zevairouter 1.0.0

package/open-sse/translator/request/openai-to-kiro.js

@@ -0,0 +1,629 @@
+/**
+ * OpenAI to Kiro Request Translator
+ * Converts OpenAI Chat Completions format to Kiro/AWS CodeWhisperer format
+ */
+import { register } from "../index.js";
+import { FORMATS } from "../formats.js";
+import { v4 as uuidv4 } from "uuid";
+import {
+  resolveKiroModel,
+  isThinkingEnabled,
+  buildThinkingSystemPrefix,
+  KIRO_AGENTIC_SYSTEM_PROMPT,
+  shouldExposeKiroReasoning
+} from "../../config/kiroConstants.js";
+
+const KIRO_TOOL_DESCRIPTION_MAX_CHARS = 4096;
+const KIRO_TOOL_DESCRIPTION_TIGHT_CHARS = 1024;
+const KIRO_TOOL_DESCRIPTION_MIN_CHARS = 256;
+const KIRO_TOOLS_JSON_BUDGET_CHARS = 256 * 1024;
+
+function truncateText(text, maxChars) {
+  if (typeof text !== "string" || text.length <= maxChars) return text;
+  return `${text.slice(0, maxChars)}\n\n[truncated for Kiro payload limit]`;
+}
+
+function normalizeToolDescription(description, name, maxChars = KIRO_TOOL_DESCRIPTION_MAX_CHARS) {
+  const raw = typeof description === "string" && description.trim()
+    ? description
+    : `Tool: ${name}`;
+  return truncateText(raw, maxChars);
+}
+
+function mapToolToKiroSpec(t) {
+  const name = t.function?.name || t.name;
+  const description = normalizeToolDescription(t.function?.description || t.description || "", name);
+  const schema = t.function?.parameters || t.parameters || t.input_schema || {};
+  // Normalize schema: Kiro requires required[] and proper type/properties
+  const normalizedSchema = Object.keys(schema).length === 0
+    ? { type: "object", properties: {}, required: [] }
+    : { ...schema, required: schema.required ?? [] };
+
+  return {
+    toolSpecification: {
+      name,
+      description,
+      inputSchema: { json: normalizedSchema }
+    }
+  };
+}
+
+function resizeToolDescriptions(tools, maxChars) {
+  return tools.map(tool => ({
+    ...tool,
+    toolSpecification: {
+      ...tool.toolSpecification,
+      description: normalizeToolDescription(tool.toolSpecification?.description || "", tool.toolSpecification?.name, maxChars)
+    }
+  }));
+}
+
+function fitKiroToolsToPayloadBudget(tools) {
+  if (JSON.stringify(tools).length <= KIRO_TOOLS_JSON_BUDGET_CHARS) return tools;
+
+  let resized = resizeToolDescriptions(tools, KIRO_TOOL_DESCRIPTION_TIGHT_CHARS);
+  if (JSON.stringify(resized).length <= KIRO_TOOLS_JSON_BUDGET_CHARS) return resized;
+
+  resized = resizeToolDescriptions(tools, KIRO_TOOL_DESCRIPTION_MIN_CHARS);
+  return resized;
+}
+
+/** Render a single tool call as a readable text line. */
+function toolCallToText(name, input) {
+  let argStr;
+  try {
+    argStr = typeof input === "string" ? input : JSON.stringify(input ?? {});
+  } catch {
+    argStr = "{}";
+  }
+  return `[Tool call: ${name || "unknown"}(${argStr})]`;
+}
+
+/** Render a tool result (string or content-block array) as a text line. */
+function toolResultToText(content) {
+  const text = Array.isArray(content)
+    ? content.map(c => (typeof c === "string" ? c : c.text || "")).join("\n")
+    : (typeof content === "string" ? content : "");
+  return `[Tool result: ${text}]`;
+}
+
+/**
+ * Flatten all tool calls/results in a conversation into plain text.
+ *
+ * Kiro's schema validator requires a non-empty
+ * currentMessage.userInputMessageContext.tools array whenever the history
+ * references any tool use; otherwise it returns "Improperly formed request"
+ * (HTTP 400). A client can hit this by omitting the `tools` array on a
+ * follow-up request — typically after client-side compaction (e.g. OpenCode).
+ *
+ * Rather than fabricate stub tool specs — which would advertise tool-calling
+ * capability the client never requested and may not handle, risking a phantom
+ * tool call on an otherwise plain turn — we collapse the tool interaction into
+ * text. The request stays honest, and since no structured tool content
+ * remains, the validator's "tools required" rule never fires.
+ *
+ * Only invoked when the client did NOT send tools; when tools are present the
+ * structured form is preserved.
+ */
+function flattenToolInteractions(messages) {
+  const out = [];
+
+  for (const msg of messages) {
+    // OpenAI tool-result message → user text line
+    if (msg.role === "tool") {
+      out.push({ role: "user", content: toolResultToText(msg.content) });
+      continue;
+    }
+
+    if (msg.role === "assistant") {
+      const parts = [];
+      if (Array.isArray(msg.content)) {
+        for (const c of msg.content) {
+          if (c.type === "tool_use") {
+            parts.push(toolCallToText(c.name, c.input));
+          } else if (c.type === "text" || c.text) {
+            parts.push(c.text || "");
+          }
+        }
+      } else if (typeof msg.content === "string") {
+        parts.push(msg.content);
+      }
+      for (const tc of msg.tool_calls || []) {
+        parts.push(toolCallToText(tc.function?.name, tc.function?.arguments));
+      }
+      out.push({ role: "assistant", content: parts.filter(Boolean).join("\n") });
+      continue;
+    }
+
+    // User messages: replace tool_result blocks with text, keep text + images.
+    if (msg.role === "user" && Array.isArray(msg.content)) {
+      const newContent = msg.content.map(c =>
+        c.type === "tool_result"
+          ? { type: "text", text: toolResultToText(c.content) }
+          : c
+      );
+      out.push({ ...msg, content: newContent });
+      continue;
+    }
+
+    out.push(msg);
+  }
+
+  return out;
+}
+
+/**
+ * Reconcile orphaned toolResults — those whose toolUseId has no matching
+ * toolUse in any assistant message. This happens when client-side compaction
+ * truncates the conversation and removes the assistant message containing the
+ * tool_use, but keeps the user message with the corresponding tool_result.
+ *
+ * A dangling structured reference makes Kiro return 400, so it must be removed.
+ * But the client deliberately kept the result content through compaction, so
+ * rather than discard it we fold it back into the user message as text — the
+ * same shape flattenToolInteractions() produces. The 400 trigger (the
+ * structured reference) is gone; the content survives.
+ *
+ * `messages` is every carrier that can hold toolResults — both history items
+ * and the popped-out currentMessage (orphans can land on either).
+ */
+function reconcileOrphanedToolResults(history, currentMessage) {
+  // Phase 1: collect all valid toolUseIds from assistant messages in history.
+  // (currentMessage is always a user turn, so it carries no toolUses.)
+  const validIds = new Set();
+  for (const h of history) {
+    const arm = h.assistantResponseMessage;
+    if (!arm) continue;
+    for (const tu of arm.toolUses || []) {
+      if (tu.toolUseId) validIds.add(tu.toolUseId);
+    }
+  }
+
+  // Phase 2: across history + currentMessage, keep results with a matching
+  // toolUse and salvage the rest as text.
+  const carriers = currentMessage ? [...history, currentMessage] : history;
+  for (const item of carriers) {
+    const uim = item.userInputMessage;
+    const ctx = uim?.userInputMessageContext;
+    if (!ctx?.toolResults?.length) continue;
+
+    const kept = [];
+    const salvaged = [];
+    for (const tr of ctx.toolResults) {
+      if (validIds.has(tr.toolUseId)) {
+        kept.push(tr);
+      } else {
+        salvaged.push(toolResultToText(tr.content));
+      }
+    }
+
+    if (salvaged.length === 0) continue; // no orphans — leave untouched
+
+    // Fold orphaned result content into the user text so it is not lost
+    const extra = salvaged.join("\n");
+    uim.content = uim.content ? `${uim.content}\n\n${extra}` : extra;
+
+    ctx.toolResults = kept;
+    if (kept.length === 0 && !ctx.tools?.length) {
+      delete uim.userInputMessageContext;
+    }
+  }
+}
+
+/**
+ * Safely parse JSON string, returning fallback on failure.
+ */
+function safeJSONParse(str, fallback) {
+  if (typeof str !== "string") return str ?? fallback;
+  try { return JSON.parse(str); } catch { return fallback; }
+}
+
+/**
+ * Convert OpenAI messages to Kiro format
+ * Rules: system/tool/user -> user role, merge consecutive same roles.
+ *
+ * Returns { history, currentMessage }.
+ */
+function convertMessages(messages, tools, model) {
+  let history = [];
+  let currentMessage = null;
+
+  const clientProvidedTools = tools && tools.length > 0;
+
+  // When the client did not send tools, flatten any tool calls/results in the
+  // history into plain text (see flattenToolInteractions). This keeps the
+  // request honest and sidesteps Kiro's "tools required" 400, since no
+  // structured tool content survives to trigger it.
+  if (!clientProvidedTools) {
+    messages = flattenToolInteractions(messages);
+  }
+
+  let pendingUserContent = [];
+  let pendingAssistantContent = [];
+  let pendingToolResults = [];
+  let pendingImages = [];
+  let currentRole = null;
+  let toolsInjectedToFirstUserMsg = false;
+
+  const flushPending = () => {
+    if (currentRole === "user") {
+      const content = pendingUserContent.join("\n\n").trim() || "continue";
+      const userMsg = {
+        userInputMessage: {
+          content: content,
+          modelId: ""
+        }
+      };
+
+      // Attach images if present (Kiro API supports images field)
+      if (pendingImages.length > 0) {
+        userMsg.userInputMessage.images = pendingImages;
+      }
+
+      if (pendingToolResults.length > 0) {
+        userMsg.userInputMessage.userInputMessageContext = {
+          toolResults: pendingToolResults
+        };
+      }
+
+      // Add tools to the user message that has no preceding assistant messages,
+      // OR the first user message (whichever comes first after any opening
+      // assistant messages). We track whether any user message has already
+      // received tools via a flag on the history array.
+      if (clientProvidedTools && !toolsInjectedToFirstUserMsg) {
+        if (!userMsg.userInputMessage.userInputMessageContext) {
+          userMsg.userInputMessage.userInputMessageContext = {};
+        }
+        userMsg.userInputMessage.userInputMessageContext.tools = fitKiroToolsToPayloadBudget(
+          tools.map(mapToolToKiroSpec)
+        );
+        toolsInjectedToFirstUserMsg = true;
+      }
+
+      history.push(userMsg);
+      currentMessage = userMsg;
+      pendingUserContent = [];
+      pendingToolResults = [];
+      pendingImages = [];
+    } else if (currentRole === "assistant") {
+      const content = pendingAssistantContent.join("\n\n").trim() || "...";
+      const assistantMsg = {
+        assistantResponseMessage: {
+          content: content
+        }
+      };
+      history.push(assistantMsg);
+      pendingAssistantContent = [];
+    }
+  };
+
+  for (let i = 0; i < messages.length; i++) {
+    const msg = messages[i];
+    let role = msg.role;
+
+    // Normalize: system/tool -> user
+    if (role === "system" || role === "tool") {
+      role = "user";
+    }
+
+    // If role changes, flush pending
+    if (role !== currentRole && currentRole !== null) {
+      flushPending();
+    }
+    currentRole = role;
+
+    if (role === "user") {
+      // Extract content
+      let content = "";
+      if (typeof msg.content === "string") {
+        content = msg.content;
+      } else if (Array.isArray(msg.content)) {
+        const textParts = [];
+        for (const c of msg.content) {
+          if (c.type === "text" || c.text) {
+            textParts.push(c.text || "");
+          } else if (c.type === "image_url") {
+            // OpenAI format: image_url.url with data URI
+            const url = c.image_url?.url || "";
+            const base64Match = url.match(/^data:([^;]+);base64,(.+)$/);
+            if (base64Match) {
+              const mediaType = base64Match[1];
+              const format = mediaType.split("/")[1] || mediaType;
+              pendingImages.push({ format, source: { bytes: base64Match[2] } });
+            } else if (url.startsWith("http://") || url.startsWith("https://")) {
+              // Kiro only supports base64 — fallback to URL text
+              textParts.push(`[Image: ${url}]`);
+            }
+          } else if (c.type === "image") {
+            // Claude format: source.type = "base64", source.media_type, source.data
+            if (c.source?.type === "base64" && c.source?.data) {
+              const mediaType = c.source.media_type || "image/png";
+              const format = mediaType.split("/")[1] || mediaType;
+              pendingImages.push({ format, source: { bytes: c.source.data } });
+            }
+          }
+        }
+        content = textParts.join("\n");
+
+        // Check for tool_result blocks
+        const toolResultBlocks = msg.content.filter(c => c.type === "tool_result");
+        if (toolResultBlocks.length > 0) {
+          toolResultBlocks.forEach(block => {
+            const text = Array.isArray(block.content)
+              ? block.content.map(c => c.text || "").join("\n")
+              : (typeof block.content === "string" ? block.content : "");
+
+            pendingToolResults.push({
+              toolUseId: block.tool_use_id,
+              status: "success",
+              content: [{ text: text }]
+            });
+          });
+        }
+      }
+
+      // Handle tool role (from normalized)
+      if (msg.role === "tool") {
+        const toolContent = typeof msg.content === "string" ? msg.content : "";
+        pendingToolResults.push({
+          toolUseId: msg.tool_call_id,
+          status: "success",
+          content: [{ text: toolContent }]
+        });
+      } else if (content) {
+        pendingUserContent.push(content);
+      }
+    } else if (role === "assistant") {
+      // Extract text content and tool uses
+      let textContent = "";
+      let toolUses = [];
+
+      if (Array.isArray(msg.content)) {
+        const textBlocks = msg.content.filter(c => c.type === "text");
+        textContent = textBlocks.map(b => b.text).join("\n").trim();
+
+        const toolUseBlocks = msg.content.filter(c => c.type === "tool_use");
+        toolUses = toolUseBlocks;
+      } else if (typeof msg.content === "string") {
+        textContent = msg.content.trim();
+      }
+
+      if (msg.tool_calls && msg.tool_calls.length > 0) {
+        toolUses = msg.tool_calls;
+      }
+
+      if (textContent) {
+        pendingAssistantContent.push(textContent);
+      }
+
+      // Store tool uses in last assistant message
+      if (toolUses.length > 0) {
+        // Flush to create assistant message with toolUses
+        flushPending();
+
+        const lastMsg = history[history.length - 1];
+        if (lastMsg?.assistantResponseMessage) {
+          lastMsg.assistantResponseMessage.toolUses = toolUses.map(tc => {
+            if (tc.function) {
+              return {
+                toolUseId: tc.id || uuidv4(),
+                name: tc.function.name,
+                input: safeJSONParse(tc.function.arguments, {})
+              };
+            } else {
+              return {
+                toolUseId: tc.id || uuidv4(),
+                name: tc.name,
+                input: tc.input || {}
+              };
+            }
+          });
+        }
+
+        currentRole = null;
+      }
+    }
+  }
+
+  // Flush remaining
+  if (currentRole !== null) {
+    flushPending();
+  }
+
+  // Pop last userInputMessage as currentMessage (search from end, skip trailing assistant messages)
+  for (let i = history.length - 1; i >= 0; i--) {
+    if (history[i].userInputMessage) {
+      currentMessage = history.splice(i, 1)[0];
+      break;
+    }
+  }
+
+  // Grab tools from first history item BEFORE cleanup removes them
+  const firstHistoryTools = history[0]?.userInputMessage?.userInputMessageContext?.tools;
+
+  // Clean up history for Kiro API compatibility
+  history.forEach(item => {
+    if (item.userInputMessage?.userInputMessageContext?.tools) {
+      delete item.userInputMessage.userInputMessageContext.tools;
+    }
+    if (item.userInputMessage?.userInputMessageContext &&
+        Object.keys(item.userInputMessage.userInputMessageContext).length === 0) {
+      delete item.userInputMessage.userInputMessageContext;
+    }
+    if (item.userInputMessage && !item.userInputMessage.modelId) {
+      item.userInputMessage.modelId = model;
+    }
+  });
+
+  // Merge consecutive user messages (Kiro requires alternating user/assistant)
+  // When merging, also combine userInputMessageContext fields so toolResults
+  // and images from the second message are not silently dropped.
+  const mergedHistory = [];
+  for (let i = 0; i < history.length; i++) {
+    const current = history[i];
+    if (current.userInputMessage &&
+        mergedHistory.length > 0 &&
+        mergedHistory[mergedHistory.length - 1].userInputMessage) {
+      const prev = mergedHistory[mergedHistory.length - 1];
+      prev.userInputMessage.content += "\n\n" + current.userInputMessage.content;
+      // Merge context: combine toolResults, images, etc.
+      const prevCtx = prev.userInputMessage.userInputMessageContext;
+      const curCtx = current.userInputMessage.userInputMessageContext;
+      if (curCtx) {
+        if (!prevCtx) {
+          prev.userInputMessage.userInputMessageContext = curCtx;
+        } else {
+          if (curCtx.toolResults?.length > 0) {
+            prevCtx.toolResults = [...(prevCtx.toolResults || []), ...curCtx.toolResults];
+          }
+          if (curCtx.tools?.length > 0) {
+            prevCtx.tools = [...(prevCtx.tools || []), ...curCtx.tools];
+          }
+        }
+      }
+    } else {
+      mergedHistory.push(current);
+    }
+  }
+
+  // When currentMessage is null (no user messages at all — edge case where
+  // input is only assistant messages), create a minimal currentMessage so
+  // tools and content can be injected.
+  if (!currentMessage) {
+    currentMessage = {
+      userInputMessage: {
+        content: "",
+        modelId: model,
+      }
+    };
+  }
+
+  // Reconcile orphaned toolResults across history AND currentMessage — when
+  // client-side compaction removes assistant messages containing tool_use but
+  // keeps the tool_result, the dangling reference triggers a Kiro 400. Fold the
+  // content back into the user text instead of discarding it. Run after
+  // currentMessage is finalized (an orphan can be merged into it) and before
+  // tool injection (which may re-add userInputMessageContext).
+  //
+  // Only needed on the tools-present path: when the client sent no tools,
+  // flattenToolInteractions already collapsed every toolResult to text, so
+  // there is nothing structured left to orphan.
+  if (clientProvidedTools) {
+    reconcileOrphanedToolResults(mergedHistory, currentMessage);
+  }
+
+  // Inject tools into currentMessage AFTER cleanup. Tools only exist here when
+  // the client explicitly sent them (otherwise flattenToolInteractions already
+  // collapsed all tool content to text upstream, so there is nothing to carry).
+  const resolvedTools = firstHistoryTools;
+
+  if (resolvedTools?.length > 0 &&
+      !currentMessage.userInputMessage.userInputMessageContext?.tools) {
+    if (!currentMessage.userInputMessage.userInputMessageContext) {
+      currentMessage.userInputMessage.userInputMessageContext = {};
+    }
+    currentMessage.userInputMessage.userInputMessageContext.tools = resolvedTools;
+  }
+
+  return { history: mergedHistory, currentMessage };
+}
+
+/**
+ * Build Kiro payload from OpenAI format
+ *
+ * Two 9router-specific behaviours implemented here:
+ *
+ * 1. `-agentic` model suffix. Synthetic variant — same upstream model, but we
+ *    inject a chunked-write system prompt to keep large file writes under
+ *    Kiro's 2-3 minute server timeout. The suffix is stripped before being
+ *    sent upstream.
+ *
+ * 2. Thinking / reasoning. Kiro does not accept `thinking.type` or
+ *    `reasoning_effort` natively. The only way to enable reasoning is to
+ *    inject `<thinking_mode>enabled</thinking_mode>` into the user content
+ *    sent upstream. Detection covers Anthropic-Beta header, Claude API
+ *    `thinking`, OpenAI `reasoning_effort`, AMP/Cursor magic tags, and model
+ *    name hints.
+ */
+export function buildKiroPayload(model, body, stream, credentials) {
+  const messages = body.messages || [];
+  const tools = body.tools || [];
+  const maxTokens = 32000;
+  const temperature = body.temperature;
+  const topP = body.top_p;
+
+  const { upstream: upstreamModel, agentic } = resolveKiroModel(model);
+  const thinkingEnabled = isThinkingEnabled(body, null, model);
+  const exposeReasoning = shouldExposeKiroReasoning(body);
+
+  const { history, currentMessage } = convertMessages(messages, tools, upstreamModel);
+
+  const profileArn = credentials?.providerSpecificData?.profileArn || "";
+
+  let finalContent = currentMessage?.userInputMessage?.content || "";
+
+  const timestamp = new Date().toISOString();
+
+  // Build the system-prompt prefix that goes ABOVE the user message body.
+  // Order: thinking_mode tag first (so Kiro sees it before any user text),
+  // then context/timestamp marker, then optional agentic chunked-write prompt.
+  const prefixParts = [];
+  if (thinkingEnabled) {
+    prefixParts.push(buildThinkingSystemPrefix());
+  }
+  prefixParts.push(`[Context: Current time is ${timestamp}]`);
+  if (agentic) {
+    prefixParts.push(KIRO_AGENTIC_SYSTEM_PROMPT);
+  }
+  finalContent = `${prefixParts.join("\n\n")}\n\n${finalContent}`;
+
+  const payload = {
+    conversationState: {
+      chatTriggerType: "MANUAL",
+      conversationId: uuidv4(),
+      currentMessage: {
+        userInputMessage: {
+          content: finalContent,
+          modelId: upstreamModel,
+          origin: "AI_EDITOR",
+          ...(currentMessage?.userInputMessage?.images?.length > 0 && {
+            images: currentMessage.userInputMessage.images
+          }),
+          ...(currentMessage?.userInputMessage?.userInputMessageContext && {
+            userInputMessageContext: currentMessage.userInputMessage.userInputMessageContext
+          })
+        }
+      },
+      history: history
+    }
+  };
+
+  if (profileArn) {
+    payload.profileArn = profileArn;
+  }
+
+  if (maxTokens || temperature !== undefined || topP !== undefined) {
+    payload.inferenceConfig = {};
+    if (maxTokens) payload.inferenceConfig.maxTokens = maxTokens;
+    if (temperature !== undefined) payload.inferenceConfig.temperature = temperature;
+    if (topP !== undefined) payload.inferenceConfig.topP = topP;
+  }
+
+  // Tag payload so the executor can route the upstream model id correctly.
+  Object.defineProperty(payload, "_kiroUpstreamModel", {
+    value: upstreamModel,
+    enumerable: false
+  });
+  Object.defineProperty(payload, "_kiroThinkingEnabled", {
+    value: thinkingEnabled,
+    enumerable: false
+  });
+  Object.defineProperty(payload, "_kiroExposeReasoning", {
+    value: exposeReasoning,
+    enumerable: false
+  });
+
+  return payload;
+}
+
+register(FORMATS.OPENAI, FORMATS.KIRO, buildKiroPayload, null);