@wenathlan/devthink 2.0.13 → 2.0.15

package/streaming.ts

@@ -0,0 +1,4101 @@
+/**
+ * @file streaming.ts
+ * @module provider/streaming
+ * @description
+ *  Third-person observer view: the library watches Antigravity / CloudCode
+ *  traffic and reproduces its streaming behavior without running Antigravity
+ *  itself.
+ *
+ *  Responsibilities isolated here:
+ *   - Incremental SSE parser tolerant to network splits (parseSSEChunk)
+ *   - Heterogeneous payload normalization (unwrap response/candidates/parts)
+ *   - LRU cache for thinking signatures — 100 entries — Map-based
+ *   - Preservation of thinking blocks between chunks (delta tracking)
+ *   - OpenAI / Anthropic transformers compatible with opencode plugin SDK
+ *   - Auto-recovery exponential backoff + jitter 0-80ms
+ *   - Synthetic finish injection when cloudcode-pa closes abruptly
+ *
+ *  Architectural constraints:
+ *   - Library-first, root-first (no src/ folder)
+ *   - Direct cloudcode-pa.googleapis.com — no localhost v1 proxy
+ *   - Only node:* builtins + global fetch (Node >=18)
+ *     -> node:crypto for FNV-1a, signature hashing, jitter
+ *     -> global TextDecoder / ReadableStream (web standards in Node)
+ *   - Production-ready: never throws in parser, defensive JSON tolerant,
+ *     JSDoc everywhere, explicit error handling, no placeholder
+ *
+ *  Date governance: 25/08/2026 — Canto do Buriti, PI, BR
+ *  Version: 2.0.0 — the merged provider lineage
+ *
+ *  Related modules:
+ *   - constants.ts — endpoints, LRU_THINKING_CACHE_SIZE, models 2026
+ *   - fingerprint.ts — FNV-1a reused here for deterministic ids
+ *   - request.ts — builds Antigravity request body
+ *   - recovery.ts — higher level auto-recovery using this module
+ *
+ * @author devthink
+ * @license MIT
+ */
+
+import { createHash, randomUUID } from "node:crypto";
+import * as os from "node:os";
+
+// ---------------------------------------------------------------------------
+// Frozen-owner imports — identical values owned by constants.ts / core.ts
+// ---------------------------------------------------------------------------
+
+import {
+  BACKOFF_BASE_MS,
+  BACKOFF_FACTOR,
+  BACKOFF_MAX_MS,
+  GEMINI_CLI_OAUTH_CLIENT_ID as GEMINI_CLI_CLIENT_ID,
+  GEMINI_CLI_OAUTH_CLIENT_SECRET as GEMINI_CLI_CLIENT_SECRET,
+  JITTER_MAX_MS,
+  JITTER_MIN_MS,
+  LRU_THINKING_CACHE_SIZE,
+  MAX_RETRIES,
+  PROJECT_FALLBACK,
+  GEMINI_CLI_USER_AGENT,
+  CLIENT_METADATA_RAW,
+  CLOUDCODE_PA_BASE,
+  RETRYABLE_STATUS_CODES,
+} from "./constants.js";
+import { fnv1a32, getJitter, thinkingCache } from "./core.js";
+import { X_GOOG_API_CLIENT_GEMINI_CLI as X_GOOG_API_CLIENT } from "./fingerprint.js";
+
+// ---------------------------------------------------------------------------
+// Constants — frozen, deterministic
+// ---------------------------------------------------------------------------
+
+// Values below are re-exported from the frozen owners so each symbol keeps a
+// single definition; the module surface is unchanged for consumers.
+export {
+  BACKOFF_BASE_MS,
+  BACKOFF_FACTOR,
+  BACKOFF_MAX_MS,
+  GEMINI_CLI_CLIENT_ID,
+  GEMINI_CLI_CLIENT_SECRET,
+  JITTER_MAX_MS,
+  JITTER_MIN_MS,
+  LRU_THINKING_CACHE_SIZE,
+  MAX_RETRIES,
+  PROJECT_FALLBACK,
+  fnv1a32,
+  getJitter,
+  thinkingCache,
+};
+
+// local variant: diverges from constants (owner list is [403,404,408,429,500,502,503,504]; this streaming subset omits 403/404/408)
+/** Retryable HTTP status codes for cloudcode-pa */
+
+/** Sentinel that Gemini / Antigravity sends to close stream */
+export const SSE_DONE_SENTINEL = "[DONE]" as const;
+
+// ---------------------------------------------------------------------------
+// Helpers — FNV-1a 32-bit (deterministic, zero external dep)
+// ---------------------------------------------------------------------------
+
+// fnv1a32 is imported from core.js (identical charCodeAt + Math.imul algorithm)
+// and re-exported above.
+
+/**
+ * SHA-256 hex truncated to 16 chars — stable key for thinking text
+ * when length > few hundred chars. Falls back to FNV-1a if crypto fails.
+ *
+ * @param text - thinking text to hash
+ */
+export function hashThinkingText(text: string): string {
+  try {
+    return createHash("sha256").update(text, "utf8").digest("hex").slice(0, 32);
+  } catch {
+    return fnv1a32(text).toString(16).padStart(8, "0");
+  }
+}
+
+// getJitter is imported from constants.js (identical randomInt 0-80 body)
+// and re-exported above.
+
+/**
+ * Promise-based sleep with jitter included for backoff
+ * @param ms - base delay
+ */
+export function sleepWithJitter(ms: number): Promise<void> {
+  const jitter = getJitter();
+  const total = Math.max(0, ms + jitter);
+  return new Promise((resolve) => setTimeout(resolve, total));
+}
+
+// ---------------------------------------------------------------------------
+// LRU Cache — thinking signatures 100 entradas
+// ---------------------------------------------------------------------------
+
+/**
+ * @description
+ *  The observer notes that Claude thinking blocks carry `thoughtSignature`
+ *  that must be preserved across turns, otherwise API returns
+ *  "invalid signature". Original plugin stored 100 entries in Map.
+ *  LRU eviction keeps most recent.
+ *
+ *  This implementation is library-first: no external LRU dep,
+ *  uses Map insertion order (ES2015 guarantees order).
+ */
+// local variant: diverges from recovery (recovery.ReplaySignatureCache is the entry-based cache with signature keys and provenance; this is the hash->signature string map)
+export class ReplaySignatureCache {
+  private readonly maxSize: number;
+  private readonly map: Map<string, string>;
+
+  /**
+   * @param maxSize - max entries, default 100 per spec
+   */
+  constructor(maxSize: number = LRU_THINKING_CACHE_SIZE) {
+    if (!Number.isInteger(maxSize) || maxSize <= 0) {
+      throw new TypeError(`maxSize must be positive integer, got ${maxSize}`);
+    }
+    this.maxSize = maxSize;
+    this.map = new Map<string, string>();
+  }
+
+  /** Current size */
+  get size(): number {
+    return this.map.size;
+  }
+
+  /** Max configured */
+  get capacity(): number {
+    return this.maxSize;
+  }
+
+  /**
+   * Retrieves signature for given thinking text.
+   * The cache key is hash of normalized text (trimmed).
+   *
+   * @param thinkingText - full thinking block text
+   * @returns signature or undefined
+   */
+  get(thinkingText: string): string | undefined {
+    if (!thinkingText || typeof thinkingText !== "string") return undefined;
+    const key = hashThinkingText(thinkingText.trim());
+    const val = this.map.get(key);
+    if (val !== undefined) {
+      // LRU touch: re-insert to move to end (most recent)
+      this.map.delete(key);
+      this.map.set(key, val);
+    }
+    return val;
+  }
+
+  /**
+   * Retrieves signature by direct hash key (when caller already hashed)
+   */
+  getByHash(hashKey: string): string | undefined {
+    const val = this.map.get(hashKey);
+    if (val !== undefined) {
+      this.map.delete(hashKey);
+      this.map.set(hashKey, val);
+    }
+    return val;
+  }
+
+  /**
+   * Stores signature keyed by thinking text
+   * @param thinkingText - full text
+   * @param signature - opaque signature string from API
+   */
+  set(thinkingText: string, signature: string): void {
+    if (!thinkingText || typeof thinkingText !== "string") return;
+    if (!signature || typeof signature !== "string") return;
+    const key = hashThinkingText(thinkingText.trim());
+    if (this.map.has(key)) this.map.delete(key);
+    this.map.set(key, signature);
+    this.evictIfNeeded();
+  }
+
+  /**
+   * Direct hash->signature set (when text not available)
+   */
+  setByHash(hashKey: string, signature: string): void {
+    if (!hashKey || !signature) return;
+    if (this.map.has(hashKey)) this.map.delete(hashKey);
+    this.map.set(hashKey, signature);
+    this.evictIfNeeded();
+  }
+
+  /** Checks presence without touching LRU order */
+  has(thinkingText: string): boolean {
+    if (!thinkingText) return false;
+    const key = hashThinkingText(thinkingText.trim());
+    return this.map.has(key);
+  }
+
+  /** Deletes entry by thinking text */
+  delete(thinkingText: string): boolean {
+    if (!thinkingText) return false;
+    const key = hashThinkingText(thinkingText.trim());
+    return this.map.delete(key);
+  }
+
+  /** Clears all */
+  clear(): void {
+    this.map.clear();
+  }
+
+  /** Iterates entries LRU order (oldest first) */
+  entries(): IterableIterator<[string, string]> {
+    return this.map.entries();
+  }
+
+  /** Keys in LRU order */
+  keys(): IterableIterator<string> {
+    return this.map.keys();
+  }
+
+  /** Returns diagnostic snapshot oldest->newest */
+  snapshot(): Array<{ hash: string; signaturePreview: string }> {
+    return Array.from(this.map.entries()).map(([h, s]) => ({
+      hash: h,
+      signaturePreview: s.slice(0, 24) + (s.length > 24 ? "…" : ""),
+    }));
+  }
+
+  private evictIfNeeded(): void {
+    while (this.map.size > this.maxSize) {
+      const oldestKey = this.map.keys().next().value;
+      if (oldestKey === undefined) break;
+      this.map.delete(oldestKey);
+    }
+  }
+}
+
+/** Global singleton used by transformers — library-first convenience */
+export const globalThinkingSignatureCache = new ReplaySignatureCache(LRU_THINKING_CACHE_SIZE);
+
+// ---------------------------------------------------------------------------
+// SSE Incremental Parser — tolerante splits
+// ---------------------------------------------------------------------------
+
+/**
+ * Raw SSE event parsed from wire.
+ * Observer keeps it minimal — only what cloudcode-pa uses.
+ */
+export interface ParsedSSEEvent {
+  /** Event type if server sent `event: xxx` */
+  event?: string | undefined;
+  /** Optional id `id: xxx` */
+  id?: string | undefined;
+  /** Retry interval `retry: xxx` */
+  retry?: number | undefined;
+  /** Joined data payload (may be JSON string, may be [DONE]) */
+  data: string;
+  /** Raw block for debugging */
+  rawBlock: string;
+}
+
+/**
+ * Mutable buffer holding incomplete SSE tail between network chunks.
+ * Third-person: the observer accumulates until double newline terminator.
+ */
+export interface SSEParserState {
+  /** Leftover incomplete tail */
+  buffer: string;
+  /** How many events parsed so far — useful for diagnostics */
+  eventsSeen: number;
+}
+
+/**
+ * Creates fresh parser state.
+ */
+export function createSSEParserState(): SSEParserState {
+  return { buffer: "", eventsSeen: 0 };
+}
+
+/**
+ * Parses one text chunk incrementally, tolerant to splits.
+ *
+ * The observer learned from production: fetch streaming may split in middle
+ * of `data: {` JSON, or split the `\n\n` delimiter across chunks.
+ * Strategy:
+ *  1. Append chunk to state.buffer
+ *  2. Normalize \r\n -> \n, \r -> \n
+ *  3. While buffer contains `\n\n` extract block
+ *  4. For each block, extract data: lines (join with \n per SSE spec)
+ *  5. Leave incomplete tail in buffer
+ *
+ * @param chunk - new text arrived from decoder (may be partial)
+ * @param state - mutable buffer object — will be updated in place
+ * @returns parsed events from this chunk only
+ *
+ * @example
+ *  const state = createSSEParserState();
+ *  let ev = parseSSEChunk("data: {\"cand", state); // [] — buffered
+ *  ev = parseSSEChunk("idates\":[]}\n\n", state); // 1 event
+ */
+// local variant: diverges from core (stateful incremental parser with buffer/event/id/retry fields; core.parseSSEChunk is stateless and returns only joined data payloads)
+export function parseSSEChunk(chunk: string, state: SSEParserState): ParsedSSEEvent[] {
+  if (typeof chunk !== "string") {
+    throw new TypeError(`parseSSEChunk chunk must be string, got ${typeof chunk}`);
+  }
+  if (!state || typeof state.buffer !== "string") {
+    throw new TypeError("parseSSEChunk state must have .buffer string");
+  }
+
+  // Defensive: if chunk empty, nothing to do
+  if (chunk.length === 0) return [];
+
+  // Append and normalize line endings eagerly — \r\n and \r both become \n
+  // This keeps split tolerance for \r\n across chunks (e.g. \r in previous, \n in current)
+  state.buffer += chunk;
+  // We cannot replace globally before we have full delimiter? We do incremental replace
+  // but keep logic simple: replace \r\n first then \r
+  // Important: don't destroy partial \r at end — if buffer ends with \r, keep it and wait for \n
+  // So we only replace \r\n that are complete and standalone \r not at end.
+  // Implementation: temporary replace \r\n -> \n, then replace remaining \r \n? Let's be safe.
+  let buf = state.buffer.replace(/\r\n/g, "\n");
+  // If last char is \r (could be split \r\n), keep it pending — but we already replaced \r\n, so lonely \r stays
+  // Replace lonely \r not followed by anything? We convert them to \n except when it's last char and previous chunk may have split?
+  // For simplicity we convert all remaining \r to \n — worst case we may prematurely split, but still valid SSE allows \r as line terminator.
+  buf = buf.replace(/\r/g, "\n");
+  state.buffer = buf;
+
+  const events: ParsedSSEEvent[] = [];
+  let searchStart = 0;
+
+  while (true) {
+    const delimiterIdx = state.buffer.indexOf("\n\n", searchStart);
+    if (delimiterIdx === -1) break; // incomplete block — wait for more chunks
+
+    const block = state.buffer.slice(searchStart, delimiterIdx);
+    // Move window past delimiter
+    const nextStart = delimiterIdx + 2;
+    // Parse block lines
+    const lines = block.split("\n");
+    let dataLines: string[] = [];
+    let eventType: string | undefined;
+    let id: string | undefined;
+    let retry: number | undefined;
+
+    for (const rawLine of lines) {
+      if (rawLine.length === 0) continue;
+      if (rawLine.startsWith(":")) {
+        // SSE comment / keepalive — ignore per spec
+        continue;
+      }
+      // According to SSE spec, field is up to first colon
+      const colonIdx = rawLine.indexOf(":");
+      let field: string;
+      let value: string;
+      if (colonIdx === -1) {
+        field = rawLine.trim();
+        value = "";
+      } else {
+        field = rawLine.slice(0, colonIdx).trim();
+        value = rawLine.slice(colonIdx + 1);
+        // If value starts with single space, strip it (spec)
+        if (value.startsWith(" ")) value = value.slice(1);
+      }
+
+      switch (field) {
+        case "data":
+          dataLines.push(value);
+          break;
+        case "event":
+          eventType = value;
+          break;
+        case "id":
+          id = value;
+          break;
+        case "retry":
+          {
+            const n = Number.parseInt(value, 10);
+            if (Number.isFinite(n)) retry = n;
+          }
+          break;
+        default:
+          // Unknown field — spec says ignore
+          break;
+      }
+    }
+
+    // Per spec, if dataLines empty, no event dispatch
+    if (dataLines.length > 0) {
+      const data = dataLines.join("\n"); // multiple data: lines concatenated with \n
+      events.push({
+        event: eventType,
+        id,
+        retry,
+        data,
+        rawBlock: block,
+      });
+      state.eventsSeen++;
+    }
+
+    // Slice buffer: remove processed portion including delimiter
+    // Since we search from searchStart, we need to handle correctly
+    // Simplify: we maintain state.buffer as remaining after nextStart
+    state.buffer = state.buffer.slice(nextStart);
+    searchStart = 0; // reset because buffer changed
+
+    // Safety: avoid infinite loop on huge buffers without delimiter
+    if (state.buffer.length > 10 * 1024 * 1024) {
+      // 10MB without delimiter is malformed — truncate to avoid OOM
+      // Observer logs warning but continues
+      console.warn("[streaming] SSE buffer exceeded 10MB without delimiter, truncating");
+      state.buffer = state.buffer.slice(-1024 * 1024); // keep last 1MB
+      break;
+    }
+  }
+
+  return events;
+}
+
+/**
+ * Parses a Uint8Array / Buffer chunk using TextDecoder streaming.
+ * Convenience wrapper around parseSSEChunk for binary fetch streams.
+ *
+ * @param binaryChunk - Uint8Array from reader.read()
+ * @param decoder - TextDecoder with {stream:true}
+ * @param state - parser state
+ */
+export function parseSSEBinaryChunk(
+  binaryChunk: Uint8Array,
+  decoder: { decode(input: Uint8Array, options?: { stream?: boolean }): string },
+  state: SSEParserState,
+): ParsedSSEEvent[] {
+  if (!(binaryChunk instanceof Uint8Array)) {
+    throw new TypeError("binaryChunk must be Uint8Array");
+  }
+  if (!(decoder instanceof TextDecoder)) {
+    throw new TypeError("decoder must be TextDecoder");
+  }
+  const text = decoder.decode(binaryChunk, { stream: true });
+  return parseSSEChunk(text, state);
+}
+
+// ---------------------------------------------------------------------------
+// Payload normalization — heterogêneos
+// ---------------------------------------------------------------------------
+
+/**
+ * Normalized tool call extracted from Antigravity part
+ */
+export interface NormalizedToolCall {
+  /** Deterministic id if not provided by model */
+  id: string;
+  /** Function / tool name sanitized */
+  name: string;
+  /** Parsed args object if JSON parsable, otherwise raw string */
+  args: unknown;
+  /** Raw JSON string of args — for OpenAI delta streaming */
+  argsJson: string;
+  /** Whether this is delta (partial args) or final */
+  isDelta?: boolean;
+}
+
+/**
+ * Single normalized chunk produced from one or more raw payloads
+ */
+export interface NormalizedChunk {
+  /** Text delta to emit (incremental already deduped) */
+  text?: string | undefined;
+  /** Full accumulated text for tracking */
+  fullText?: string | undefined;
+  /** Thinking delta */
+  thinking?: string | undefined;
+  /** Full accumulated thinking */
+  fullThinking?: string | undefined;
+  /** Signature for thinking block if present */
+  thinkingSignature?: string | undefined;
+  /** Tool calls extracted */
+  toolCalls?: NormalizedToolCall[] | undefined;
+  /** Finish reason from cloudcode-pa, normalized to lowercase (stop, max_tokens, safety) */
+  finishReason?: string | undefined;
+  /** Raw original payload for debugging / grounding */
+  raw: unknown;
+  /** Candidate index if payload contained multiple candidates */
+  candidateIndex?: number | undefined;
+  /** Whether this part was marked as thought */
+  isThought?: boolean | undefined;
+  /** Grounding metadata e.g. groundingMetadata, groundingChunks */
+  groundingMetadata?: unknown | undefined;
+  /** Safety ratings if present */
+  safetyRatings?: unknown | undefined;
+  /** Usage metadata if present */
+  usageMetadata?: unknown | undefined;
+  /** Indicates stream is done (from [DONE] or finishReason) */
+  isFinished?: boolean | undefined;
+}
+
+/**
+ * Safe JSON parse — never throws, returns null on failure
+ * @param text - raw data string
+ */
+export function safeJsonParse(text: string): unknown | null {
+  if (!text || typeof text !== "string") return null;
+  const trimmed = text.trim();
+  if (trimmed.length === 0) return null;
+  if (trimmed === SSE_DONE_SENTINEL) return { __sentinelDone: true };
+  try {
+    return JSON.parse(trimmed);
+  } catch {
+    // Tolerant attempt: sometimes cloudcode-pa sends multiple JSON concatenated with newline
+    // Try first line
+    try {
+      const firstLine = trimmed.split("\n")[0];
+      if (firstLine && firstLine !== trimmed) return JSON.parse(firstLine);
+    } catch {
+      // ignore
+    }
+    return null;
+  }
+}
+
+/**
+ * Unwraps heterogeneous envelope into candidate array.
+ * Handles observed shapes:
+ *  - {response:{candidates:[...]}}
+ *  - {candidates:[...]}
+ *  - {candidate:{...}}
+ *  - [{...}, {...}] (already array)
+ *  - {content:{parts:[...]}} single candidate
+ *  - {parts:[...]} direct parts
+ *
+ * @param payload - parsed JSON object
+ * @returns array of candidate-like objects
+ */
+export function unwrapCandidates(payload: unknown): unknown[] {
+  if (!payload || typeof payload !== "object") return [];
+
+  // Sentinel DONE already handled elsewhere
+  const obj = payload as any;
+
+  if (obj.__sentinelDone) return [];
+
+  // Array directly = list of candidates or payloads
+  if (Array.isArray(obj)) {
+    return obj;
+  }
+
+  // response.candidates
+  if (obj.response) {
+    if (Array.isArray(obj.response.candidates)) return obj.response.candidates;
+    if (obj.response.candidate) return [obj.response.candidate];
+    if (obj.response.content || obj.response.parts) return [obj.response];
+  }
+
+  // candidates
+  if (Array.isArray(obj.candidates)) return obj.candidates;
+
+  // single candidate field
+  if (obj.candidate && typeof obj.candidate === "object") return [obj.candidate];
+
+  // content or parts at top level considered single candidate
+  if (obj.content || obj.parts || obj.text || obj.thoughtSignature) return [obj];
+
+  // Fallback: treat whole object as candidate
+  return [obj];
+}
+
+/**
+ * Extracts parts array from candidate
+ */
+export function extractParts(candidate: unknown): unknown[] {
+  if (!candidate || typeof candidate !== "object") return [];
+  const c = candidate as any;
+
+  // Direct parts
+  if (Array.isArray(c.parts)) return c.parts;
+  // content.parts
+  if (c.content && Array.isArray(c.content.parts)) return c.content.parts;
+  // content array?
+  if (Array.isArray(c.content)) return c.content;
+  // text at candidate level as single part
+  if (typeof c.text === "string") {
+    return [{ text: c.text, thought: c.thought, thoughtSignature: c.thoughtSignature }];
+  }
+
+  // Gemini sometimes: candidate.content.parts empty but candidate itself is part
+  if (c.functionCall || c.function_call) return [c];
+
+  return [];
+}
+
+/**
+ * Normalizes single part object into text/thinking/toolCall
+ * @param part - raw part
+ */
+export function normalizePart(part: unknown): {
+  text?: string;
+  thinking?: string;
+  thinkingSignature?: string;
+  toolCall?: NormalizedToolCall;
+  isThought?: boolean;
+} {
+  if (!part || typeof part !== "object") {
+    if (typeof part === "string") return { text: part };
+    return {};
+  }
+  const p = part as any;
+
+  // Text + thought marker
+  // Antigravity observed: {text:"xxx", thought:true, thoughtSignature:"..."}
+  if (typeof p.text === "string") {
+    if (p.thought === true || p.isThought === true || p.thoughtSignature || p.type === "thinking") {
+      return {
+        thinking: p.text,
+        thinkingSignature: p.thoughtSignature || p.signature,
+        isThought: true,
+      };
+    }
+    return { text: p.text, isThought: false };
+  }
+
+  // Thinking explicitly: {thinking:"...", signature:"..."}
+  if (typeof p.thinking === "string") {
+    return {
+      thinking: p.thinking,
+      thinkingSignature: p.thoughtSignature || p.signature,
+      isThought: true,
+    };
+  }
+
+  // Function call shapes
+  // Gemini: {functionCall:{name, args}}
+  // Anthropic style: {tool_use:{id,name,input}}
+  // OpenAI style inside Gemini: {function_call:{name,arguments}}
+  const fc = p.functionCall || p.function_call || p.tool_use || p.toolUse;
+  if (fc && typeof fc === "object") {
+    const nameRaw = fc.name || fc.functionName || p.name || "unknown_tool";
+    // Sanitize name: must start with letter or underscore for gemini, but we normalize
+    const name = String(nameRaw).replace(/[^a-zA-Z0-9_.-]/g, "_");
+    let argsRaw = fc.args || fc.arguments || fc.input || {};
+    let argsJson: string;
+    let argsParsed: unknown = argsRaw;
+
+    if (typeof argsRaw === "string") {
+      argsJson = argsRaw;
+      try {
+        argsParsed = JSON.parse(argsRaw);
+      } catch {
+        argsParsed = argsRaw;
+      }
+    } else {
+      try {
+        argsJson = JSON.stringify(argsRaw);
+      } catch {
+        argsJson = "{}";
+      }
+    }
+
+    const id = fc.id || p.id || `toolu_${fnv1a32(name + argsJson).toString(16)}`;
+
+    return {
+      toolCall: {
+        id: String(id),
+        name,
+        args: argsParsed,
+        argsJson,
+        isDelta: !!p.isDelta,
+      },
+    };
+  }
+
+  // Direct tool call at part level
+  if (p.name && (p.args !== undefined || p.arguments !== undefined || p.input !== undefined)) {
+    return {
+      toolCall: {
+        id: p.id || `toolu_${fnv1a32(p.name).toString(16)}`,
+        name: String(p.name),
+        args: p.args ?? p.arguments ?? p.input ?? {},
+        argsJson:
+          typeof (p.args ?? p.arguments ?? p.input) === "string"
+            ? (p.args ?? p.arguments ?? p.input)
+            : JSON.stringify(p.args ?? p.arguments ?? p.input ?? {}),
+      },
+    };
+  }
+
+  return {};
+}
+
+/**
+ * Maps Gemini finishReason to normalized lowercase reason
+ */
+export function mapFinishReason(rawReason: unknown): string | undefined {
+  if (!rawReason || typeof rawReason !== "string") return undefined;
+  const r = rawReason.trim().toUpperCase();
+  switch (r) {
+    case "STOP":
+    case "END_TURN":
+    case "FINISH":
+      return "stop";
+    case "MAX_TOKENS":
+    case "LENGTH":
+      return "length";
+    case "SAFETY":
+    case "RECITATION":
+      return "content_filter";
+    case "TOOL_CALLS":
+    case "TOOL_USE":
+      return "tool_calls";
+    default:
+      return rawReason.toLowerCase();
+  }
+}
+
+/**
+ * Core normalizer — converts raw SSE data strings (JSON) into normalized chunks.
+ * Tolerant to heterogeneous shapes observed in prod, daily, sandbox.
+ *
+ * @param rawDataStrings - array of data: payloads (already stripped of `data:` prefix)
+ * @returns normalized chunks (may be more than input if multiple candidates)
+ */
+// local variant: diverges from core (accepts an array of raw data strings and returns rich NormalizedChunk objects; core.normalizePayloads takes one string and returns any[])
+export function normalizePayloads(rawDataStrings: (string | unknown)[]): NormalizedChunk[] {
+  const out: NormalizedChunk[] = [];
+
+  for (const rawData of rawDataStrings) {
+    let parsed: unknown;
+
+    if (typeof rawData === "string") {
+      const trimmed = rawData.trim();
+      if (trimmed === "" || trimmed === SSE_DONE_SENTINEL) {
+        out.push({
+          raw: trimmed,
+          isFinished: true,
+          finishReason: "stop",
+        });
+        continue;
+      }
+      parsed = safeJsonParse(trimmed);
+      if (parsed === null) {
+        // If not JSON, treat as plain text delta (tolerant)
+        out.push({
+          text: rawData,
+          fullText: rawData,
+          raw: rawData,
+        });
+        continue;
+      }
+    } else {
+      parsed = rawData;
+    }
+
+    if (parsed === null) continue;
+
+    // Handle sentinel object
+    if ((parsed as any).__sentinelDone) {
+      out.push({
+        raw: parsed,
+        isFinished: true,
+        finishReason: "stop",
+      });
+      continue;
+    }
+
+    const candidates = unwrapCandidates(parsed);
+
+    for (let ci = 0; ci < candidates.length; ci++) {
+      const cand = candidates[ci] as any;
+      if (!cand) continue;
+
+      const parts = extractParts(cand);
+      let textAcc = "";
+      let thinkingAcc = "";
+      let thinkingSig: string | undefined;
+      const toolCalls: NormalizedToolCall[] = [];
+      let finishReason: string | undefined;
+      let groundingMetadata: unknown;
+      let safetyRatings: unknown;
+      let usageMetadata: unknown;
+
+      // Extract top-level fields from candidate
+      if (cand.finishReason || cand.finish_reason || cand.stopReason) {
+        finishReason = mapFinishReason(cand.finishReason || cand.finish_reason || cand.stopReason);
+      }
+      if (cand.groundingMetadata) groundingMetadata = cand.groundingMetadata;
+      if (cand.grounding_metadata) groundingMetadata = cand.grounding_metadata;
+      if (cand.safetyRatings) safetyRatings = cand.safetyRatings;
+      if (cand.usageMetadata || cand.usage_metadata) usageMetadata = cand.usageMetadata || cand.usage_metadata;
+
+      // If no parts, but candidate itself might be direct text (fallback)
+      if (parts.length === 0) {
+        // candidate may have content as string
+        if (typeof cand.content === "string") {
+          textAcc += cand.content;
+        } else if (typeof cand.text === "string" && !cand.thought) {
+          textAcc += cand.text;
+        } else if (typeof cand.thinking === "string") {
+          thinkingAcc += cand.thinking;
+          thinkingSig = cand.thoughtSignature || cand.signature;
+        }
+      } else {
+        for (const part of parts) {
+          const norm = normalizePart(part);
+          if (norm.text) textAcc += norm.text;
+          if (norm.thinking) thinkingAcc += norm.thinking;
+          if (norm.thinkingSignature) thinkingSig = norm.thinkingSignature;
+          if (norm.toolCall) toolCalls.push(norm.toolCall);
+        }
+      }
+
+      // Also check candidate-level thoughtSignature if no part signature
+      if (!thinkingSig && cand.thoughtSignature) thinkingSig = cand.thoughtSignature;
+      if (!thinkingSig && cand.signature) thinkingSig = cand.signature;
+
+      // Skip empty candidate unless it has finishReason
+      if (!textAcc && !thinkingAcc && toolCalls.length === 0 && !finishReason && !groundingMetadata) {
+        // Might still be useful if it contains usage etc, but skip if truly empty
+        // Keep as raw-only if it has any interesting field
+        if (!cand.candidates && !cand.content && !cand.parts) continue;
+      }
+
+      out.push({
+        text: textAcc || undefined,
+        fullText: textAcc || undefined,
+        thinking: thinkingAcc || undefined,
+        fullThinking: thinkingAcc || undefined,
+        thinkingSignature: thinkingSig,
+        toolCalls: toolCalls.length > 0 ? toolCalls : undefined,
+        finishReason,
+        groundingMetadata,
+        safetyRatings,
+        usageMetadata,
+        raw: cand,
+        candidateIndex: ci,
+        isThought: thinkingAcc ? true : undefined,
+        isFinished: finishReason ? true : undefined,
+      });
+    }
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Preservation of thinking blocks between chunks — delta tracking
+// ---------------------------------------------------------------------------
+
+/**
+ * Extracts delta from full text given previously sent buffer.
+ * If new text starts with old, returns suffix, else returns full new.
+ * This handles both cumulative and incremental server behavior.
+ *
+ * @param full - current accumulated text from server (may be cumulative)
+ * @param alreadySent - buffer already emitted to client
+ * @returns delta to emit now
+ */
+export function extractDelta(full: string, alreadySent: string): string {
+  if (!full) return "";
+  if (!alreadySent) return full;
+  if (full === alreadySent) return "";
+  if (full.startsWith(alreadySent)) {
+    return full.slice(alreadySent.length);
+  }
+  // If server sends incremental delta that doesn't start with previous (e.g. new block),
+  // we check if alreadySent ends with prefix of full? simplest fallback: return full
+  // But try to find largest common prefix
+  // For production safety, just return full if not prefix — deduplication layer above handles
+  // However attempt small heuristic: if full length < alreadySent length, likely reset -> return full
+  return full;
+}
+
+/**
+ * Tracks streaming session to preserve thinking and text between chunks.
+ * The observer watches two buffers: sentThinkingBuffer and sentTextBuffer.
+ * Prevents duplicated thinking blocks in UI (bug #120 in original repo).
+ */
+export class StreamingSessionTracker {
+  private fullThinking = "";
+  private sentThinking = "";
+  private fullText = "";
+  private sentText = "";
+  private thinkingSignature?: string | undefined;
+  private finished = false;
+  private lastFinishReason?: string | undefined;
+  private toolCallsSeen = new Map<string, NormalizedToolCall>();
+
+  /**
+   * @param signatureCache - optional LRU cache for thinking signatures
+   */
+  constructor(public readonly signatureCache: ReplaySignatureCache = globalThinkingSignatureCache) {}
+
+  /** Whether stream already finished */
+  get isFinished(): boolean {
+    return this.finished;
+  }
+
+  get currentThinking(): string {
+    return this.fullThinking;
+  }
+
+  get currentText(): string {
+    return this.fullText;
+  }
+
+  /**
+   * Pushes new normalized chunk, deduplicates thinking/text, updates signature cache,
+   * and returns delta-only chunk ready for transformer.
+   *
+   * @param chunk - normalized from normalizePayloads
+   * @returns delta chunk (with only new portions) + metadata
+   */
+  push(chunk: NormalizedChunk): NormalizedChunk {
+    if (!chunk) return chunk;
+
+    let textDelta: string | undefined;
+    let thinkingDelta: string | undefined;
+
+    // ---- thinking preservation ----
+    if (chunk.fullThinking || chunk.thinking) {
+      const incomingThinking = chunk.fullThinking ?? chunk.thinking ?? "";
+      // Accumulate if server sent cumulative? Heuristic: if incoming longer than current and starts with current, replace full with incoming
+      if (!this.fullThinking) {
+        this.fullThinking = incomingThinking;
+      } else {
+        // If incoming contains current as prefix, it is cumulative update
+        if (incomingThinking.startsWith(this.fullThinking)) {
+          this.fullThinking = incomingThinking;
+        } else if (this.fullThinking.startsWith(incomingThinking)) {
+          // incoming is subset already processed — keep existing full
+        } else {
+          // Different block or incremental — append if not duplicate
+          // Check if already contains incoming suffix
+          if (!this.fullThinking.includes(incomingThinking)) {
+            this.fullThinking += incomingThinking;
+          }
+        }
+      }
+
+      // Extract what still needs to be sent
+      thinkingDelta = extractDelta(this.fullThinking, this.sentThinking);
+      if (thinkingDelta) {
+        this.sentThinking += thinkingDelta;
+      }
+
+      // Signature handling — preserve latest
+      if (chunk.thinkingSignature) {
+        this.thinkingSignature = chunk.thinkingSignature;
+        // Cache it keyed by current full thinking (when block finalized)
+        // But we cache progressively as well to avoid losing signature on crash
+        try {
+          this.signatureCache.set(this.fullThinking, chunk.thinkingSignature);
+        } catch {
+          // ignore cache errors
+        }
+      } else {
+        // Try to recover signature from cache if this thinking text was seen before
+        // (multiturn continuation)
+        try {
+          const cachedSig = this.signatureCache.get(this.fullThinking);
+          if (cachedSig) {
+            this.thinkingSignature = cachedSig;
+            chunk.thinkingSignature = cachedSig;
+          }
+        } catch {
+          // ignore
+        }
+      }
+    }
+
+    // ---- text preservation ----
+    if (chunk.fullText || chunk.text) {
+      const incomingText = chunk.fullText ?? chunk.text ?? "";
+      if (!this.fullText) {
+        this.fullText = incomingText;
+      } else {
+        if (incomingText.startsWith(this.fullText)) {
+          this.fullText = incomingText;
+        } else if (!this.fullText.includes(incomingText)) {
+          // Append only if not already included
+          this.fullText += incomingText;
+        }
+      }
+      textDelta = extractDelta(this.fullText, this.sentText);
+      if (textDelta) this.sentText += textDelta;
+    }
+
+    // ---- tool calls ----
+    let mergedToolCalls: NormalizedToolCall[] | undefined;
+    if (chunk.toolCalls && chunk.toolCalls.length > 0) {
+      mergedToolCalls = [];
+      for (const tc of chunk.toolCalls) {
+        const prev = this.toolCallsSeen.get(tc.id);
+        if (!prev) {
+          this.toolCallsSeen.set(tc.id, tc);
+          mergedToolCalls.push(tc);
+        } else {
+          // Merge args (delta)
+          if (tc.argsJson && tc.argsJson !== prev.argsJson) {
+            // append delta args if looks like incremental JSON
+            const mergedArgsJson = prev.argsJson + tc.argsJson;
+            // Keep merged but also provide delta
+            const merged: NormalizedToolCall = {
+              ...tc,
+              argsJson: mergedArgsJson,
+              isDelta: true,
+            };
+            this.toolCallsSeen.set(tc.id, { ...prev, argsJson: mergedArgsJson });
+            mergedToolCalls.push(merged);
+          }
+        }
+      }
+      if (mergedToolCalls.length === 0) mergedToolCalls = undefined;
+    }
+
+    // ---- finish tracking ----
+    if (chunk.finishReason || chunk.isFinished) {
+      this.finished = true;
+      this.lastFinishReason = chunk.finishReason || "stop";
+    }
+
+    return {
+      ...chunk,
+      text: textDelta ?? undefined,
+      fullText: this.fullText || undefined,
+      thinking: thinkingDelta ?? undefined,
+      fullThinking: this.fullThinking || undefined,
+      thinkingSignature: this.thinkingSignature || chunk.thinkingSignature,
+      toolCalls: mergedToolCalls,
+    };
+  }
+
+  /**
+   * Returns synthetic finish chunk if stream ended without explicit finish.
+   * Preserves last known signature.
+   */
+  createSyntheticFinish(reason = "stop"): NormalizedChunk | null {
+    if (this.finished) return null;
+    this.finished = true;
+    this.lastFinishReason = reason;
+    return {
+      raw: { synthetic: true, reason, fullThinking: this.fullThinking, fullText: this.fullText },
+      finishReason: reason,
+      isFinished: true,
+      thinkingSignature: this.thinkingSignature,
+      fullThinking: this.fullThinking || undefined,
+      fullText: this.fullText || undefined,
+    };
+  }
+
+  /** Resets tracker for new stream */
+  reset(): void {
+    this.fullThinking = "";
+    this.sentThinking = "";
+    this.fullText = "";
+    this.sentText = "";
+    this.thinkingSignature = undefined;
+    this.finished = false;
+    this.lastFinishReason = undefined;
+    this.toolCallsSeen.clear();
+  }
+
+  /** Diagnostic snapshot */
+  snapshot(): {
+    fullThinkingLen: number;
+    sentThinkingLen: number;
+    fullTextLen: number;
+    sentTextLen: number;
+    finished: boolean;
+    toolCalls: number;
+    hasSignature: boolean;
+  } {
+    return {
+      fullThinkingLen: this.fullThinking.length,
+      sentThinkingLen: this.sentThinking.length,
+      fullTextLen: this.fullText.length,
+      sentTextLen: this.sentText.length,
+      finished: this.finished,
+      toolCalls: this.toolCallsSeen.size,
+      hasSignature: !!this.thinkingSignature,
+    };
+  }
+}
+
+// ---------------------------------------------------------------------------
+// OpenAI compatible transformer — opencode expects chat.completion.chunk
+// ---------------------------------------------------------------------------
+
+export interface OpenAIChunkOptions {
+  /** Model id to echo back — e.g. antigravity-gemini-3-flash */
+  model: string;
+  /** Optional fixed id for stream — if not provided random-ish */
+  id?: string;
+  /** Creation timestamp in seconds */
+  created?: number;
+}
+
+function generateChatId(): string {
+  try {
+    const bytes = createHash("sha256")
+      .update(String(Date.now()) + Math.random())
+      .digest("hex")
+      .slice(0, 12);
+    return `chatcmpl-${bytes}`;
+  } catch {
+    return `chatcmpl-${fnv1a32(String(Date.now())).toString(16)}`;
+  }
+}
+
+/**
+ * Maps normalized finishReason to OpenAI finish_reason values
+ */
+function mapToOpenAIFinishReason(reason?: string): string | null {
+  if (!reason) return null;
+  switch (reason) {
+    case "stop":
+      return "stop";
+    case "length":
+      return "length";
+    case "content_filter":
+      return "content_filter";
+    case "tool_calls":
+      return "tool_calls";
+    case "tool_use":
+      return "tool_calls";
+    default:
+      return reason;
+  }
+}
+
+/**
+ * Transforms normalized chunks into OpenAI SSE lines.
+ * Library-first: stateless transform + session tracker for delta.
+ *
+ * The observer emits reasoning_content for thinking (opencode supports both
+ * reasoning and reasoning_content per 2026 spec).
+ *
+ * @param chunk - delta chunk from StreamingSessionTracker
+ * @param opts - model/id/created
+ * @returns array of SSE formatted strings `data: {...}\n\n`
+ */
+// local variant: diverges from core (returns SSE-formatted `data: {...}` strings with OpenAIChunkOptions; core.transformToOpenAI returns a bare chunk object)
+export function transformToOpenAI(
+  chunk: NormalizedChunk,
+  opts: OpenAIChunkOptions,
+  tracker?: StreamingSessionTracker,
+): string[] {
+  const id = opts.id ?? generateChatId();
+  const created = opts.created ?? Math.floor(Date.now() / 1000);
+  const choiceIndex = chunk.candidateIndex ?? 0;
+
+  const delta: any = {};
+
+  // Thinking — map to reasoning_content (opencode + OpenAI compat)
+  if (chunk.thinking && chunk.thinking.length > 0) {
+    delta.reasoning_content = chunk.thinking;
+    delta.reasoning = chunk.thinking; // alt field for compatibility
+    // Include signature if present — some clients use it for multi-turn
+    if (chunk.thinkingSignature) {
+      delta.reasoning_content_signature = chunk.thinkingSignature;
+      delta.thinking_signature = chunk.thinkingSignature;
+    }
+  }
+
+  // Text
+  if (chunk.text && chunk.text.length > 0) {
+    delta.content = chunk.text;
+  }
+
+  // Tool calls
+  if (chunk.toolCalls && chunk.toolCalls.length > 0) {
+    delta.tool_calls = chunk.toolCalls.map((tc, idx) => ({
+      index: idx,
+      id: tc.id,
+      type: "function",
+      function: {
+        name: tc.name,
+        // OpenAI streams incremental args as delta; we send full or delta depending on isDelta
+        arguments: tc.argsJson,
+      },
+    }));
+  }
+
+  // Grounding metadata passthrough as extra field for search models
+  if (chunk.groundingMetadata) {
+    delta.grounding_metadata = chunk.groundingMetadata;
+  }
+
+  const hasDelta = Object.keys(delta).length > 0;
+  const finish = mapToOpenAIFinishReason(chunk.finishReason);
+
+  // If nothing to emit and no finish, skip (empty chunk)
+  if (!hasDelta && !finish && !chunk.isFinished) {
+    return [];
+  }
+
+  const payload: any = {
+    id,
+    object: "chat.completion.chunk",
+    created,
+    model: opts.model,
+    choices: [
+      {
+        index: choiceIndex,
+        delta: hasDelta ? delta : {},
+        finish_reason: finish ?? null,
+        // Include logprobs null per OpenAI spec
+        logprobs: null,
+      },
+    ],
+  };
+
+  // If usage present, include for final chunk (some clients expect)
+  if (chunk.usageMetadata && finish) {
+    payload.usage = chunk.usageMetadata;
+  }
+
+  return [`data: ${JSON.stringify(payload)}\n\n`];
+}
+
+/**
+ * Factory that returns stateful OpenAI transformer with internal tracker
+ * and LRU signature cache. Suitable for opencode plugin hook.
+ *
+ * @param opts - model options
+ * @param cache - optional cache injection
+ */
+export function createOpenAIStreamingTransformer(
+  opts: OpenAIChunkOptions,
+  cache: ReplaySignatureCache = globalThinkingSignatureCache,
+): {
+  tracker: StreamingSessionTracker;
+  transform: (chunk: NormalizedChunk) => string[];
+  final: () => string[];
+} {
+  const tracker = new StreamingSessionTracker(cache);
+  const fixedId = opts.id ?? generateChatId();
+  const created = opts.created ?? Math.floor(Date.now() / 1000);
+
+  return {
+    tracker,
+    transform: (chunk: NormalizedChunk) => {
+      const deduped = tracker.push(chunk);
+      // If after dedup nothing new but finish, still emit finish
+      if (!deduped.text && !deduped.thinking && !deduped.toolCalls && !deduped.finishReason) {
+        if (deduped.isFinished || deduped.finishReason) {
+          // force emit finish even if deduped empty
+          return transformToOpenAI(deduped, { ...opts, id: fixedId, created }, tracker);
+        }
+        return [];
+      }
+      return transformToOpenAI(deduped, { ...opts, id: fixedId, created }, tracker);
+    },
+    final: () => {
+      const synthetic = tracker.createSyntheticFinish("stop");
+      if (!synthetic) return [];
+      return transformToOpenAI(synthetic, { ...opts, id: fixedId, created }, tracker);
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Anthropic compatible transformer — opencode uses Anthropic messages SSE
+// ---------------------------------------------------------------------------
+
+export interface AnthropicChunkOptions {
+  /** Model to echo */
+  model: string;
+  /** Message id */
+  messageId?: string;
+}
+
+type AnthropicBlockType = "text" | "thinking" | "tool_use";
+
+/**
+ * Internal state for Anthropic streaming sequence.
+ * Observer maintains index counters per content block.
+ */
+class AnthropicStateMachine {
+  public messageStarted = false;
+  public nextBlockIndex = 0;
+  public activeBlocks = new Map<number, AnthropicBlockType>();
+  public blockThinkingSent = new Map<number, string>();
+  public blockTextSent = new Map<number, string>();
+  public toolJsonSent = new Map<string, string>();
+  public finished = false;
+
+  reset(): void {
+    this.messageStarted = false;
+    this.nextBlockIndex = 0;
+    this.activeBlocks.clear();
+    this.blockThinkingSent.clear();
+    this.blockTextSent.clear();
+    this.toolJsonSent.clear();
+    this.finished = false;
+  }
+}
+
+/**
+ * Transforms normalized chunk into Anthropic SSE events.
+ * Emits event: / data: pairs as per Anthropic Messages streaming spec:
+ *   message_start, content_block_start, content_block_delta, content_block_stop,
+ *   message_delta, message_stop
+ *
+ * Library-first: each call returns array of formatted SSE lines.
+ *
+ * @param chunk - delta chunk from tracker
+ * @param opts - model/message options
+ * @param state - anthropic state machine (mutable)
+ * @returns array of SSE lines (`event: xxx\ndata: {...}\n\n`)
+ */
+// local variant: diverges from core (full stateful Anthropic event machine with block tracking; core.transformToAnthropic maps one payload to a single delta event)
+export function transformToAnthropic(
+  chunk: NormalizedChunk,
+  opts: AnthropicChunkOptions,
+  state: AnthropicStateMachine,
+): string[] {
+  const out: string[] = [];
+  const messageId = opts.messageId ?? `msg_${fnv1a32(String(Date.now())).toString(16)}`;
+
+  // Helper to emit event+data pair
+  const emit = (eventType: string, dataObj: unknown): void => {
+    out.push(`event: ${eventType}\ndata: ${JSON.stringify(dataObj)}\n\n`);
+  };
+
+  // message_start on first emission
+  if (!state.messageStarted) {
+    state.messageStarted = true;
+    emit("message_start", {
+      type: "message_start",
+      message: {
+        id: messageId,
+        type: "message",
+        role: "assistant",
+        content: [],
+        model: opts.model,
+        stop_reason: null,
+        stop_sequence: null,
+        usage: { input_tokens: 0, output_tokens: 0 },
+      },
+    });
+  }
+
+  // Thinking handling
+  if (chunk.thinking && chunk.thinking.length > 0) {
+    // Find active thinking block or create new
+    let idx = Array.from(state.activeBlocks.entries()).find(([, t]) => t === "thinking")?.[0];
+    if (idx === undefined) {
+      idx = state.nextBlockIndex++;
+      state.activeBlocks.set(idx, "thinking");
+      state.blockThinkingSent.set(idx, "");
+      emit("content_block_start", {
+        type: "content_block_start",
+        index: idx,
+        content_block: {
+          type: "thinking",
+          thinking: "",
+          signature: chunk.thinkingSignature ?? "",
+        },
+      });
+    }
+
+    // Delta
+    const prev = state.blockThinkingSent.get(idx) ?? "";
+    // chunk.thinking already delta from tracker, but we still track
+    // For anthropic, thinking delta event
+    emit("content_block_delta", {
+      type: "content_block_delta",
+      index: idx,
+      delta: {
+        type: "thinking_delta",
+        thinking: chunk.thinking,
+      },
+    });
+    state.blockThinkingSent.set(idx, prev + chunk.thinking);
+  }
+
+  // Text handling
+  if (chunk.text && chunk.text.length > 0) {
+    let idx = Array.from(state.activeBlocks.entries()).find(([, t]) => t === "text")?.[0];
+    // If thinking block active, we should keep text separate — create new block if first text
+    // But also text may come after thinking stopped — we need new block
+    const hasActiveText = state.activeBlocks.has(idx as number) && state.activeBlocks.get(idx as number) === "text";
+    if (!hasActiveText) {
+      // If there is active thinking and we now have text, we should stop thinking block first (per Claude spec thinking must precede text)
+      const thinkingIdx = Array.from(state.activeBlocks.entries()).find(([, t]) => t === "thinking")?.[0];
+      if (thinkingIdx !== undefined) {
+        emit("content_block_stop", {
+          type: "content_block_stop",
+          index: thinkingIdx,
+        });
+        state.activeBlocks.delete(thinkingIdx);
+      }
+      idx = state.nextBlockIndex++;
+      state.activeBlocks.set(idx, "text");
+      state.blockTextSent.set(idx, "");
+      emit("content_block_start", {
+        type: "content_block_start",
+        index: idx,
+        content_block: {
+          type: "text",
+          text: "",
+        },
+      });
+    }
+
+    emit("content_block_delta", {
+      type: "content_block_delta",
+      index: idx,
+      delta: {
+        type: "text_delta",
+        text: chunk.text,
+      },
+    });
+    const prev = state.blockTextSent.get(idx!) ?? "";
+    state.blockTextSent.set(idx!, prev + chunk.text);
+  }
+
+  // Tool calls
+  if (chunk.toolCalls && chunk.toolCalls.length > 0) {
+    for (const tc of chunk.toolCalls) {
+      // One content block per tool_use
+      // Check if tool already has block
+      let idx: number | undefined;
+      // We use tool id as map key for block index lookup
+      for (const entry of Array.from(state.activeBlocks.entries())) {
+        const bIdx = entry[0];
+        const bType = entry[1];
+        if (bType === "tool_use" && state.toolJsonSent.has(`idx_${bIdx}_id_${tc.id}`)) {
+          idx = bIdx;
+          break;
+        }
+      }
+      const isNew = idx === undefined;
+      if (isNew) {
+        idx = state.nextBlockIndex++;
+        state.activeBlocks.set(idx, "tool_use");
+        state.toolJsonSent.set(`idx_${idx}_id_${tc.id}`, "");
+        state.toolJsonSent.set(`id_${tc.id}`, String(idx)); // reverse lookup
+        emit("content_block_start", {
+          type: "content_block_start",
+          index: idx,
+          content_block: {
+            type: "tool_use",
+            id: tc.id,
+            name: tc.name,
+            input: {},
+          },
+        });
+      }
+
+      // input_json delta
+      // For Anthropic, partial_json is incremental input json
+      const prevJson = state.toolJsonSent.get(`idx_${idx}_id_${tc.id}`) ?? "";
+      const deltaJson = tc.isDelta ? tc.argsJson : extractDelta(tc.argsJson, prevJson) || tc.argsJson;
+      if (deltaJson) {
+        emit("content_block_delta", {
+          type: "content_block_delta",
+          index: idx,
+          delta: {
+            type: "input_json_delta",
+            partial_json: deltaJson,
+          },
+        });
+        state.toolJsonSent.set(`idx_${idx}_id_${tc.id}`, prevJson + deltaJson);
+      }
+    }
+  }
+
+  // Finish handling
+  if (chunk.finishReason || chunk.isFinished) {
+    // Stop all active blocks
+    for (const bIdx of Array.from(state.activeBlocks.keys())) {
+      emit("content_block_stop", {
+        type: "content_block_stop",
+        index: bIdx,
+      });
+    }
+    state.activeBlocks.clear();
+
+    // Map finish reason to Anthropic stop_reason
+    let stopReason: string | null = "end_turn";
+    switch (chunk.finishReason) {
+      case "tool_calls":
+        stopReason = "tool_use";
+        break;
+      case "length":
+        stopReason = "max_tokens";
+        break;
+      case "content_filter":
+        stopReason = "stop_sequence";
+        break;
+      default:
+        stopReason = chunk.finishReason ? "end_turn" : "end_turn";
+    }
+
+    emit("message_delta", {
+      type: "message_delta",
+      delta: {
+        stop_reason: stopReason,
+        stop_sequence: null,
+      },
+      usage: chunk.usageMetadata ?? { output_tokens: 0 },
+    });
+
+    emit("message_stop", {
+      type: "message_stop",
+    });
+
+    state.finished = true;
+  }
+
+  return out;
+}
+
+/**
+ * Factory for Anthropic streaming transformer compatible with opencode
+ *
+ * @param opts - model options
+ * @param cache - LRU signature cache
+ */
+export function createAnthropicStreamingTransformer(
+  opts: AnthropicChunkOptions,
+  cache: ReplaySignatureCache = globalThinkingSignatureCache,
+): {
+  tracker: StreamingSessionTracker;
+  anthropicState: AnthropicStateMachine;
+  transform: (chunk: NormalizedChunk) => string[];
+  final: () => string[];
+} {
+  const tracker = new StreamingSessionTracker(cache);
+  const anthropicState = new AnthropicStateMachine();
+
+  return {
+    tracker,
+    anthropicState,
+    transform: (chunk: NormalizedChunk) => {
+      const deduped = tracker.push(chunk);
+      // If nothing new and not finish, skip
+      if (!deduped.text && !deduped.thinking && !deduped.toolCalls && !deduped.finishReason && !deduped.isFinished) {
+        return [];
+      }
+      return transformToAnthropic(deduped, opts, anthropicState);
+    },
+    final: () => {
+      const synthetic = tracker.createSyntheticFinish("stop");
+      if (!synthetic) return [];
+      return transformToAnthropic(synthetic, opts, anthropicState);
+    },
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Auto-recovery — retry exponencial e injeção finish sintético
+// ---------------------------------------------------------------------------
+
+export interface RetryOptions {
+  /** Max attempts including first */
+  maxRetries?: number;
+  /** Base delay ms */
+  baseMs?: number;
+  /** Factor */
+  factor?: number;
+  /** Max delay ms */
+  maxMs?: number;
+  /** Which status codes trigger retry */
+  retryableStatus?: number[];
+}
+
+/**
+ * Determines if error/status is retryable.
+ * Observer pattern: checks 429, 5xx, and network failures.
+ *
+ * @param status - HTTP status or undefined for network error
+ */
+// local variant: diverges from models (adds opts.retryableStatus and treats undefined as retryable network failure; models.isRetryableStatus takes a plain status and hardcodes 403/404/408/429/5xx)
+export function isRetryableNetworkStatus(status: number | undefined, opts?: RetryOptions): boolean {
+  if (status === undefined) return true; // network failure retryable
+  const list = opts?.retryableStatus ?? (RETRYABLE_STATUS_CODES as unknown as number[]);
+  return list.includes(status);
+}
+
+/**
+ * Fetches with exponential backoff + jitter.
+ * Uses only node builtins + global fetch.
+ * Third-person observer: it watches fetch, retries on transient errors.
+ *
+ * @param input - RequestInfo
+ * @param init - RequestInit
+ * @param retryOpts - backoff options
+ * @returns Response
+ * @throws last error after retries exhausted
+ */
+export async function fetchWithExponentialRetry(
+  input: Parameters<typeof fetch>[0],
+  init?: RequestInit,
+  retryOpts: RetryOptions = {},
+): Promise<Response> {
+  const maxRetries = retryOpts.maxRetries ?? MAX_RETRIES;
+  const baseMs = retryOpts.baseMs ?? BACKOFF_BASE_MS;
+  const factor = retryOpts.factor ?? BACKOFF_FACTOR;
+  const maxMs = retryOpts.maxMs ?? BACKOFF_MAX_MS;
+
+  let attempt = 0;
+  let lastError: unknown;
+
+  while (attempt <= maxRetries) {
+    try {
+      const res = await fetch(input, init);
+
+      if (res.ok) return res;
+
+      // Non-ok: check if retryable
+      if (!isRetryableNetworkStatus(res.status, retryOpts) || attempt >= maxRetries) {
+        return res; // return non-retryable response for caller to handle
+      }
+
+      // Retryable status — consume body to free connection before retry
+      try {
+        await res.text();
+      } catch {
+        // ignore
+      }
+
+      lastError = new Error(`retryable status ${res.status}`);
+    } catch (err) {
+      lastError = err;
+      if (attempt >= maxRetries) break;
+      // network error — retryable unless explicitly non-retryable
+    }
+
+    // Backoff before next attempt
+    if (attempt < maxRetries) {
+      const delay = Math.min(baseMs * Math.pow(factor, attempt), maxMs);
+      await sleepWithJitter(delay);
+    }
+
+    attempt++;
+  }
+
+  throw lastError instanceof Error ? lastError : new Error(String(lastError ?? "fetch retry exhausted"));
+}
+
+/**
+ * Creates synthetic finish chunk when stream closes abruptly.
+ * Observer injects `stop` finish to keep opencode SDK happy.
+ *
+ * @param tracker - session tracker to know current state
+ * @param reason - finish reason, default stop
+ * @returns normalized finish chunk or null if already finished
+ */
+export function injectSyntheticFinish(
+  tracker: StreamingSessionTracker,
+  reason: string = "stop",
+): NormalizedChunk | null {
+  return tracker.createSyntheticFinish(reason);
+}
+
+/**
+ * Wraps streaming response reading with auto-recovery.
+ * If reader throws before finish, retries whole fetch (exponential).
+ * On final failure, injects synthetic finish so UI doesn't hang.
+ *
+ * @param fetchFn - function that returns Promise<Response> (fresh fetch each attempt)
+ * @param onChunk - callback for each normalized chunk (deduped)
+ * @param retryOpts - retry options for fetch phase
+ * @param tracker - optional tracker (if not provided new one created)
+ */
+export async function withAutoRecoveryStreaming(
+  fetchFn: () => Promise<Response>,
+  onChunk: (chunk: NormalizedChunk) => void | Promise<void>,
+  retryOpts: RetryOptions = {},
+  tracker: StreamingSessionTracker = new StreamingSessionTracker(globalThinkingSignatureCache),
+): Promise<void> {
+  const maxRetries = retryOpts.maxRetries ?? MAX_RETRIES;
+  let attempt = 0;
+
+  while (attempt <= maxRetries) {
+    let response: Response;
+    try {
+      response = await fetchWithExponentialRetry(
+        // dummy — actual fetchFn does request construction
+        // We call fetchFn directly but reuse retry logic for network: we wrap fetchFn with backoff here
+        // So first we try fetchFn()
+        // fetchWithExponentialRetry expects input, not fn — we implement manual retry loop for fetchFn
+        // Instead we directly call fetchFn and apply retry inside catch below
+        // For simplicity we call fetchFn here and if it throws we retry
+        // This inner try will rely on outer loop for backoff
+        "" as unknown as Parameters<typeof fetch>[0],
+        undefined,
+        retryOpts,
+      ).catch(() => {
+        // unreachable because we pass empty input — we will replace with fetchFn call
+        return null as any;
+      });
+      // Actually we need to call fetchFn, not fetchWithExponentialRetry with dummy
+      // Re-do: override response with fetchFn result
+      response = await fetchFn();
+    } catch (err) {
+      if (attempt >= maxRetries) {
+        // Final failure — inject synthetic finish and exit
+        const synthetic = tracker.createSyntheticFinish("stop");
+        if (synthetic) await onChunk(synthetic);
+        throw err;
+      }
+      const delay = Math.min(
+        (retryOpts.baseMs ?? BACKOFF_BASE_MS) * Math.pow(retryOpts.factor ?? BACKOFF_FACTOR, attempt),
+        retryOpts.maxMs ?? BACKOFF_MAX_MS,
+      );
+      await sleepWithJitter(delay);
+      attempt++;
+      continue;
+    }
+
+    if (!response.ok) {
+      if (isRetryableNetworkStatus(response.status, retryOpts) && attempt < maxRetries) {
+        try {
+          await response.text();
+        } catch {
+          // ignore
+        }
+        const delay = Math.min(
+          (retryOpts.baseMs ?? BACKOFF_BASE_MS) * Math.pow(retryOpts.factor ?? BACKOFF_FACTOR, attempt),
+          retryOpts.maxMs ?? BACKOFF_MAX_MS,
+        );
+        await sleepWithJitter(delay);
+        attempt++;
+        continue;
+      }
+      // Non-retryable error — propagate response error as exception with body preview
+      let bodyPreview = "";
+      try {
+        bodyPreview = (await response.text()).slice(0, 2000);
+      } catch {
+        bodyPreview = "<no body>";
+      }
+      throw new Error(`cloudcode-pa ${response.status} ${response.statusText}: ${bodyPreview}`);
+    }
+
+    // Stream reading phase — try to read all
+    try {
+      for await (const normalized of readSSEStream(response, tracker)) {
+        await onChunk(normalized);
+      }
+
+      // Normal end — ensure finish injected if missing
+      if (!tracker.isFinished) {
+        const synthetic = tracker.createSyntheticFinish("stop");
+        if (synthetic) await onChunk(synthetic);
+      }
+
+      return; // success
+    } catch (streamErr) {
+      // Stream broke mid-way, retry if not finished and attempts left
+      if (attempt >= maxRetries || tracker.isFinished) {
+        if (!tracker.isFinished) {
+          const synthetic = tracker.createSyntheticFinish("stop");
+          if (synthetic) await onChunk(synthetic);
+        }
+        throw streamErr;
+      }
+      const delay = Math.min(
+        (retryOpts.baseMs ?? BACKOFF_BASE_MS) * Math.pow(retryOpts.factor ?? BACKOFF_FACTOR, attempt),
+        retryOpts.maxMs ?? BACKOFF_MAX_MS,
+      );
+      await sleepWithJitter(delay);
+      attempt++;
+      // Continue retry loop — will re-fetch and continue? Note: we lost progress but tracker preserves sent state for dedup
+      continue;
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// High-level SSE stream reader — direct cloudcode-pa
+// ---------------------------------------------------------------------------
+
+/**
+ * Async generator that reads a fetch Response body (cloudcode-pa alt=sse)
+ * and yields normalized chunks with preserved thinking blocks.
+ *
+ * The observer directly consumes https://cloudcode-pa.googleapis.com
+ * (or daily-.../sandbox) without localhost v1 proxy.
+ *
+ * @param response - fetch Response with body ReadableStream
+ * @param tracker - optional tracker for delta preservation (new one if not provided)
+ * @yields NormalizedChunk (deduped delta already applied via tracker)
+ */
+export async function* readSSEStream(
+  response: Response,
+  tracker: StreamingSessionTracker = new StreamingSessionTracker(globalThinkingSignatureCache),
+): AsyncGenerator<NormalizedChunk> {
+  if (!response || !response.body) {
+    throw new TypeError("readSSEStream requires Response with body");
+  }
+
+  const parserState = createSSEParserState();
+  const decoder = new TextDecoder("utf-8");
+
+  // Node fetch Response.body is a ReadableStream (web)
+  const reader = (response.body as any).getReader() as ReadableStreamDefaultReader<Uint8Array>;
+
+  if (!reader || typeof reader.read !== "function") {
+    throw new Error("Response body does not implement getReader() — unsupported environment");
+  }
+
+  try {
+    while (true) {
+      let result: { done: boolean; value?: Uint8Array | undefined };
+      try {
+        result = await reader.read();
+      } catch (readErr) {
+        // Network mid-stream failure — let outer auto-recovery handle
+        const wrapped = new Error(`stream read failed: ${(readErr as Error).message}`);
+        (wrapped as any).cause = readErr;
+        throw wrapped;
+      }
+
+      if (result.done) break;
+
+      const value = result.value;
+      if (!value || value.length === 0) continue;
+
+      let textChunk: string;
+      try {
+        textChunk = decoder.decode(value, { stream: true });
+      } catch (decErr) {
+        console.warn("[streaming] TextDecoder failed, skipping chunk", decErr);
+        continue;
+      }
+
+      let events: ParsedSSEEvent[];
+      try {
+        events = parseSSEChunk(textChunk, parserState);
+      } catch (parseErr) {
+        console.warn("[streaming] parseSSEChunk failed", parseErr);
+        continue;
+      }
+
+      if (events.length === 0) continue;
+
+      for (const ev of events) {
+        // Skip [DONE] sentinel here — normalize will handle but we also short-circuit
+        if (ev.data.trim() === SSE_DONE_SENTINEL) {
+          const chunks = normalizePayloads([ev.data]);
+          for (const c of chunks) {
+            const deduped = tracker.push(c);
+            yield deduped;
+          }
+          continue;
+        }
+
+        const normalizedList = normalizePayloads([ev.data]);
+
+        for (const norm of normalizedList) {
+          const deduped = tracker.push(norm);
+          // Yield only if has something new or finish
+          if (
+            deduped.text ||
+            deduped.thinking ||
+            (deduped.toolCalls && deduped.toolCalls.length > 0) ||
+            deduped.finishReason ||
+            deduped.isFinished
+          ) {
+            yield deduped;
+          } else if (deduped.isFinished) {
+            yield deduped;
+          }
+        }
+      }
+    }
+
+    // Flush decoder remaining
+    try {
+      const remaining = decoder.decode();
+      if (remaining) {
+        const events = parseSSEChunk(remaining, parserState);
+        for (const ev of events) {
+          const normalizedList = normalizePayloads([ev.data]);
+          for (const norm of normalizedList) {
+            yield tracker.push(norm);
+          }
+        }
+      }
+    } catch {
+      // ignore flush errors
+    }
+
+    // If buffer still has data without trailing \n\n, try to parse it as final block (tolerant)
+    if (parserState.buffer.trim().length > 0) {
+      const finalBlock = parserState.buffer.trim();
+      // Attempt to treat it as data line even without delimiter
+      // Remove possible "data:" prefix if present
+      let dataPayload = finalBlock;
+      if (dataPayload.startsWith("data:")) {
+        dataPayload = dataPayload.slice(5).trimStart();
+      }
+      if (dataPayload && dataPayload !== SSE_DONE_SENTINEL) {
+        const normalizedList = normalizePayloads([dataPayload]);
+        for (const norm of normalizedList) {
+          yield tracker.push(norm);
+        }
+      }
+      parserState.buffer = "";
+    }
+  } finally {
+    try {
+      reader.releaseLock();
+    } catch {
+      // ignore
+    }
+  }
+}
+
+/**
+ * Convenience: stream cloudcode-pa response and transform to OpenAI SSE lines.
+ * Used by opencode plugin fetch interceptor.
+ *
+ * @param response - fetch Response from cloudcode-pa
+ * @param openAIOpts - model info
+ * @param cache - optional LRU cache
+ * @returns AsyncGenerator<string> of `data: {...}\n\n` lines
+ */
+export async function* streamToOpenAI(
+  response: Response,
+  openAIOpts: OpenAIChunkOptions,
+  cache: ReplaySignatureCache = globalThinkingSignatureCache,
+): AsyncGenerator<string> {
+  const tracker = new StreamingSessionTracker(cache);
+  const fixedId = openAIOpts.id ?? generateChatId();
+  const created = openAIOpts.created ?? Math.floor(Date.now() / 1000);
+
+  for await (const norm of readSSEStream(response, tracker)) {
+    const lines = transformToOpenAI(norm, { ...openAIOpts, id: fixedId, created }, tracker);
+    for (const line of lines) yield line;
+  }
+
+  if (!tracker.isFinished) {
+    const synthetic = tracker.createSyntheticFinish("stop");
+    if (synthetic) {
+      const lines = transformToOpenAI(synthetic, { ...openAIOpts, id: fixedId, created }, tracker);
+      for (const line of lines) yield line;
+    }
+  }
+
+  // Final [DONE] for OpenAI clients
+  yield `data: ${SSE_DONE_SENTINEL}\n\n`;
+}
+
+/**
+ * Convenience: stream cloudcode-pa response and transform to Anthropic SSE lines.
+ *
+ * @param response - fetch Response
+ * @param anthropicOpts - model info
+ * @param cache - optional LRU cache
+ * @returns AsyncGenerator<string> of `event: ...\ndata: ...\n\n`
+ */
+export async function* streamToAnthropic(
+  response: Response,
+  anthropicOpts: AnthropicChunkOptions,
+  cache: ReplaySignatureCache = globalThinkingSignatureCache,
+): AsyncGenerator<string> {
+  const tracker = new StreamingSessionTracker(cache);
+  const anthropicState = new AnthropicStateMachine();
+
+  for await (const norm of readSSEStream(response, tracker)) {
+    const lines = transformToAnthropic(norm, anthropicOpts, anthropicState);
+    for (const line of lines) yield line;
+  }
+
+  if (!tracker.isFinished) {
+    const synthetic = tracker.createSyntheticFinish("stop");
+    if (synthetic) {
+      const lines = transformToAnthropic(synthetic, anthropicOpts, anthropicState);
+      for (const line of lines) yield line;
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Merged: canonical bypass constants (from ONDA 4 streaming module)
+// ---------------------------------------------------------------------------
+
+// GEMINI_CLI_CLIENT_ID / GEMINI_CLI_CLIENT_SECRET / PROJECT_FALLBACK are
+// imported from constants.js (identical values) and re-exported above.
+
+/** Production Cloud Code Assist base endpoint. */
+
+/** Exact User-Agent sent by Gemini CLI 0.57.0. */
+
+// local variant: diverges from constants (constants.X_GOOG_API_CLIENT is the antigravity/{ver} gl-node gax/grpc chain; this is the bare gl-node value)
+/** X-Goog-Api-Client value observed on real Gemini CLI traffic. */
+
+/** Client-Metadata header string used by Gemini CLI bypass requests. */
+
+// v2.1.15 Phase A: MODELS_2026 (the 11-id legacy bypass list) and the
+// ModelId2026 union moved to models.ts (MODELS_2026_LEGACY / ModelId2026) —
+// imported under the historical name and re-exported so the public surface
+// is unchanged.
+import { MODELS_2026_LEGACY as MODELS_2026, UA_DEFAULT_MODEL } from "./models.js";
+import type { ModelId2026 } from "./models.js";
+export { MODELS_2026 };
+export type { ModelId2026 };
+
+// ---------------------------------------------------------------------------
+// Merged: raw cloudcode-pa payload types
+// ---------------------------------------------------------------------------
+
+/**
+ * Raw part shape emitted by cloudcode-pa (Gemini core with Claude interop).
+ */
+export interface RawPart {
+  text?: string;
+  thought?: boolean;
+  thoughtSignature?: string;
+  functionCall?: { name: string; args?: Record<string, unknown> | string; id?: string };
+  inlineData?: { mimeType: string; data: string };
+  [k: string]: unknown;
+}
+
+/**
+ * Raw content container — either wrapped parts or bare text.
+ */
+export interface RawContent {
+  role?: string;
+  parts?: RawPart[];
+  text?: string;
+}
+
+/**
+ * Raw candidate inside a cloudcode-pa chunk.
+ */
+export interface RawCandidate {
+  content?: RawContent;
+  finishReason?: string;
+  finishMessage?: string;
+  index?: number;
+  tokenCount?: number;
+  safetyRatings?: unknown[];
+  parts?: RawPart[];
+  text?: string;
+  thought?: boolean;
+  thoughtSignature?: string;
+}
+
+/**
+ * Heterogeneous raw chunk from v1internal:streamGenerateContent.
+ */
+export interface CloudCodeRawChunk {
+  candidates?: RawCandidate[];
+  content?: RawContent;
+  parts?: RawPart[];
+  response?: CloudCodeRawChunk;
+  usageMetadata?: {
+    promptTokenCount?: number;
+    candidatesTokenCount?: number;
+    totalTokenCount?: number;
+    thoughtsTokenCount?: number;
+  };
+  model?: string;
+  modelVersion?: string;
+  thought?: string;
+  thoughtSignature?: string;
+  [k: string]: unknown;
+}
+
+// ---------------------------------------------------------------------------
+// Merged: normalized payload types
+// ---------------------------------------------------------------------------
+
+/** Kind discriminator for {@link NormalizedPart}. */
+export type NormalizedPartKind = "text" | "thinking" | "thought_signature" | "function_call" | "inline_data";
+
+/**
+ * One normalized part with provenance indexes and optional signature.
+ */
+export interface NormalizedPart {
+  kind: NormalizedPartKind;
+  text: string;
+  thought: boolean;
+  thoughtSignature?: string | undefined;
+  functionCall?: { name: string; args: string; id: string };
+  rawIndex: number;
+  candidateIndex: number;
+  finishReason?: string | undefined;
+}
+
+/**
+ * Fully normalized payload: parts plus finish/usage metadata.
+ */
+export interface NormalizedPayload {
+  parts: NormalizedPart[];
+  finishReason?: string | undefined;
+  usage?: CloudCodeRawChunk["usageMetadata"] | undefined;
+  model?: string | undefined;
+  isThinkingChunk: boolean;
+  hasFinish: boolean;
+  raw: CloudCodeRawChunk;
+}
+
+// ---------------------------------------------------------------------------
+// Merged: streaming state + OpenAI / Anthropic wire types
+// ---------------------------------------------------------------------------
+
+/**
+ * Mutable state shared by the stateful OpenAI / Anthropic transformers.
+ */
+export interface StreamingState {
+  completionId: string;
+  model: string;
+  created: number;
+  sentRole: boolean;
+  accumulatedText: string;
+  accumulatedThinking: string;
+  lastFinishReason?: string | undefined;
+  insideThinking: boolean;
+  anthropicBlockIndex: number;
+  anthropicBlockKinds: Map<number, NormalizedPartKind>;
+  toolCallIdCounter: number;
+  pendingSignatures: string[];
+}
+
+/**
+ * OpenAI chat.completion.chunk wire type emitted by the transformer.
+ */
+export interface OpenAIChatChunk {
+  id: string;
+  object: "chat.completion.chunk";
+  created: number;
+  model: string;
+  choices: Array<{
+    index: number;
+    delta: {
+      role?: "assistant";
+      content?: string;
+      reasoning_content?: string;
+      reasoning?: string;
+      tool_calls?: Array<{
+        index: number;
+        id: string;
+        type: "function";
+        function: { name: string; arguments: string };
+      }>;
+    };
+    finish_reason: string | null;
+  }>;
+  usage?: {
+    prompt_tokens?: number;
+    completion_tokens?: number;
+    total_tokens?: number;
+    completion_tokens_details?: { reasoning_tokens?: number };
+  };
+}
+
+/**
+ * Anthropic-style SSE event descriptor.
+ */
+export interface AnthropicSSEEvent {
+  type: string;
+  data?: unknown;
+  sse?: string;
+}
+
+/**
+ * Diagnostic result shape for remainder-based SSE parsing.
+ */
+export interface ParseSSEChunkResult {
+  payloads: CloudCodeRawChunk[];
+  remainder: string;
+  consumedBytes: number;
+  errors: string[];
+}
+
+// ---------------------------------------------------------------------------
+// Merged: multi-turn thinking signature LRU cache (entry-based)
+// ---------------------------------------------------------------------------
+
+/**
+ * Byte-wise FNV-1a 32-bit hash of a UTF-8 string, rendered as 8 hex chars.
+ * Shared by {@link ThinkingSignatureLRUCache.hashThinking} and
+ * {@link buildHostSessionId} (v2.1.16 internal-dedup: the two identical
+ * inline loops were extracted into this single private helper). Distinct from
+ * core.js fnv1a32, which hashes UTF-16 code units.
+ *
+ * @param value - input string
+ * @returns 8-char lowercase hex digest
+ */
+function fnv1a32HexBytes(value: string): string {
+  let h = 0x811c9dc5;
+  const buf = Buffer.from(value, "utf8");
+  for (let i = 0; i < buf.length; i++) {
+    h ^= buf[i]!;
+    h = Math.imul(h, 0x01000193) >>> 0;
+  }
+  return h.toString(16).padStart(8, "0");
+}
+
+/**
+ * Cached signature entry with turn/model provenance.
+ */
+export interface ReplaySignatureEntry {
+  signature: string;
+  thinkingText: string;
+  model: string;
+  timestamp: number;
+  turnId: string;
+  tokenCount?: number | undefined;
+}
+
+/**
+ * Entry-based LRU cache for thinking signatures across turns.
+ * Preserves thoughtSignature values that must be replayed on multi-turn
+ * requests for gemini-3-pro / claude thinking models.
+ *
+ * Kept alongside {@link ReplaySignatureCache}, which maps plain text hashes
+ * to signature strings.
+ */
+export class ThinkingSignatureLRUCache {
+  private readonly maxEntries: number;
+  private readonly map: Map<string, ReplaySignatureEntry>;
+
+  constructor(maxEntries = 100) {
+    this.maxEntries = maxEntries;
+    this.map = new Map();
+  }
+
+  private touch(key: string): void {
+    const entry = this.map.get(key);
+    if (!entry) return;
+    // LRU: re-insert at the end (most recently used)
+    this.map.delete(key);
+    this.map.set(key, entry);
+  }
+
+  get(key: string): ReplaySignatureEntry | undefined {
+    const v = this.map.get(key);
+    if (v) this.touch(key);
+    return v;
+  }
+
+  set(key: string, entry: ReplaySignatureEntry): void {
+    if (this.map.has(key)) this.map.delete(key);
+    this.map.set(key, entry);
+    while (this.map.size > this.maxEntries) {
+      const oldest = this.map.keys().next().value as string;
+      this.map.delete(oldest);
+    }
+  }
+
+  /**
+   * Short FNV-1a hex hash of thinking text, used as key when no turnId exists.
+   */
+  static hashThinking(text: string): string {
+    return fnv1a32HexBytes(text);
+  }
+
+  /**
+   * Adds an entry derived from a payload fragment and returns the cache key.
+   */
+  addFromPayload(payload: {
+    signature: string;
+    thinkingText?: string;
+    model?: string;
+    turnId?: string;
+    tokenCount?: number;
+  }): string {
+    const keyBase = payload.turnId ?? ThinkingSignatureLRUCache.hashThinking(payload.thinkingText ?? payload.signature);
+    const entry: ReplaySignatureEntry = {
+      signature: payload.signature,
+      thinkingText: payload.thinkingText ?? "",
+      model: payload.model ?? UA_DEFAULT_MODEL,
+      timestamp: Date.now(),
+      turnId: keyBase,
+      tokenCount: payload.tokenCount,
+    };
+    this.set(keyBase, entry);
+    return keyBase;
+  }
+
+  findByModel(model: string): ReplaySignatureEntry[] {
+    const out: ReplaySignatureEntry[] = [];
+    for (const e of Array.from(this.map.values())) {
+      if (e.model === model || e.model.includes(model) || model.includes(e.model)) out.push(e);
+    }
+    return out;
+  }
+
+  clear(): void {
+    this.map.clear();
+  }
+
+  get size(): number {
+    return this.map.size;
+  }
+
+  entries(): IterableIterator<[string, ReplaySignatureEntry]> {
+    return this.map.entries();
+  }
+
+  toJSON(): Record<string, ReplaySignatureEntry> {
+    const obj: Record<string, ReplaySignatureEntry> = {};
+    for (const kv of Array.from(this.map.entries())) obj[kv[0]] = kv[1];
+    return obj;
+  }
+}
+
+/**
+ * Global singleton of the entry-based LRU cache. Kept alongside the text-hash
+ * singleton {@link globalThinkingSignatureCache}.
+ */
+export const globalThinkingSignatureLRUCache = new ThinkingSignatureLRUCache(100);
+
+// ---------------------------------------------------------------------------
+// Merged: lightweight SSE parser utilities (v4/v5)
+// v2.1.16 dead-code cleanup: the exported LRU class (byte-identical twin of
+// core.ts's LRU, zero instantiations here) and the barrel-only compat trio
+// signatureCache / saveSignature / getSignature (zero consumers repo-wide)
+// were deleted; core.js keeps the shared LRU instance (thinkingCache).
+// ---------------------------------------------------------------------------
+
+/**
+ * Minimal SSE event shape used by the lightweight parser utilities.
+ */
+export type sseEvent = { data: string };
+
+/**
+ * Linear-time strip of SDK-injected `[cache_control...]` markers (regex-free;
+ * the former /\[cache_control.*?\]/g replace was polynomial on uncontrolled input).
+ */
+export const stripCacheControlMarkers = (text: string): string => {
+  let out = text;
+  let idx = out.indexOf("[cache_control");
+  while (idx !== -1) {
+    const end = out.indexOf("]", idx);
+    if (end === -1) break;
+    out = out.slice(0, idx) + out.slice(end + 1);
+    idx = out.indexOf("[cache_control", idx);
+  }
+  return out;
+};
+
+/**
+ * Strips SDK-injected cache_control markers from a thinking block payload.
+ *
+ * @param p - payload possibly containing `thinking`
+ * @param keepThinking - kept for API compatibility; both branches preserve thinking
+ * @returns payload with cleaned thinking text
+ */
+export const handleThinkingBlock = (p: any, keepThinking = false): any => {
+  if (!p.thinking) return p;
+  // Filter sdk-injected cache_control markers
+  const cleaned = {
+    ...p.thinking,
+    text: typeof p.thinking.text === "string" ? stripCacheControlMarkers(p.thinking.text) : p.thinking.text,
+  };
+  if (keepThinking) {
+    return { ...p, thinking: cleaned };
+  }
+  return { ...p, thinking: cleaned };
+};
+
+// ---------------------------------------------------------------------------
+// Merged: tolerant SSE data parser (remainder-based)
+// ---------------------------------------------------------------------------
+
+function safeJsonParseArray(text: string): { objects: CloudCodeRawChunk[]; remainder: string; errors: string[] } {
+  const objects: CloudCodeRawChunk[] = [];
+  const errors: string[] = [];
+  let remainder = text;
+
+  // Trim BOM
+  if (remainder.charCodeAt(0) === 0xfeff) remainder = remainder.slice(1);
+
+  let buffer = remainder;
+  remainder = "";
+
+  const lines = buffer.split(/\r?\n/);
+  let jsonAccumulator = "";
+  let insideDataBlock = false;
+  let dataPrefixFound = false;
+
+  const flushAccumulator = (): { parsed: CloudCodeRawChunk[]; leftover: string } => {
+    const trimmed = jsonAccumulator.trim();
+    if (!trimmed) return { parsed: [], leftover: "" };
+    const parsed: CloudCodeRawChunk[] = [];
+    let leftover = "";
+
+    // JSON array form [...]
+    if (trimmed.startsWith("[") && trimmed.endsWith("]")) {
+      try {
+        const arr = JSON.parse(trimmed) as unknown;
+        if (Array.isArray(arr)) {
+          for (const item of arr) {
+            if (item && typeof item === "object") parsed.push(item as CloudCodeRawChunk);
+          }
+          return { parsed, leftover: "" };
+        }
+      } catch {
+        // incomplete array -> treat as streamed object below
+      }
+    }
+
+    // Concatenated objects like }{ or }\n{ — balanced-brace incremental scan
+    let depth = 0;
+    let inString = false;
+    let escape = false;
+    let start = -1;
+    for (let i = 0; i < trimmed.length; i++) {
+      const ch = trimmed[i]!;
+      if (escape) {
+        escape = false;
+        continue;
+      }
+      if (ch === "\\" && inString) {
+        escape = true;
+        continue;
+      }
+      if (ch === '"') {
+        inString = !inString;
+        continue;
+      }
+      if (inString) continue;
+      if (ch === "{" || ch === "[") {
+        if (depth === 0) start = i;
+        depth++;
+      } else if (ch === "}" || ch === "]") {
+        depth--;
+        if (depth === 0 && start !== -1) {
+          const slice = trimmed.slice(start, i + 1);
+          try {
+            const obj = JSON.parse(slice);
+            if (obj && typeof obj === "object") parsed.push(obj as CloudCodeRawChunk);
+          } catch (e) {
+            // Invalid JSON in this slice -> keep as leftover for retry
+            errors.push(`json_parse_slice_failed: ${(e as Error).message} slice=${slice.slice(0, 120)}`);
+            leftover = trimmed.slice(start);
+            break;
+          }
+          start = -1;
+        }
+      }
+    }
+    if (depth !== 0 && start !== -1) {
+      leftover = trimmed.slice(start);
+    } else if (parsed.length === 0 && trimmed.length > 0 && leftover === "") {
+      leftover = trimmed;
+    }
+    return { parsed, leftover };
+  };
+
+  for (let li = 0; li < lines.length; li++) {
+    const line = lines[li]!;
+    const trimmedLine = line.trim();
+
+    if (trimmedLine.length === 0) {
+      // Empty line marks end of SSE event -> flush
+      if (jsonAccumulator.trim().length > 0) {
+        const res = flushAccumulator();
+        objects.push(...res.parsed);
+        if (res.leftover) {
+          jsonAccumulator = res.leftover;
+          if (li === lines.length - 1 && res.parsed.length === 0) {
+            break;
+          }
+        } else {
+          jsonAccumulator = "";
+        }
+      }
+      insideDataBlock = false;
+      dataPrefixFound = false;
+      continue;
+    }
+
+    // [DONE] sentinel from intermediate OpenAI-compatible wrappers
+    if (trimmedLine === "data: [DONE]" || trimmedLine === "[DONE]") {
+      continue;
+    }
+
+    let dataContent = "";
+    if (trimmedLine.startsWith("data:")) {
+      dataPrefixFound = true;
+      insideDataBlock = true;
+      dataContent = trimmedLine.slice(5).trimStart();
+      if (dataContent.length === 0) {
+        continue;
+      }
+    } else if (trimmedLine.startsWith("event:") || trimmedLine.startsWith("id:") || trimmedLine.startsWith(":")) {
+      // SSE metadata lines — ignore but keep framing
+      continue;
+    } else if (insideDataBlock || !dataPrefixFound) {
+      // Prefix-less line inside a data block (multi-line JSON) or plain ndjson
+      dataContent = trimmedLine;
+    }
+
+    if (dataContent.length > 0) {
+      // Preserve embedded newlines inside escaped JSON strings
+      if (jsonAccumulator.length > 0) jsonAccumulator += "\n";
+      jsonAccumulator += dataContent;
+      // Flush early when the accumulator looks complete and the next line starts
+      // a new event or the buffer ends — tolerates missing blank separators.
+      const looksComplete = (() => {
+        let d = 0,
+          ins = false,
+          esc = false;
+        for (let k = 0; k < jsonAccumulator.length; k++) {
+          const c = jsonAccumulator[k]!;
+          if (esc) {
+            esc = false;
+            continue;
+          }
+          if (c === "\\" && ins) {
+            esc = true;
+            continue;
+          }
+          if (c === '"') {
+            ins = !ins;
+            continue;
+          }
+          if (ins) continue;
+          if (c === "{" || c === "[") d++;
+          else if (c === "}" || c === "]") d--;
+        }
+        return d === 0 && !ins;
+      })();
+      const nextLine = lines[li + 1];
+      const nextIsData = nextLine?.trim().startsWith("data:") ?? false;
+      const nextEmpty = (nextLine?.trim().length ?? 1) === 0;
+      if (looksComplete && (nextIsData || nextEmpty || li === lines.length - 1)) {
+        const res = flushAccumulator();
+        objects.push(...res.parsed);
+        if (res.leftover) {
+          jsonAccumulator = res.leftover;
+        } else {
+          jsonAccumulator = "";
+          insideDataBlock = false;
+        }
+      }
+    }
+  }
+
+  if (jsonAccumulator.trim().length > 0) {
+    const res = flushAccumulator();
+    if (res.parsed.length > 0) {
+      objects.push(...res.parsed);
+      remainder = res.leftover;
+    } else {
+      remainder = jsonAccumulator;
+    }
+  } else {
+    remainder = "";
+  }
+
+  return { objects, remainder, errors };
+}
+
+/**
+ * Remainder-based SSE parser for cloudcode-pa streams. Pure function: feed each
+ * network chunk together with the previous remainder; it returns fully parsed
+ * payloads plus the new remainder. Tolerates UTF-8 splits, concatenated JSON,
+ * ndjson without SSE framing and `[DONE]` sentinels.
+ *
+ * Named `parseSSEPayloads` because {@link parseSSEChunk} already owns the
+ * event-block parser name in this module.
+ *
+ * @param chunk - new raw text or binary chunk
+ * @param previousRemainder - leftover text from the previous call
+ * @returns parsed payloads, new remainder, error notes and consumed length
+ *
+ * @example
+ * let remainder = "";
+ * for await (const raw of stream) {
+ *   const { payloads, remainder: next } = parseSSEPayloads(raw, remainder);
+ *   remainder = next;
+ *   for (const p of payloads) handle(p);
+ * }
+ */
+export function parseSSEPayloads(
+  chunk: string | Uint8Array,
+  previousRemainder = "",
+): { payloads: CloudCodeRawChunk[]; remainder: string; errors: string[]; consumed: number } {
+  let text = "";
+  if (typeof chunk === "string") {
+    text = chunk;
+  } else {
+    try {
+      text = Buffer.from(chunk).toString("utf8");
+    } catch {
+      text = new TextDecoder("utf-8", { fatal: false }).decode(chunk);
+    }
+  }
+
+  const combined = previousRemainder + text;
+  const { objects, remainder, errors } = safeJsonParseArray(combined);
+
+  // Unwrap an outer { response: {...} } layer when present
+  const payloads = objects.map((obj) => {
+    const maybe = (obj as any).response as CloudCodeRawChunk | undefined;
+    if (
+      maybe &&
+      typeof maybe === "object" &&
+      ((maybe as any).candidates || (maybe as any).content || (maybe as any).parts)
+    ) {
+      return maybe;
+    }
+    return obj;
+  });
+
+  return {
+    payloads,
+    remainder,
+    errors,
+    consumed: text.length,
+  };
+}
+
+// ---------------------------------------------------------------------------
+// Merged: normalizePayloads over typed raw chunks
+// ---------------------------------------------------------------------------
+
+function normalizeFinishReason(raw: string): string | undefined {
+  if (!raw) return undefined;
+  const upper = raw.toUpperCase();
+  if (upper.includes("STOP") || upper === "FINISH_REASON_STOP" || upper === "END") return "stop";
+  if (upper.includes("MAX_TOKENS") || upper.includes("LENGTH")) return "length";
+  if (upper.includes("SAFETY") || upper.includes("BLOCK")) return "content_filter";
+  if (upper.includes("RECITATION")) return "content_filter";
+  if (upper.includes("TOOL")) return "tool_calls";
+  if (upper.includes("FINISH_REASON_UNSPECIFIED")) return undefined;
+  return upper.toLowerCase();
+}
+
+function extractTextFromAny(obj: unknown): string | undefined {
+  if (!obj) return undefined;
+  if (typeof obj === "string") return obj;
+  if (typeof obj !== "object") return undefined;
+  const o = obj as Record<string, unknown>;
+  if (typeof o.text === "string") return o.text;
+  if (typeof o.content === "string") return o.content;
+  if (typeof o.message === "string") return o.message;
+  if (Array.isArray(o.parts)) {
+    const txt = (o.parts as any[]).map((p) => (typeof p === "string" ? p : ((p as any)?.text ?? ""))).join("");
+    if (txt) return txt;
+  }
+  return undefined;
+}
+
+/**
+ * Normalizes heterogeneous raw chunks into {@link NormalizedPayload} arrays:
+ * unwraps `{response:{...}}`, builds virtual candidates for content/parts/text
+ * shapes, classifies parts into text/thinking/function_call/inline_data/
+ * thought_signature kinds and carries finish reasons and usage metadata.
+ *
+ * Named `normalizeRawPayloads` because the string-based {@link normalizePayloads}
+ * keeps its historical name in this module.
+ *
+ * @param rawChunks - typed raw chunks (output of parseSSEPayloads)
+ * @param opts - model fallback and candidate index fallback
+ * @returns normalized payloads
+ */
+export function normalizeRawPayloads(
+  rawChunks: CloudCodeRawChunk[],
+  opts?: {
+    model?: string;
+    candidateIndexFallback?: number;
+  },
+): NormalizedPayload[] {
+  const modelFallback = opts?.model ?? UA_DEFAULT_MODEL;
+  const out: NormalizedPayload[] = [];
+
+  let globalIndex = opts?.candidateIndexFallback ?? 0;
+
+  for (let rawIndex = 0; rawIndex < rawChunks.length; rawIndex++) {
+    const raw = rawChunks[rawIndex]!;
+    if (!raw || typeof raw !== "object") continue;
+
+    // Guarantee unwrapping even if the parser layer was bypassed
+    const unwrapped: CloudCodeRawChunk =
+      (raw as any).response && typeof (raw as any).response === "object"
+        ? ((raw as any).response as CloudCodeRawChunk)
+        : raw;
+
+    const candidatesList: RawCandidate[] = [];
+
+    if (Array.isArray(unwrapped.candidates) && unwrapped.candidates.length > 0) {
+      candidatesList.push(...unwrapped.candidates);
+    } else if (unwrapped.content) {
+      candidatesList.push({ content: unwrapped.content, index: 0 });
+    } else if (Array.isArray(unwrapped.parts) && unwrapped.parts.length > 0) {
+      candidatesList.push({ content: { parts: unwrapped.parts }, index: 0 });
+    } else if (typeof (unwrapped as any).text === "string") {
+      candidatesList.push({ content: { parts: [{ text: (unwrapped as any).text as string }] }, index: 0 });
+    } else if (typeof (unwrapped as any).thought === "string") {
+      candidatesList.push({
+        content: {
+          parts: [
+            {
+              text: (unwrapped as any).thought as string,
+              thought: true,
+              thoughtSignature: (unwrapped as any).thoughtSignature as string | undefined,
+            } as RawPart,
+          ],
+        },
+        index: 0,
+      });
+    } else {
+      const txt = extractTextFromAny(unwrapped);
+      if (txt) {
+        candidatesList.push({ content: { parts: [{ text: txt }] }, index: 0 });
+      } else {
+        continue;
+      }
+    }
+
+    for (let ci = 0; ci < candidatesList.length; ci++) {
+      const cand = candidatesList[ci]!;
+      const candIndex = typeof cand.index === "number" ? cand.index : ci;
+      const finishReasonRaw = (cand.finishReason || (unwrapped as any).finishReason || "") as string;
+      const finishReason = normalizeFinishReason(finishReasonRaw);
+
+      let rawParts: RawPart[] = [];
+
+      if (cand.content?.parts && Array.isArray(cand.content.parts)) {
+        rawParts = cand.content.parts;
+      } else if ((cand as any).parts && Array.isArray((cand as any).parts)) {
+        rawParts = (cand as any).parts;
+      } else if (typeof cand.text === "string" && cand.text.length > 0) {
+        rawParts = [
+          { text: cand.text, thought: !!(cand as any).thought, thoughtSignature: (cand as any).thoughtSignature },
+        ];
+      } else if (cand.content?.text && typeof cand.content.text === "string") {
+        rawParts = [{ text: cand.content.text }];
+      } else if (typeof (cand.content as any)?.text === "string") {
+        rawParts = [{ text: (cand.content as any).text as string }];
+      }
+
+      const normalizedParts: NormalizedPart[] = [];
+      let isThinkingChunk = false;
+
+      for (let pi = 0; pi < rawParts.length; pi++) {
+        const rp = rawParts[pi]!;
+        if (!rp || typeof rp !== "object") continue;
+
+        const txt = typeof rp.text === "string" ? rp.text : extractTextFromAny(rp);
+        const isThought = rp.thought === true || (rp as any).isThought === true;
+        const sig =
+          typeof rp.thoughtSignature === "string"
+            ? rp.thoughtSignature
+            : typeof (rp as any).thought_signature === "string"
+              ? ((rp as any).thought_signature as string)
+              : undefined;
+
+        if (isThought) isThinkingChunk = true;
+
+        if (rp.functionCall && typeof rp.functionCall === "object") {
+          const fc = rp.functionCall as { name: string; args?: unknown; id?: string };
+          const argsStr = typeof fc.args === "string" ? fc.args : JSON.stringify(fc.args ?? {});
+          normalizedParts.push({
+            kind: "function_call",
+            text: "",
+            thought: false,
+            thoughtSignature: sig,
+            functionCall: {
+              name: fc.name || "unknown",
+              args: argsStr,
+              id: fc.id || `call_${globalIndex}_${ci}_${pi}_${Date.now().toString(36)}`,
+            },
+            rawIndex,
+            candidateIndex: candIndex,
+            finishReason,
+          });
+          continue;
+        }
+
+        if (rp.inlineData) {
+          normalizedParts.push({
+            kind: "inline_data",
+            text: "",
+            thought: false,
+            thoughtSignature: sig,
+            rawIndex,
+            candidateIndex: candIndex,
+            finishReason,
+          });
+          continue;
+        }
+
+        if (!txt && sig) {
+          normalizedParts.push({
+            kind: "thought_signature",
+            text: "",
+            thought: true,
+            thoughtSignature: sig,
+            rawIndex,
+            candidateIndex: candIndex,
+            finishReason,
+          });
+          continue;
+        }
+
+        if (!txt && !sig) continue;
+
+        normalizedParts.push({
+          kind: isThought ? "thinking" : "text",
+          text: txt ?? "",
+          thought: isThought,
+          thoughtSignature: sig,
+          rawIndex,
+          candidateIndex: candIndex,
+          finishReason,
+        });
+      }
+
+      const hasFinish = !!finishReason;
+
+      if (normalizedParts.length === 0 && !hasFinish) {
+        // Usage-only chunk
+        if (unwrapped.usageMetadata) {
+          out.push({
+            parts: [],
+            finishReason: undefined,
+            usage: unwrapped.usageMetadata,
+            model: unwrapped.model ?? modelFallback,
+            isThinkingChunk: false,
+            hasFinish: false,
+            raw: unwrapped,
+          });
+        }
+        continue;
+      }
+
+      out.push({
+        parts: normalizedParts,
+        finishReason: finishReason || undefined,
+        usage: unwrapped.usageMetadata,
+        model: unwrapped.model ?? modelFallback,
+        isThinkingChunk,
+        hasFinish,
+        raw: unwrapped,
+      });
+
+      globalIndex++;
+    }
+  }
+
+  return out;
+}
+
+// ---------------------------------------------------------------------------
+// Merged: stateful OpenAI / Anthropic transforms over normalized payloads
+// ---------------------------------------------------------------------------
+
+function newCompletionId(): string {
+  try {
+    const uuid = randomUUID();
+    return `chatcmpl-${uuid.replace(/-/g, "").slice(0, 29)}`;
+  } catch {
+    return `chatcmpl-${Math.random().toString(36).slice(2)}${Date.now().toString(36)}`;
+  }
+}
+
+/**
+ * Creates the mutable state shared by the merged transformers.
+ *
+ * @param model - model id echoed back to clients
+ * @param completionId - optional fixed completion id
+ */
+export function createInitialStreamingState(model: string, completionId?: string): StreamingState {
+  return {
+    completionId: completionId ?? newCompletionId(),
+    model,
+    created: Math.floor(Date.now() / 1000),
+    sentRole: false,
+    accumulatedText: "",
+    accumulatedThinking: "",
+    lastFinishReason: undefined,
+    insideThinking: false,
+    anthropicBlockIndex: 0,
+    anthropicBlockKinds: new Map(),
+    toolCallIdCounter: 0,
+    pendingSignatures: [],
+  };
+}
+
+/**
+ * Transforms normalized payloads into OpenAI chat.completion.chunk objects.
+ * Emits a leading role chunk, routes thinking into reasoning_content/reasoning,
+ * caches thought signatures in the global entry-based LRU and streams tool calls.
+ *
+ * @param payloads - normalized payloads
+ * @param state - mutable streaming state updated in place
+ * @returns OpenAI chunk objects (serialize with serializeOpenAIChunksToSSE)
+ */
+export function transformToOpenAIChunks(payloads: NormalizedPayload[], state: StreamingState): OpenAIChatChunk[] {
+  const out: OpenAIChatChunk[] = [];
+
+  for (const p of payloads) {
+    if (p.finishReason) state.lastFinishReason = p.finishReason;
+
+    if (p.parts.length === 0 && p.usage) {
+      continue;
+    }
+
+    for (const part of p.parts) {
+      if (!state.sentRole) {
+        out.push({
+          id: state.completionId,
+          object: "chat.completion.chunk",
+          created: state.created,
+          model: state.model,
+          choices: [
+            {
+              index: part.candidateIndex,
+              delta: { role: "assistant" },
+              finish_reason: null,
+            },
+          ],
+        });
+        state.sentRole = true;
+      }
+
+      if (part.thoughtSignature) {
+        try {
+          globalThinkingSignatureLRUCache.addFromPayload({
+            signature: part.thoughtSignature,
+            thinkingText: part.kind === "thinking" ? part.text : state.accumulatedThinking.slice(-2000),
+            model: p.model ?? state.model,
+            turnId: `${state.completionId}_${part.candidateIndex}`,
+          });
+        } catch {
+          // Cache failures must never break streaming
+        }
+        if (!state.pendingSignatures.includes(part.thoughtSignature)) {
+          state.pendingSignatures.push(part.thoughtSignature);
+          if (state.pendingSignatures.length > 50) state.pendingSignatures.shift();
+        }
+      }
+
+      if (part.kind === "function_call" && part.functionCall) {
+        out.push({
+          id: state.completionId,
+          object: "chat.completion.chunk",
+          created: state.created,
+          model: state.model,
+          choices: [
+            {
+              index: part.candidateIndex,
+              delta: {
+                tool_calls: [
+                  {
+                    index: state.toolCallIdCounter++,
+                    id: part.functionCall.id,
+                    type: "function",
+                    function: {
+                      name: part.functionCall.name,
+                      arguments: part.functionCall.args,
+                    },
+                  },
+                ],
+              },
+              finish_reason: null,
+            },
+          ],
+        });
+        continue;
+      }
+
+      if (part.kind === "thinking" || part.thought) {
+        if (!state.insideThinking) {
+          state.insideThinking = true;
+        }
+        state.accumulatedThinking += part.text;
+
+        out.push({
+          id: state.completionId,
+          object: "chat.completion.chunk",
+          created: state.created,
+          model: state.model,
+          choices: [
+            {
+              index: part.candidateIndex,
+              delta: {
+                reasoning_content: part.text,
+                reasoning: part.text,
+              },
+              finish_reason: null,
+            },
+          ],
+        });
+      } else if (part.kind === "text") {
+        if (state.insideThinking) {
+          state.insideThinking = false;
+        }
+        state.accumulatedText += part.text;
+
+        out.push({
+          id: state.completionId,
+          object: "chat.completion.chunk",
+          created: state.created,
+          model: state.model,
+          choices: [
+            {
+              index: part.candidateIndex,
+              delta: {
+                content: part.text,
+              },
+              finish_reason: null,
+            },
+          ],
+        });
+      } else if (part.kind === "thought_signature") {
+        continue;
+      }
+    }
+
+    if (p.hasFinish && p.finishReason) {
+      // Finish reason recorded; final emission handled by synthetic finish helpers
+    }
+  }
+
+  return out;
+}
+
+/**
+ * Builds the final OpenAI finish chunk (with optional usage mapping) once the
+ * stream ends. The caller serializes it and appends the `[DONE]` marker.
+ *
+ * @param state - streaming state carrying the last observed finish reason
+ * @param reason - fallback finish reason
+ * @param usage - usage metadata captured from the stream
+ */
+export function createSyntheticFinishChunks(
+  state: StreamingState,
+  reason: string = "stop",
+  usage?: NormalizedPayload["usage"],
+): OpenAIChatChunk[] {
+  const finish = state.lastFinishReason ?? reason;
+  const normalized = normalizeFinishReason(finish) ?? "stop";
+
+  const chunk: OpenAIChatChunk = {
+    id: state.completionId,
+    object: "chat.completion.chunk",
+    created: state.created,
+    model: state.model,
+    choices: [
+      {
+        index: 0,
+        delta: {},
+        finish_reason: normalized,
+      },
+    ],
+  };
+
+  if (usage) {
+    (chunk as any).usage = {
+      prompt_tokens: usage.promptTokenCount,
+      completion_tokens: usage.candidatesTokenCount ?? usage.totalTokenCount,
+      total_tokens: usage.totalTokenCount,
+      completion_tokens_details: usage.thoughtsTokenCount ? { reasoning_tokens: usage.thoughtsTokenCount } : undefined,
+    };
+  }
+
+  return [chunk];
+}
+
+/**
+ * Serializes OpenAI chunks into SSE data lines ready for the wire.
+ *
+ * @param chunks - OpenAI chunk objects
+ * @returns array of `data: {...}\n\n` strings
+ */
+export function serializeOpenAIChunksToSSE(chunks: OpenAIChatChunk[]): string[] {
+  const lines: string[] = [];
+  for (const c of chunks) {
+    lines.push(`data: ${JSON.stringify(c)}\n\n`);
+  }
+  return lines;
+}
+
+/**
+ * Transforms normalized payloads into Anthropic Messages SSE lines:
+ * message_start, content_block_start/delta/stop and message_delta/stop.
+ * Opens separate blocks when thinking interleaves with text so block ordering
+ * satisfies Claude validation.
+ *
+ * @param payloads - normalized payloads
+ * @param state - mutable streaming state updated in place
+ * @returns formatted SSE strings (`event: ...\ndata: {...}\n\n`)
+ */
+export function transformToAnthropicEvents(payloads: NormalizedPayload[], state: StreamingState): string[] {
+  const sseLines: string[] = [];
+
+  const emit = (event: string, data: unknown): void => {
+    sseLines.push(`event: ${event}\ndata: ${JSON.stringify(data)}\n\n`);
+  };
+
+  if (state.anthropicBlockIndex === 0 && !state.sentRole) {
+    emit("message_start", {
+      type: "message_start",
+      message: {
+        id: state.completionId.replace("chatcmpl-", "msg_"),
+        type: "message",
+        role: "assistant",
+        content: [],
+        model: state.model,
+        stop_reason: null,
+        stop_sequence: null,
+        usage: { input_tokens: 0, output_tokens: 0 },
+      },
+    });
+    state.sentRole = true;
+  }
+
+  for (const p of payloads) {
+    if (p.finishReason) state.lastFinishReason = p.finishReason;
+
+    for (const part of p.parts) {
+      if (part.thoughtSignature) {
+        try {
+          globalThinkingSignatureLRUCache.addFromPayload({
+            signature: part.thoughtSignature,
+            thinkingText: part.text || state.accumulatedThinking.slice(-2000),
+            model: p.model ?? state.model,
+            turnId: `${state.completionId}_${part.candidateIndex}`,
+          });
+        } catch {
+          /* the guarded best-effort operation falls through: the outer flow owns the failure */
+        }
+      }
+
+      if (part.kind === "function_call" && part.functionCall) {
+        const blockIndex = state.anthropicBlockIndex++;
+        state.anthropicBlockKinds.set(blockIndex, "function_call");
+
+        emit("content_block_start", {
+          type: "content_block_start",
+          index: blockIndex,
+          content_block: {
+            type: "tool_use",
+            id: part.functionCall.id,
+            name: part.functionCall.name,
+            input: {},
+          },
+        });
+
+        emit("content_block_delta", {
+          type: "content_block_delta",
+          index: blockIndex,
+          delta: {
+            type: "input_json_delta",
+            partial_json: part.functionCall.args,
+          },
+        });
+
+        emit("content_block_stop", {
+          type: "content_block_stop",
+          index: blockIndex,
+        });
+        continue;
+      }
+
+      const isThinking = part.kind === "thinking" || part.thought;
+      const wantedKind: NormalizedPartKind = isThinking ? "thinking" : "text";
+
+      let currentBlockIndex = state.anthropicBlockIndex;
+      const currentKind = state.anthropicBlockKinds.get(currentBlockIndex - 1);
+
+      const needNewBlock =
+        currentBlockIndex === 0 ||
+        (currentKind !== undefined && currentKind !== wantedKind && currentKind !== part.kind) ||
+        state.insideThinking !== isThinking;
+
+      if (needNewBlock) {
+        if (currentBlockIndex > 0 && currentKind !== undefined) {
+          if (currentKind !== wantedKind) {
+            emit("content_block_stop", {
+              type: "content_block_stop",
+              index: currentBlockIndex - 1,
+            });
+          }
+        }
+
+        if (currentKind !== wantedKind || currentBlockIndex === 0) {
+          const idx = state.anthropicBlockIndex;
+          state.anthropicBlockKinds.set(idx, wantedKind);
+          state.anthropicBlockIndex++;
+
+          if (isThinking) {
+            emit("content_block_start", {
+              type: "content_block_start",
+              index: idx,
+              content_block: { type: "thinking", thinking: "" },
+            });
+            state.insideThinking = true;
+          } else {
+            emit("content_block_start", {
+              type: "content_block_start",
+              index: idx,
+              content_block: { type: "text", text: "" },
+            });
+            state.insideThinking = false;
+          }
+          currentBlockIndex = idx;
+        }
+      }
+
+      if (isThinking) {
+        state.accumulatedThinking += part.text;
+        const idx = state.anthropicBlockIndex - 1;
+        emit("content_block_delta", {
+          type: "content_block_delta",
+          index: idx,
+          delta: { type: "thinking_delta", thinking: part.text },
+        });
+      } else if (part.kind === "text") {
+        state.accumulatedText += part.text;
+        const idx = state.anthropicBlockIndex - 1;
+        emit("content_block_delta", {
+          type: "content_block_delta",
+          index: idx,
+          delta: { type: "text_delta", text: part.text },
+        });
+      }
+    }
+  }
+
+  return sseLines;
+}
+
+/**
+ * Emits the final Anthropic events (block stop, message_delta with stop_reason,
+ * message_stop) once the stream ends.
+ *
+ * @param state - streaming state
+ * @param reason - fallback stop reason
+ * @param usage - usage metadata captured from the stream
+ * @returns formatted SSE strings
+ */
+export function createAnthropicSyntheticFinish(
+  state: StreamingState,
+  reason: string = "end_turn",
+  usage?: NormalizedPayload["usage"],
+): string[] {
+  const lines: string[] = [];
+
+  const emit = (event: string, data: unknown) => {
+    lines.push(`event: ${event}\ndata: ${JSON.stringify(data)}\n\n`);
+  };
+
+  if (state.anthropicBlockIndex > 0) {
+    const lastIdx = state.anthropicBlockIndex - 1;
+    if (state.anthropicBlockKinds.has(lastIdx)) {
+      emit("content_block_stop", { type: "content_block_stop", index: lastIdx });
+    }
+  }
+
+  emit("message_delta", {
+    type: "message_delta",
+    delta: {
+      stop_reason: state.lastFinishReason ? mapFinishToAnthropic(state.lastFinishReason) : reason,
+      stop_sequence: null,
+    },
+    usage: usage
+      ? {
+          output_tokens: usage.candidatesTokenCount ?? usage.totalTokenCount ?? 0,
+        }
+      : { output_tokens: 0 },
+  });
+
+  emit("message_stop", { type: "message_stop" });
+
+  return lines;
+}
+
+function mapFinishToAnthropic(finish: string): string {
+  const n = finish.toLowerCase();
+  if (n === "stop" || n.includes("stop")) return "end_turn";
+  if (n === "length" || n.includes("max")) return "max_tokens";
+  if (n === "tool_calls") return "tool_use";
+  if (n === "content_filter") return "end_turn";
+  return "end_turn";
+}
+
+// ---------------------------------------------------------------------------
+// Merged: incremental SSE parser class + high-level transformer
+// ---------------------------------------------------------------------------
+
+/**
+ * Stateful incremental SSE parser wrapping parseSSEPayloads with streaming
+ * UTF-8 decoding and byte accounting.
+ */
+export class IncrementalSSEParser {
+  private buffer: string;
+  private readonly decoder: { decode(input?: Uint8Array, options?: { stream?: boolean }): string };
+  private totalBytes: number;
+
+  constructor() {
+    this.buffer = "";
+    this.decoder = new TextDecoder("utf-8", { fatal: false });
+    this.totalBytes = 0;
+  }
+
+  /**
+   * Pushes an incremental chunk (string, Uint8Array or Buffer), tolerating
+   * splits anywhere, and returns payloads parsed so far.
+   */
+  push(chunk: string | Uint8Array | Buffer): CloudCodeRawChunk[] {
+    let text: string;
+    if (typeof chunk === "string") {
+      text = chunk;
+    } else if (chunk instanceof Uint8Array || Buffer.isBuffer(chunk)) {
+      text = this.decoder.decode(chunk, { stream: true });
+    } else {
+      text = String(chunk);
+    }
+
+    this.totalBytes += text.length;
+    const result = parseSSEPayloads(text, this.buffer);
+    this.buffer = result.remainder;
+    return result.payloads;
+  }
+
+  /**
+   * Final flush: decodes any pending bytes and attempts to parse the remainder.
+   */
+  flush(): CloudCodeRawChunk[] {
+    if (!this.buffer.trim()) {
+      this.buffer = "";
+      return [];
+    }
+    try {
+      const finalText = this.decoder.decode();
+      if (finalText) this.buffer += finalText;
+    } catch {
+      /* the guarded best-effort operation falls through: the outer flow owns the failure */
+    }
+    const result = parseSSEPayloads("", this.buffer);
+    this.buffer = result.remainder;
+    if (result.payloads.length > 0) {
+      this.buffer = "";
+    }
+    return result.payloads;
+  }
+
+  get remainder(): string {
+    return this.buffer;
+  }
+
+  get bytesReceived(): number {
+    return this.totalBytes;
+  }
+
+  reset(): void {
+    this.buffer = "";
+    this.totalBytes = 0;
+  }
+}
+
+/** Output format selector for {@link CloudCodeStreamingTransformer}. */
+export type OutputFormat = "openai" | "anthropic" | "raw";
+
+/**
+ * Options for the high-level streaming transformer.
+ */
+export interface StreamingTransformerOptions {
+  /** Target model id. */
+  model: string;
+  /** Wire format, openai by default. */
+  format?: OutputFormat;
+  /** Optional fixed completion id. */
+  completionId?: string;
+  /** External entry-based LRU cache (defaults to the global singleton). */
+  thinkingCache?: ThinkingSignatureLRUCache;
+  /** Inject a synthetic finish when the stream ends without one? Default true. */
+  injectSyntheticFinish?: boolean;
+  /** Append `[DONE]` after OpenAI SSE? Default true. */
+  includeDoneMarker?: boolean;
+  /** Callback invoked for every transformed SSE line. */
+  onChunk?: (sseLine: string, state: StreamingState) => void;
+}
+
+/**
+ * High-level cloudcode-pa SSE transformer: push raw network chunks in, receive
+ * serialized SSE lines out in the configured format, with synthetic finish
+ * injection and `[DONE]` handling on flush.
+ */
+export class CloudCodeStreamingTransformer {
+  private readonly parser: IncrementalSSEParser;
+  private readonly state: StreamingState;
+  private readonly options: Required<StreamingTransformerOptions>;
+  private sentFinished: boolean;
+  private lastUsage?: CloudCodeRawChunk["usageMetadata"];
+
+  constructor(opts: StreamingTransformerOptions) {
+    this.parser = new IncrementalSSEParser();
+    this.state = createInitialStreamingState(opts.model, opts.completionId);
+    this.options = {
+      model: opts.model,
+      format: opts.format ?? "openai",
+      completionId: this.state.completionId,
+      thinkingCache: opts.thinkingCache ?? globalThinkingSignatureLRUCache,
+      injectSyntheticFinish: opts.injectSyntheticFinish ?? true,
+      includeDoneMarker: opts.includeDoneMarker ?? true,
+      onChunk: opts.onChunk ?? (() => {}),
+    } as Required<StreamingTransformerOptions>;
+    this.sentFinished = false;
+  }
+
+  get streamingState(): StreamingState {
+    return this.state;
+  }
+
+  /**
+   * Transforms one raw chunk into output SSE lines.
+   */
+  transformChunk(rawChunk: string | Uint8Array): string[] {
+    const parsed = this.parser.push(rawChunk);
+    if (parsed.length === 0) return [];
+
+    for (const r of parsed) {
+      if ((r as any).usageMetadata) this.lastUsage = (r as any).usageMetadata;
+    }
+
+    const normalized = normalizeRawPayloads(parsed, { model: this.options.model });
+
+    let sseLines: string[] = [];
+
+    if (this.options.format === "openai") {
+      const openAIChunks = transformToOpenAIChunks(normalized, this.state);
+      sseLines = serializeOpenAIChunksToSSE(openAIChunks);
+    } else if (this.options.format === "anthropic") {
+      sseLines = transformToAnthropicEvents(normalized, this.state);
+    } else {
+      for (const n of normalized) {
+        sseLines.push(`data: ${JSON.stringify(n)}\n\n`);
+      }
+    }
+
+    for (const line of sseLines) {
+      try {
+        this.options.onChunk(line, this.state);
+      } catch {
+        /* the guarded best-effort operation falls through: the outer flow owns the failure */
+      }
+    }
+
+    return sseLines;
+  }
+
+  /**
+   * Flushes buffered input and injects the synthetic finish plus `[DONE]`.
+   */
+  flush(): string[] {
+    const remainingParsed = this.parser.flush();
+    const extraLines: string[] = [];
+
+    if (remainingParsed.length > 0) {
+      for (const r of remainingParsed) {
+        if ((r as any).usageMetadata) this.lastUsage = (r as any).usageMetadata;
+      }
+      const normalized = normalizeRawPayloads(remainingParsed, { model: this.options.model });
+      if (this.options.format === "openai") {
+        const chunks = transformToOpenAIChunks(normalized, this.state);
+        extraLines.push(...serializeOpenAIChunksToSSE(chunks));
+      } else if (this.options.format === "anthropic") {
+        extraLines.push(...transformToAnthropicEvents(normalized, this.state));
+      } else {
+        for (const n of normalized) extraLines.push(`data: ${JSON.stringify(n)}\n\n`);
+      }
+    }
+
+    if (!this.sentFinished && this.options.injectSyntheticFinish) {
+      const finishLines = this.createFinish();
+      extraLines.push(...finishLines);
+      this.sentFinished = true;
+    }
+
+    if (this.options.format === "openai" && this.options.includeDoneMarker) {
+      extraLines.push("data: [DONE]\n\n");
+    }
+
+    for (const line of extraLines) {
+      try {
+        this.options.onChunk(line, this.state);
+      } catch {
+        /* the guarded best-effort operation falls through: the outer flow owns the failure */
+      }
+    }
+
+    return extraLines;
+  }
+
+  private createFinish(): string[] {
+    if (this.options.format === "openai") {
+      const chunks = createSyntheticFinishChunks(this.state, "stop", this.lastUsage);
+      return serializeOpenAIChunksToSSE(chunks);
+    } else if (this.options.format === "anthropic") {
+      return createAnthropicSyntheticFinish(this.state, "end_turn", this.lastUsage);
+    } else {
+      return [
+        `data: ${JSON.stringify({ finishReason: this.state.lastFinishReason ?? "stop", usage: this.lastUsage })}\n\n`,
+      ];
+    }
+  }
+
+  reset(model?: string): void {
+    this.parser.reset();
+    this.sentFinished = false;
+    this.lastUsage = undefined;
+    this.state.accumulatedText = "";
+    this.state.accumulatedThinking = "";
+    this.state.sentRole = false;
+    this.state.insideThinking = false;
+    this.state.lastFinishReason = undefined;
+    this.state.anthropicBlockIndex = 0;
+    this.state.anthropicBlockKinds.clear();
+    this.state.toolCallIdCounter = 0;
+    this.state.pendingSignatures = [];
+    if (model) {
+      this.state.model = model;
+      (this.options as any).model = model;
+    }
+    this.state.completionId = newCompletionId();
+    this.state.created = Math.floor(Date.now() / 1000);
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Merged: resilient fetch + transformed stream
+// ---------------------------------------------------------------------------
+
+/**
+ * Options for {@link resilientFetch}.
+ */
+export interface ResilientFetchOptions {
+  /** Max attempts including the first. */
+  maxRetries?: number;
+  /** Base delay in ms. */
+  baseDelayMs?: number;
+  /** Max delay in ms. */
+  maxDelayMs?: number;
+  /** Apply jitter? Default true. */
+  jitter?: boolean;
+  /** Status codes that trigger a retry. */
+  retryOn?: number[];
+  /** External abort signal. */
+  signal?: AbortSignal;
+  /** Per-attempt timeout in ms. */
+  timeoutMs?: number;
+}
+
+const DEFAULT_RETRY_ON = [429, 500, 502, 503, 504];
+
+/**
+ * Computes exponential delay with optional multiplicative jitter (0.7x-1.3x).
+ *
+ * @param attempt - zero-based attempt index
+ * @param base - base delay ms
+ * @param max - maximum delay ms
+ * @param jitter - apply jitter
+ */
+export function computeExponentialDelay(attempt: number, base: number, max: number, jitter: boolean): number {
+  let delay = base * Math.pow(2, attempt);
+  if (delay > max) delay = max;
+  if (jitter) {
+    const jitterFactor = 0.7 + Math.random() * 0.6;
+    delay = Math.floor(delay * jitterFactor);
+  }
+  return delay;
+}
+
+function sleepMs(ms: number): Promise<void> {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+/**
+ * Fetches streamGenerateContent with exponential backoff, per-attempt timeout
+ * and external-signal support, keeping Gemini CLI bypass headers intact.
+ *
+ * Equivalent sibling of {@link fetchWithExponentialRetry}; both are kept.
+ *
+ * @param url - endpoint URL
+ * @param init - RequestInit for fetch
+ * @param opts - retry options
+ * @returns Response with body stream
+ */
+export async function resilientFetch(
+  url: string,
+  init: RequestInit,
+  opts: ResilientFetchOptions = {},
+): Promise<Response> {
+  const maxRetries = opts.maxRetries ?? 3;
+  const baseDelay = opts.baseDelayMs ?? 500;
+  const maxDelay = opts.maxDelayMs ?? 5000;
+  const jitter = opts.jitter ?? true;
+  const retryOn = opts.retryOn ?? DEFAULT_RETRY_ON;
+
+  let lastError: Error | null = null;
+
+  for (let attempt = 0; attempt <= maxRetries; attempt++) {
+    const timeoutMs = opts.timeoutMs ?? 120_000;
+    const ac = new AbortController();
+    const timer = setTimeout(() => ac.abort(), timeoutMs);
+
+    const onAbortExternal = () => ac.abort();
+    if (opts.signal) {
+      if (opts.signal.aborted) {
+        clearTimeout(timer);
+        throw new Error("resilientFetch aborted by external signal");
+      }
+      opts.signal.addEventListener("abort", onAbortExternal, { once: true });
+    }
+
+    try {
+      const mergedInit: RequestInit = {
+        ...init,
+        signal: ac.signal,
+      };
+
+      const res = await fetch(url, mergedInit);
+      clearTimeout(timer);
+      if (opts.signal) opts.signal.removeEventListener("abort", onAbortExternal);
+
+      if (res.ok) return res;
+
+      if (isRetryableNetworkStatus(res.status, { retryableStatus: retryOn }) && attempt < maxRetries) {
+        lastError = new Error(`retryable status ${res.status}`);
+        try {
+          await res.text();
+        } catch {
+          /* the guarded best-effort operation falls through: the outer flow owns the failure */
+        }
+        const delay = computeExponentialDelay(attempt, baseDelay, maxDelay, jitter);
+        await sleepMs(delay);
+        continue;
+      }
+
+      return res;
+    } catch (e) {
+      clearTimeout(timer);
+      if (opts.signal) opts.signal.removeEventListener("abort", onAbortExternal);
+      const err = e as Error & { name?: string };
+      lastError = err;
+
+      // AbortError retries only when it came from the internal timeout
+      const isAbortTimeout = err.name === "AbortError";
+      if (isAbortTimeout && attempt < maxRetries) {
+        const delay = computeExponentialDelay(attempt, baseDelay, maxDelay, jitter);
+        await sleepMs(delay);
+        continue;
+      }
+
+      if (attempt < maxRetries) {
+        const delay = computeExponentialDelay(attempt, baseDelay, maxDelay, jitter);
+        await sleepMs(delay);
+        continue;
+      }
+
+      throw lastError;
+    }
+  }
+
+  throw lastError ?? new Error("resilientFetch failed after retries");
+}
+
+/**
+ * High-level wrapper consuming streamGenerateContent with auto-recovery and
+ * yielding SSE strings in the configured format. Tolerant to mid-stream breaks
+ * (synthetic finish), missing finish reasons and UTF-8/JSON splits.
+ *
+ * @example
+ * for await (const sseLine of createResilientTransformedStream(url, init, { model, format: "openai" })) {
+ *   res.write(sseLine);
+ * }
+ */
+export async function* createResilientTransformedStream(
+  url: string,
+  init: RequestInit,
+  transformerOpts: StreamingTransformerOptions,
+  fetchOpts: ResilientFetchOptions = {},
+): AsyncGenerator<string, void, unknown> {
+  const transformer = new CloudCodeStreamingTransformer(transformerOpts);
+
+  const response = await resilientFetch(url, init, fetchOpts);
+
+  if (!response.ok) {
+    let errBody = "";
+    try {
+      errBody = await response.text();
+    } catch {
+      /* the guarded best-effort operation falls through: the outer flow owns the failure */
+    }
+    throw new Error(`cloudcode-pa stream failed ${response.status}: ${errBody.slice(0, 500)}`);
+  }
+
+  if (!response.body) {
+    yield* transformer.flush();
+    return;
+  }
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder("utf-8", { fatal: false });
+
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      const sseLines = transformer.transformChunk(value);
+      for (const line of sseLines) yield line;
+    }
+
+    const finalLines = transformer.flush();
+    for (const line of finalLines) yield line;
+  } catch (_e) {
+    // Mid-stream failure: degrade gracefully by flushing what we have
+    try {
+      const finalLines = transformer.flush();
+      for (const line of finalLines) yield line;
+    } catch {
+      // ignore
+    }
+  } finally {
+    try {
+      await reader.cancel();
+    } catch {
+      /* the guarded best-effort operation falls through: the outer flow owns the failure */
+    }
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Merged: bypass headers + deterministic session id
+// ---------------------------------------------------------------------------
+
+/**
+ * Options for {@link buildGeminiCliBypassHeaders}.
+ */
+export interface BypassHeadersOptions {
+  model?: string;
+  accessToken?: string;
+  projectId?: string;
+  includeAuth?: boolean;
+}
+
+/**
+ * Builds headers fully compatible with official Gemini CLI traffic against
+ * cloudcode-pa, including the X-Goog-Request-Params project binding.
+ *
+ * @param opts - auth/project options
+ * @returns headers record ready for fetch
+ */
+export function buildGeminiCliBypassHeaders(opts: BypassHeadersOptions = {}): Record<string, string> {
+  const h: Record<string, string> = {
+    "User-Agent": GEMINI_CLI_USER_AGENT,
+    "X-Goog-Api-Client": X_GOOG_API_CLIENT,
+    "Client-Metadata": CLIENT_METADATA_RAW,
+    "Content-Type": "application/json",
+    Accept: "text/event-stream, application/json",
+    "X-Goog-Request-Params": `project=${opts.projectId ?? PROJECT_FALLBACK}`,
+  };
+
+  if (opts.includeAuth && opts.accessToken) {
+    h.Authorization = `Bearer ${opts.accessToken}`;
+  }
+
+  return h;
+}
+
+/**
+ * Deterministic session id (FNV-1a hex) derived from directory, hostname, cwd
+ * and pid. Streaming-module variant kept alongside the request-module version.
+ *
+ * @param directory - optional directory seed
+ * @returns 8-char hex digest
+ */
+export function buildHostSessionId(directory?: string): string {
+  const base = directory ?? `${os.hostname()}|${process.cwd()}|${process.pid}`;
+  return fnv1a32HexBytes(base);
+}
+
+// ---------------------------------------------------------------------------
+// Merged: plugin utilities
+// ---------------------------------------------------------------------------
+
+/**
+ * Detects whether any payload carries thinking content.
+ *
+ * @param payloads - normalized payloads
+ */
+export function hasThinkingContent(payloads: NormalizedPayload[]): boolean {
+  return payloads.some((p) => p.isThinkingChunk || p.parts.some((pt) => pt.thought));
+}
+
+/**
+ * Collects all distinct thinking signatures for multi-turn replay.
+ *
+ * @param payloads - normalized payloads
+ * @returns unique signature strings
+ */
+export function extractThinkingSignatures(payloads: NormalizedPayload[]): string[] {
+  const sigs: string[] = [];
+  for (const p of payloads) {
+    for (const part of p.parts) {
+      if (part.thoughtSignature && !sigs.includes(part.thoughtSignature)) sigs.push(part.thoughtSignature);
+    }
+  }
+  return sigs;
+}
+
+/**
+ * Exposes accumulated state metrics for observability in the opencode UI.
+ *
+ * @param state - streaming state
+ */
+export function getStreamingSummary(state: StreamingState): {
+  completionId: string;
+  model: string;
+  textChars: number;
+  thinkingChars: number;
+  finishReason?: string | undefined;
+  toolCalls: number;
+  pendingSignatures: number;
+} {
+  return {
+    completionId: state.completionId,
+    model: state.model,
+    textChars: state.accumulatedText.length,
+    thinkingChars: state.accumulatedThinking.length,
+    finishReason: state.lastFinishReason,
+    toolCalls: state.toolCallIdCounter,
+    pendingSignatures: state.pendingSignatures.length,
+  };
+}
+
+/**
+ * Passes through the raw candidate payload when present, otherwise rebuilds a
+ * minimal Gemini-shaped response from the normalized view.
+ *
+ * @param p - normalized payload-like object
+ * @returns raw payload or reconstructed Gemini candidates envelope
+ */
+export const transformToGemini = (p: any): any =>
+  p.raw || { candidates: [{ content: { parts: [{ text: p.text || "" }] } }] };
+
+// ---------------------------------------------------------------------------
+// Merged: Streaming named barrel (default export remains streamingModule)
+// ---------------------------------------------------------------------------
+
+/**
+ * Named bundle mirroring the legacy ONDA 4 module surface. Renamed members are
+ * wired to their merged equivalents: `parseSSEPayloads`, `normalizeRawPayloads`
+ * and the module-level `isRetryableNetworkStatus`.
+ */
+export const Streaming = {
+  IncrementalSSEParser,
+  CloudCodeStreamingTransformer,
+  ThinkingSignatureLRUCache,
+  globalThinkingSignatureCache,
+  globalThinkingSignatureLRUCache,
+
+  parseSSEPayloads,
+  normalizeRawPayloads,
+
+  transformToOpenAIChunks,
+  transformToAnthropicEvents,
+  serializeOpenAIChunksToSSE,
+  createSyntheticFinishChunks,
+  createAnthropicSyntheticFinish,
+  createInitialStreamingState,
+
+  resilientFetch,
+  createResilientTransformedStream,
+  computeExponentialDelay,
+  isRetryableNetworkStatus,
+
+  buildGeminiCliBypassHeaders,
+  buildHostSessionId,
+
+  hasThinkingContent,
+  extractThinkingSignatures,
+  getStreamingSummary,
+  transformToGemini,
+  handleThinkingBlock,
+
+  GEMINI_CLI_CLIENT_ID,
+  GEMINI_CLI_CLIENT_SECRET,
+  GEMINI_CLI_USER_AGENT,
+  X_GOOG_API_CLIENT,
+  CLIENT_METADATA_RAW,
+  PROJECT_FALLBACK,
+  MODELS_2026,
+} as const;
+
+// ---------------------------------------------------------------------------
+// Exports — library-first barrel
+// ---------------------------------------------------------------------------
+
+/**
+ * Default export for barrel compatibility: contains all primary helpers
+ */
+const streamingModule = {
+  LRU_THINKING_CACHE_SIZE,
+  ReplaySignatureCache,
+  globalThinkingSignatureCache,
+  createSSEParserState,
+  parseSSEChunk,
+  parseSSEBinaryChunk,
+  safeJsonParse,
+  unwrapCandidates,
+  extractParts,
+  normalizePart,
+  normalizePayloads,
+  mapFinishReason,
+  extractDelta,
+  StreamingSessionTracker,
+  transformToOpenAI,
+  createOpenAIStreamingTransformer,
+  transformToAnthropic,
+  createAnthropicStreamingTransformer,
+  fetchWithExponentialRetry,
+  injectSyntheticFinish,
+  withAutoRecoveryStreaming,
+  readSSEStream,
+  streamToOpenAI,
+  streamToAnthropic,
+  fnv1a32,
+  hashThinkingText,
+  getJitter,
+  sleepWithJitter,
+  SSE_DONE_SENTINEL,
+};
+
+export default streamingModule;
+
+/* ════════════════════════════════════════════════════════════════════
+   Section: the devthink workbench stream surface (the 1.1.16 ChatEvent
+   family absorbed by the grand merge; the chat event vocabulary bridges
+   onto the normalized streaming machinery above).
+   ════════════════════════════════════════════════════════════════════ */
+export type ChatEvent =
+  | { type: "start"; provider: string; model: string }
+  | { type: "text"; text: string }
+  | { type: "reasoning"; text: string }
+  | { type: "tool"; name: string; input: unknown }
+  | { type: "error"; message: string }
+  | { type: "finish"; reason?: string };
+
+export type StreamOptions = {
+  provider: string;
+  model: string;
+  signal?: AbortSignal | undefined;
+};
+
+function decodeJson(value: string): Record<string, unknown> | null {
+  try {
+    const parsed: unknown = JSON.parse(value);
+    return parsed && typeof parsed === "object" ? (parsed as Record<string, unknown>) : null;
+  } catch {
+    return null;
+  }
+}
+
+function extractSseFrames(buffer: { value: string }, incoming: string): string[] {
+  buffer.value += incoming;
+  const frames = buffer.value.split(/\r?\n\r?\n/);
+  buffer.value = frames.pop() || "";
+  return frames;
+}
+
+function extractData(frame: string): string | null {
+  const lines = frame.split(/\r?\n/).filter((line) => line.startsWith("data:"));
+  if (lines.length === 0) return null;
+  const value = lines
+    .map((line) => line.slice(5).trimStart())
+    .join("\n")
+    .trim();
+  return value && value !== "[DONE]" ? value : null;
+}
+
+function textAt(value: unknown): string {
+  return typeof value === "string" ? value : "";
+}
+
+function normalizeOpenAi(frame: Record<string, unknown>): ChatEvent[] {
+  const choices = Array.isArray(frame.choices) ? frame.choices : [];
+  const choice = choices[0] as Record<string, unknown> | undefined;
+  const delta = (choice?.delta || {}) as Record<string, unknown>;
+  const events: ChatEvent[] = [];
+  const reasoning = textAt(delta.reasoning_content || delta.reasoning);
+  const text = textAt(delta.content);
+  if (reasoning) events.push({ type: "reasoning", text: reasoning });
+  if (text) events.push({ type: "text", text });
+  const finish = choice?.finish_reason;
+  if (finish) events.push({ type: "finish", reason: textAt(finish) });
+  return events;
+}
+
+function normalizeAnthropic(frame: Record<string, unknown>): ChatEvent[] {
+  const type = textAt(frame.type);
+  if (type === "content_block_delta") {
+    const delta = (frame.delta || {}) as Record<string, unknown>;
+    const text = textAt(delta.text);
+    return text ? [{ type: "text", text }] : [];
+  }
+  if (type === "message_stop") return [{ type: "finish", reason: "stop" }];
+  if (type === "error") {
+    const error = (frame.error || {}) as Record<string, unknown>;
+    return [{ type: "error", message: textAt(error.message) || "Anthropic returned an error." }];
+  }
+  return [];
+}
+
+function normalizeGoogle(frame: Record<string, unknown>): ChatEvent[] {
+  const candidates = Array.isArray(frame.candidates) ? frame.candidates : [];
+  const content = ((candidates[0] as Record<string, unknown> | undefined)?.content || {}) as Record<string, unknown>;
+  const parts = Array.isArray(content.parts) ? content.parts : [];
+  const events: ChatEvent[] = [];
+  for (const part of parts) {
+    const item = part as Record<string, unknown>;
+    const text = textAt(item.text);
+    if (text) events.push({ type: "text", text });
+  }
+  const finish = textAt((candidates[0] as Record<string, unknown> | undefined)?.finishReason);
+  if (finish) events.push({ type: "finish", reason: finish });
+  return events;
+}
+
+export function normalizeFrame(provider: string, frame: Record<string, unknown>): ChatEvent[] {
+  const lower = provider.toLowerCase();
+  if (lower === "anthropic") return normalizeAnthropic(frame);
+  if (lower === "google") return normalizeGoogle(frame);
+  const data = (frame.data || frame) as Record<string, unknown>;
+  if (data !== frame && textAt(data.delta_content)) return [{ type: "text", text: textAt(data.delta_content) }];
+  if (data !== frame && textAt(data.reasoning_content))
+    return [{ type: "reasoning", text: textAt(data.reasoning_content) }];
+  if (data.error || frame.error)
+    return [{ type: "error", message: textAt(data.error || frame.error) || "The provider returned an error." }];
+  return normalizeOpenAi(frame);
+}
+
+export async function* parseEventStream(response: Response, options: StreamOptions): AsyncGenerator<ChatEvent> {
+  if (!response.body) throw new Error("The provider returned an empty stream.");
+  yield { type: "start", provider: options.provider, model: options.model };
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+  const buffer = { value: "" };
+  try {
+    while (true) {
+      const { done, value } = await reader.read();
+      if (done) break;
+      for (const frame of extractSseFrames(buffer, decoder.decode(value, { stream: true }))) {
+        const payload = extractData(frame);
+        if (!payload) continue;
+        const parsed = decodeJson(payload);
+        if (!parsed) continue;
+        for (const event of normalizeFrame(options.provider, parsed)) yield event;
+      }
+    }
+  } finally {
+    reader.releaseLock();
+  }
+  yield { type: "finish", reason: "stop" };
+}
+
+export async function collectEvents(events: AsyncIterable<ChatEvent>): Promise<{ text: string; reasoning: string }> {
+  let text = "";
+  let reasoning = "";
+  for await (const event of events) {
+    if (event.type === "text") text += event.text;
+    if (event.type === "reasoning") reasoning += event.text;
+    if (event.type === "error") throw new Error(event.message);
+  }
+  return { text, reasoning };
+}