
}
@@ -180,8 +182,22 @@ function convertContent(content) {
}
}
// Content with only functionResponses → return array of tool messages
// Content with functionResponses — return array of tool result messages,
// plus an assistant message for any co-located tool calls / text.
if (toolResults.length > 0) {
if (toolCalls.length > 0 || textParts.length > 0 || reasoningContent) {
const assistantMsg = { role: ROLE.ASSISTANT };
if (textParts.length > 0) {
assistantMsg.content = collapseTextParts(textParts);
}
if (reasoningContent) {
assistantMsg.reasoning_content = reasoningContent;
}
if (toolCalls.length > 0) {
assistantMsg.tool_calls = toolCalls;
}
return [...toolResults, assistantMsg];
}
return toolResults;
}

7
open-sse/translator/request/claude-to-kiro.js
View File

@@ -393,9 +393,12 @@ export function claudeToKiroRequest(model, body, stream, credentials) {
reconcileOrphanedToolResults(history, currentMessage);
}
// API-key auth must never use the shared default ARN (403); OAuth/social fall back to it.
// api_key / idc / external_idp must never use the shared default ARN (belongs
// to another account → 403 "bearer token invalid"); OAuth/social fall back to it.
const authMethod = credentials?.providerSpecificData?.authMethod;
const profileArn = authMethod === "api_key"
const accountBoundAuth =
authMethod === "api_key" || authMethod === "idc" || authMethod === "external_idp";
const profileArn = accountBoundAuth
? (credentials?.providerSpecificData?.profileArn || "")
: (credentials?.providerSpecificData?.profileArn || resolveDefaultProfileArn(authMethod));

16
open-sse/translator/request/claude-to-openai.js
View File

@@ -129,8 +129,24 @@ function fixMissingToolResponsesOpenAI(messages) {
}
}
// Wrap mid-conversation system text so it ends as a user turn (avoids Anthropic prefill 400)
function systemReminderText(content) {
const parts = Array.isArray(content)
? content.filter(c => c?.type === CLAUDE_BLOCK.TEXT).map(c => c.text || "")
: [typeof content === "string" ? content : ""];
const text = parts.filter(Boolean).join("\n");
if (!text.trim()) return "";
return `<system-reminder>\n${text}\n</system-reminder>`;
}
// Convert single Claude message - returns single message or array of messages
function convertClaudeMessage(msg) {
// Mid-conversation system message -> user (per Anthropic placement rules)
if (msg.role === ROLE.SYSTEM) {
const text = systemReminderText(msg.content);
return text ? { role: ROLE.USER, content: text } : null;
}
const role = msg.role === ROLE.USER || msg.role === ROLE.TOOL ? ROLE.USER : ROLE.ASSISTANT;
// Simple string content

16
open-sse/translator/request/openai-to-gemini.js
View File

@@ -35,6 +35,17 @@ function sanitizeGeminiFunctionName(name) {
return sanitized.substring(0, 64);
}
function normalizeGeminiContents(contents) {
const out = [];
for (const c of contents || []) {
if (!c?.role || !Array.isArray(c.parts) || c.parts.length === 0) continue;
const last = out.at(-1);
if (last?.role === c.role) last.parts.push(...c.parts);
else out.push({ ...c, parts: [...c.parts] });
}
return out;
}
// Core: Convert OpenAI request to Gemini format (base for all variants)
function openaiToGeminiBase(model, body, stream, signature = DEFAULT_THINKING_AG_SIGNATURE) {
const result = {
@@ -217,6 +228,7 @@ function openaiToGeminiBase(model, body, stream, signature = DEFAULT_THINKING_AG
}
}
result.contents = normalizeGeminiContents(result.contents);
return result;
}
@@ -299,7 +311,7 @@ function wrapInCloudCodeEnvelope(model, geminiCLI, credentials = null, isAntigra
}
// Wrap Claude format in Cloud Code envelope for Antigravity
function wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials = null) {
function wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials = null, signature = DEFAULT_THINKING_AG_SIGNATURE) {
const projectId = credentials?.projectId || generateProjectId();
const envelope = {
@@ -343,6 +355,7 @@ function wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials = nu
parts.push({ text: block.text });
} else if (block.type === CLAUDE_BLOCK.TOOL_USE) {
parts.push({
thoughtSignature: signature,
functionCall: {
id: block.id,
name: sanitizeGeminiFunctionName(block.name),
@@ -425,6 +438,7 @@ function wrapInCloudCodeEnvelopeForClaude(model, claudeRequest, credentials = nu
envelope.request.systemInstruction = { role: GEMINI_ROLE.USER, parts: systemParts };
}
envelope.request.contents = normalizeGeminiContents(envelope.request.contents);
return envelope;
}

9
open-sse/translator/request/openai-to-kiro.js
View File

@@ -530,8 +530,15 @@ export function openaiToKiroRequest(model, body, stream, credentials) {
// (the ARN doesn't belong to the key's account). So for api_key, only send a
// profileArn that was actually resolved for this connection — never the default.
// OAuth/social keep the default fallback (their tokens accept it).
// api_key / idc / external_idp carry an account-specific (or token-bound)
// profile. The shared builder-id/social default ARN belongs to a different
// account and triggers 403 "bearer token invalid", so never fall back to it —
// send the resolved ARN, or an empty string so CodeWhisperer uses the token's
// own default profile. Only OAuth/social keep the shared placeholder.
const authMethod = credentials?.providerSpecificData?.authMethod;
const profileArn = authMethod === "api_key"
const accountBoundAuth =
authMethod === "api_key" || authMethod === "idc" || authMethod === "external_idp";
const profileArn = accountBoundAuth
? (credentials?.providerSpecificData?.profileArn || "")
: (credentials?.providerSpecificData?.profileArn || resolveDefaultProfileArn(authMethod));

40
open-sse/translator/response/claude-to-openai.js
View File

@@ -27,6 +27,25 @@ export function claudeToOpenAIResponse(chunk, state) {
state.messageId = chunk.message?.id || `msg_${Date.now()}`;
state.model = chunk.message?.model;
state.toolCallIndex = 0;
// Claude sends input_tokens + cache_read + cache_creation here; message_delta
// later carries only the final output_tokens. Capture cache now so the
// delta (output-only) doesn't reset it to zero.
const startUsage = chunk.message?.usage;
if (startUsage && typeof startUsage === "object") {
const inputTokens = typeof startUsage.input_tokens === "number" ? startUsage.input_tokens : 0;
const cacheReadTokens = typeof startUsage.cache_read_input_tokens === "number" ? startUsage.cache_read_input_tokens : 0;
const cacheCreationTokens = typeof startUsage.cache_creation_input_tokens === "number" ? startUsage.cache_creation_input_tokens : 0;
const promptTokens = inputTokens + cacheReadTokens + cacheCreationTokens;
state.usage = {
prompt_tokens: promptTokens,
completion_tokens: 0,
total_tokens: promptTokens,
input_tokens: inputTokens,
output_tokens: 0
};
if (cacheReadTokens > 0) state.usage.cache_read_input_tokens = cacheReadTokens;
if (cacheCreationTokens > 0) state.usage.cache_creation_input_tokens = cacheCreationTokens;
}
results.push(createChunk(state, { role: ROLE.ASSISTANT }));
break;
}
@@ -103,13 +122,15 @@ export function claudeToOpenAIResponse(chunk, state) {
}
case "message_delta": {
// Extract usage from message_delta event (Claude native format)
// Normalize to OpenAI format (prompt_tokens/completion_tokens) for consistent logging
// Extract usage from message_delta event (Claude native format).
// Anthropic sends input/cache in message_start and only output here, so
// fall back to cache captured in message_start when the delta omits it.
if (chunk.usage && typeof chunk.usage === "object") {
const inputTokens = typeof chunk.usage.input_tokens === "number" ? chunk.usage.input_tokens : 0;
const prev = state.usage || {};
const inputTokens = typeof chunk.usage.input_tokens === "number" ? chunk.usage.input_tokens : (prev.input_tokens || 0);
const outputTokens = typeof chunk.usage.output_tokens === "number" ? chunk.usage.output_tokens : 0;
const cacheReadTokens = typeof chunk.usage.cache_read_input_tokens === "number" ? chunk.usage.cache_read_input_tokens : 0;
const cacheCreationTokens = typeof chunk.usage.cache_creation_input_tokens === "number" ? chunk.usage.cache_creation_input_tokens : 0;
const cacheReadTokens = typeof chunk.usage.cache_read_input_tokens === "number" ? chunk.usage.cache_read_input_tokens : (prev.cache_read_input_tokens || 0);
const cacheCreationTokens = typeof chunk.usage.cache_creation_input_tokens === "number" ? chunk.usage.cache_creation_input_tokens : (prev.cache_creation_input_tokens || 0);
// prompt_tokens = input_tokens + cache_read + cache_creation (all prompt-side tokens)
const promptTokens = inputTokens + cacheReadTokens + cacheCreationTokens;
@@ -131,7 +152,14 @@ export function claudeToOpenAIResponse(chunk, state) {
const finalChunk = createChunk(state, {}, state.finishReason);
if (state.usage) {
finalChunk.usage = toOpenAIUsage(chunk.usage, "claude");
// Build OpenAI usage from the merged state (cache from message_start +
// output from message_delta), not the delta chunk alone.
finalChunk.usage = toOpenAIUsage({
input_tokens: state.usage.input_tokens || 0,
output_tokens: state.usage.output_tokens || 0,
cache_read_input_tokens: state.usage.cache_read_input_tokens,
cache_creation_input_tokens: state.usage.cache_creation_input_tokens
}, "claude");
}
results.push(finalChunk);

8
open-sse/translator/response/gemini-to-openai.js
View File

@@ -25,7 +25,10 @@ function emitFunctionCall(functionCall, state) {
type: OPENAI_BLOCK.FUNCTION,
function: { name: fcName, arguments: JSON.stringify(fcArgs) },
};
state.toolCalls.set(toolCallIndex, toolCall);
// Keep Gemini bookkeeping separate from the shared translator state.toolCalls map.
// The downstream OpenAI→Claude translator uses state.toolCalls for Claude block
// metadata; pre-populating it here makes Anthropic tool deltas lose index.
state.geminiToolCallCount = (state.geminiToolCallCount || 0) + 1;
return buildChunk(chunkMeta(state), { tool_calls: [toolCall] }, null);
}
@@ -46,6 +49,7 @@ export function geminiToOpenAIResponse(chunk, state) {
state.messageId = response.responseId || `msg_${Date.now()}`;
state.model = response.modelVersion || "gemini";
state.functionIndex = 0;
state.geminiToolCallCount = 0;
results.push(buildChunk(chunkMeta(state), { role: ROLE.ASSISTANT }, null));
}
@@ -117,7 +121,7 @@ export function geminiToOpenAIResponse(chunk, state) {
// Finish reason - include usage in final chunk
if (candidate.finishReason) {
let finishReason = toOpenAIFinish(candidate.finishReason, "gemini");
if (finishReason === OPENAI_FINISH.STOP && state.toolCalls.size > 0) {
if (finishReason === OPENAI_FINISH.STOP && state.geminiToolCallCount > 0) {
finishReason = OPENAI_FINISH.TOOL_CALLS;
}

3
open-sse/translator/response/openai-to-claude.js
View File

@@ -184,7 +184,8 @@ export function openaiToClaudeResponse(chunk, state) {
for (const tc of delta.tool_calls) {
const idx = tc.index ?? 0;
if (tc.id) {
// GLM/fireworks repeats id+null-name on every arg chunk; open block once per idx
if (tc.id && !state.toolCalls.has(idx)) {
stopThinkingBlock(state, results);
stopTextBlock(state, results);