universal-llm-client 4.2.0 → 4.5.0

package/src/stream-decoder.ts

@@ -1,361 +1,535 @@
-/**
- * Universal LLM Client v3 — Stream Decoder
- *
- * Pluggable interface for decoding raw LLM token streams into typed events.
- * Consumers select their strategy per-call: passthrough for raw speed,
- * standard-chat for structured tool calls, or interleaved-reasoning
- * for models that emit <think>/<progress> tags.
- */
-
-import type { LLMToolCall } from './interfaces.js';
-
-// ============================================================================
-// Decoded Event Types
-// ============================================================================
-
-/** Clean, typed events emitted by a stream decoder */
-export type DecodedEvent =
-    | { type: 'text'; content: string }
-    | { type: 'thinking'; content: string }
-    | { type: 'progress'; content: string }
-    | { type: 'tool_call'; calls: LLMToolCall[] };
-
-/** Callback invoked by the decoder as events become available */
-export type DecoderCallback = (event: DecodedEvent) => void;
-
-// ============================================================================
-// Decoder Interface
-// ============================================================================
-
-/**
- * Transform raw LLM tokens into clean typed events.
- *
- * Usage:
- *   const decoder = createDecoder('standard-chat', callback);
- *   for (const token of stream) decoder.push(token);
- *   decoder.flush();
- *   const clean = decoder.getCleanContent();
- */
-export interface StreamDecoder {
-    /** Feed a raw token from the LLM stream */
-    push(token: string): void;
-    /** Signal end of stream — flush any buffered state */
-    flush(): void;
-    /** Get the accumulated clean text (all structural tags stripped) */
-    getCleanContent(): string;
-    /** Get accumulated reasoning/thinking content (if any) */
-    getReasoning(): string | undefined;
-}
-
-// ============================================================================
-// Decoder Types
-// ============================================================================
-
-export type DecoderType = 'passthrough' | 'standard-chat' | 'interleaved-reasoning';
-
-// ============================================================================
-// Passthrough Decoder
-// ============================================================================
-
-/**
- * Bare-bones decoder for raw text completions.
- * No parsing, no tag awareness. All tokens → text events.
- */
-export class PassthroughDecoder implements StreamDecoder {
-    private content = '';
-    private readonly callback: DecoderCallback;
-
-    constructor(callback: DecoderCallback) {
-        this.callback = callback;
-    }
-
-    push(token: string): void {
-        this.content += token;
-        this.callback({ type: 'text', content: token });
-    }
-
-    flush(): void {
-        // Nothing to flush — all tokens emitted immediately
-    }
-
-    getCleanContent(): string {
-        return this.content;
-    }
-
-    getReasoning(): string | undefined {
-        return undefined;
-    }
-}
-
-// ============================================================================
-// Standard Chat Decoder
-// ============================================================================
-
-/**
- * Decoder for standard LLM chat patterns — text streaming with native
- * reasoning and structured API tool calls. No text-level tag parsing.
- *
- * Streamed tokens are clean text → emitted as `text` events.
- * Native reasoning tokens → accepted via `pushReasoning()`.
- * Structured tool calls → accepted via `pushToolCalls()`.
- */
-export class StandardChatDecoder implements StreamDecoder {
-    private content = '';
-    private reasoning = '';
-    private readonly callback: DecoderCallback;
-
-    constructor(callback: DecoderCallback) {
-        this.callback = callback;
-    }
-
-    push(token: string): void {
-        this.content += token;
-        this.callback({ type: 'text', content: token });
-    }
-
-    /** Feed native reasoning tokens from the provider */
-    pushReasoning(content: string): void {
-        this.reasoning += content;
-        this.callback({ type: 'thinking', content });
-    }
-
-    /** Feed structured tool calls from the provider API response */
-    pushToolCalls(calls: LLMToolCall[]): void {
-        this.callback({ type: 'tool_call', calls });
-    }
-
-    flush(): void {
-        // Nothing to flush — all events emitted as they arrive
-    }
-
-    getCleanContent(): string {
-        return this.content;
-    }
-
-    getReasoning(): string | undefined {
-        return this.reasoning || undefined;
-    }
-}
-
-// ============================================================================
-// Interleaved Reasoning Decoder
-// ============================================================================
-
-/**
- * Decoder for models that emit interleaved reasoning tags in text.
- * Parses <think>...</think> and <progress>...</progress> tags from the
- * raw token stream and emits typed events for each.
- *
- * Handles streaming where tags may be split across chunks.
- */
-export class InterleavedReasoningDecoder implements StreamDecoder {
-    private buffer = '';
-    private content = '';
-    private reasoning = '';
-    private readonly callback: DecoderCallback;
-    private inThink = false;
-    private inProgress = false;
-
-    constructor(callback: DecoderCallback) {
-        this.callback = callback;
-    }
-
-    push(token: string): void {
-        this.buffer += token;
-        this.processBuffer();
-    }
-
-    flush(): void {
-        // Emit any remaining buffer content as text
-        if (this.buffer.length > 0) {
-            if (this.inThink) {
-                this.reasoning += this.buffer;
-                this.callback({ type: 'thinking', content: this.buffer });
-            } else if (this.inProgress) {
-                this.callback({ type: 'progress', content: this.buffer });
-            } else {
-                this.content += this.buffer;
-                this.callback({ type: 'text', content: this.buffer });
-            }
-            this.buffer = '';
-        }
-    }
-
-    getCleanContent(): string {
-        return this.content;
-    }
-
-    getReasoning(): string | undefined {
-        return this.reasoning || undefined;
-    }
-
-    private processBuffer(): void {
-        let safety = 0;
-        while (this.buffer.length > 0 && safety++ < 200) {
-            if (this.inThink) {
-                const closeIdx = this.buffer.indexOf('</think>');
-                if (closeIdx === -1) {
-                    // Might have partial closing tag at end
-                    if (this.buffer.endsWith('<') || this.buffer.endsWith('</') ||
-                        this.buffer.endsWith('</t') || this.buffer.endsWith('</th') ||
-                        this.buffer.endsWith('</thi') || this.buffer.endsWith('</thin') ||
-                        this.buffer.endsWith('</think')) {
-                        return; // Wait for more data
-                    }
-                    this.reasoning += this.buffer;
-                    this.callback({ type: 'thinking', content: this.buffer });
-                    this.buffer = '';
-                    return;
-                }
-                const thinkContent = this.buffer.slice(0, closeIdx);
-                if (thinkContent) {
-                    this.reasoning += thinkContent;
-                    this.callback({ type: 'thinking', content: thinkContent });
-                }
-                this.buffer = this.buffer.slice(closeIdx + 8); // '</think>'.length
-                this.inThink = false;
-                continue;
-            }
-
-            if (this.inProgress) {
-                const closeIdx = this.buffer.indexOf('</progress>');
-                if (closeIdx === -1) {
-                    if (this.couldBePartialTag(this.buffer, '</progress>')) return;
-                    this.callback({ type: 'progress', content: this.buffer });
-                    this.buffer = '';
-                    return;
-                }
-                const progressContent = this.buffer.slice(0, closeIdx);
-                if (progressContent) {
-                    this.callback({ type: 'progress', content: progressContent });
-                }
-                this.buffer = this.buffer.slice(closeIdx + 11); // '</progress>'.length
-                this.inProgress = false;
-                continue;
-            }
-
-            // Look for opening tags
-            const thinkIdx = this.buffer.indexOf('<think>');
-            const progressIdx = this.buffer.indexOf('<progress>');
-
-            // Find earliest tag
-            const nextTag = this.findEarliest(thinkIdx, progressIdx);
-
-            if (nextTag === -1) {
-                // No complete opening tags — check for partial tag at end
-                const lastAngle = this.buffer.lastIndexOf('<');
-                if (lastAngle >= 0 && lastAngle > this.buffer.length - 12) {
-                    // Potential partial tag — emit text before it, keep the rest
-                    const textBefore = this.buffer.slice(0, lastAngle);
-                    if (textBefore) {
-                        this.content += textBefore;
-                        this.callback({ type: 'text', content: textBefore });
-                    }
-                    this.buffer = this.buffer.slice(lastAngle);
-                    return;
-                }
-                // No partial tags — emit all as text
-                this.content += this.buffer;
-                this.callback({ type: 'text', content: this.buffer });
-                this.buffer = '';
-                return;
-            }
-
-            // Emit text before the tag
-            const textBefore = this.buffer.slice(0, nextTag);
-            if (textBefore) {
-                this.content += textBefore;
-                this.callback({ type: 'text', content: textBefore });
-            }
-
-            if (nextTag === thinkIdx) {
-                this.buffer = this.buffer.slice(nextTag + 7); // '<think>'.length
-                this.inThink = true;
-            } else {
-                this.buffer = this.buffer.slice(nextTag + 10); // '<progress>'.length
-                this.inProgress = true;
-            }
-        }
-    }
-
-    private findEarliest(a: number, b: number): number {
-        if (a === -1) return b;
-        if (b === -1) return a;
-        return Math.min(a, b);
-    }
-
-    private couldBePartialTag(buffer: string, tag: string): boolean {
-        for (let i = 1; i < tag.length; i++) {
-            if (buffer.endsWith(tag.slice(0, i))) return true;
-        }
-        return false;
-    }
-}
-
-// ============================================================================
-// Pluggable Decoder Registry
-// ============================================================================
-
-export interface DecoderOptions {
-    /** Known tool names for text-based tool call recovery */
-    knownToolNames?: Set<string>;
-}
-
-/**
- * Factory function that creates a StreamDecoder instance.
- * External code registers these via `registerDecoder()`.
- */
-export type DecoderFactory = (callback: DecoderCallback, options?: DecoderOptions) => StreamDecoder;
-
-/** Internal registry of decoder factories, keyed by decoder type name */
-const decoderRegistry = new Map<string, DecoderFactory>();
-
-/**
- * Register a custom stream decoder type.
- * Once registered, it can be used via `createDecoder(name, ...)` or
- * by passing `decoderType: name` in ChatOptions.
- *
- * @example
- * ```typescript
- * import { registerDecoder } from 'universal-llm-client';
- *
- * registerDecoder('my-decoder', (callback, options) => {
- *   return new MyCustomDecoder(callback, options);
- * });
- * ```
- */
-export function registerDecoder(type: string, factory: DecoderFactory): void {
-    decoderRegistry.set(type, factory);
-}
-
-/**
- * Get all registered decoder type names.
- */
-export function getRegisteredDecoders(): string[] {
-    return Array.from(decoderRegistry.keys());
-}
-
-// Pre-register built-in decoders
-registerDecoder('passthrough', (cb) => new PassthroughDecoder(cb));
-registerDecoder('standard-chat', (cb) => new StandardChatDecoder(cb));
-registerDecoder('interleaved-reasoning', (cb) => new InterleavedReasoningDecoder(cb));
-
-/**
- * Create a stream decoder by type name.
- * Looks up the decoder in the registry (built-in + custom).
- *
- * @throws Error if the decoder type is not registered
- */
-export function createDecoder(
-    type: DecoderType | string,
-    callback: DecoderCallback,
-    options?: DecoderOptions,
-): StreamDecoder {
-    const factory = decoderRegistry.get(type);
-    if (!factory) {
-        const available = Array.from(decoderRegistry.keys()).join(', ');
-        throw new Error(`Unknown decoder type: "${type}". Available: ${available}`);
-    }
-    return factory(callback, options);
-}
-
+/**
+ * Universal LLM Client v3 — Stream Decoder
+ *
+ * Pluggable interface for decoding raw LLM token streams into typed events.
+ * Consumers select their strategy per-call: passthrough for raw speed,
+ * standard-chat for structured tool calls, or interleaved-reasoning
+ * for models that emit <think>/<progress> tags.
+ */
+
+import type { LLMToolCall } from './interfaces.js';
+import { GEMMA_THOUGHT_OPENERS, normalizeGemmaThought } from './gemma-channel.js';
+
+// ============================================================================
+// Decoded Event Types
+// ============================================================================
+
+/** Clean, typed events emitted by a stream decoder */
+export type DecodedEvent =
+    | { type: 'text'; content: string }
+    | { type: 'thinking'; content: string }
+    | { type: 'progress'; content: string }
+    | { type: 'tool_call'; calls: LLMToolCall[] };
+
+/** Callback invoked by the decoder as events become available */
+export type DecoderCallback = (event: DecodedEvent) => void;
+
+// ============================================================================
+// Decoder Interface
+// ============================================================================
+
+/**
+ * Transform raw LLM tokens into clean typed events.
+ *
+ * Usage:
+ *   const decoder = createDecoder('standard-chat', callback);
+ *   for (const token of stream) decoder.push(token);
+ *   decoder.flush();
+ *   const clean = decoder.getCleanContent();
+ */
+export interface StreamDecoder {
+    /** Feed a raw token from the LLM stream */
+    push(token: string): void;
+    /** Signal end of stream — flush any buffered state */
+    flush(): void;
+    /** Get the accumulated clean text (all structural tags stripped) */
+    getCleanContent(): string;
+    /** Get accumulated reasoning/thinking content (if any) */
+    getReasoning(): string | undefined;
+}
+
+// ============================================================================
+// Decoder Types
+// ============================================================================
+
+export type DecoderType = 'passthrough' | 'standard-chat' | 'interleaved-reasoning';
+
+// ============================================================================
+// Passthrough Decoder
+// ============================================================================
+
+/**
+ * Bare-bones decoder for raw text completions.
+ * No parsing, no tag awareness. All tokens → text events.
+ */
+export class PassthroughDecoder implements StreamDecoder {
+    private content = '';
+    private readonly callback: DecoderCallback;
+
+    constructor(callback: DecoderCallback) {
+        this.callback = callback;
+    }
+
+    push(token: string): void {
+        this.content += token;
+        this.callback({ type: 'text', content: token });
+    }
+
+    flush(): void {
+        // Nothing to flush — all tokens emitted immediately
+    }
+
+    getCleanContent(): string {
+        return this.content;
+    }
+
+    getReasoning(): string | undefined {
+        return undefined;
+    }
+}
+
+// ============================================================================
+// Standard Chat Decoder
+// ============================================================================
+
+/**
+ * Decoder for standard LLM chat patterns — text streaming with native
+ * reasoning and structured API tool calls. No text-level tag parsing.
+ *
+ * Streamed tokens are clean text → emitted as `text` events.
+ * Native reasoning tokens → accepted via `pushReasoning()`.
+ * Structured tool calls → accepted via `pushToolCalls()`.
+ */
+export class StandardChatDecoder implements StreamDecoder {
+    private content = '';
+    private reasoning = '';
+    private readonly callback: DecoderCallback;
+    private tagBuffer = '';
+    private inProgressTag = false;
+    private progressBody = '';
+    private inGemmaThought = false;
+    private gemmaThoughtBody = '';
+    private gemmaThoughtClose = '';
+    private inToolCallTag = false;
+    private toolCallBody = '';
+    private toolCallClose = '';
+
+    constructor(callback: DecoderCallback) {
+        this.callback = callback;
+    }
+
+    push(token: string): void {
+        let pos = 0;
+
+        while (pos < token.length) {
+            if (this.inGemmaThought) {
+                this.gemmaThoughtBody += token.slice(pos);
+                const closeIdx = this.gemmaThoughtBody.indexOf(this.gemmaThoughtClose);
+                if (closeIdx !== -1) {
+                    const body = this.gemmaThoughtBody.slice(0, closeIdx);
+                    const remainder = this.gemmaThoughtBody.slice(closeIdx + this.gemmaThoughtClose.length);
+                    this.emitReasoning(normalizeGemmaThought(body));
+                    this.inGemmaThought = false;
+                    this.gemmaThoughtBody = '';
+                    this.gemmaThoughtClose = '';
+                    if (remainder) this.push(remainder);
+                }
+                return;
+            }
+
+            if (this.inToolCallTag) {
+                this.toolCallBody += token.slice(pos);
+                const closeIdx = this.toolCallBody.indexOf(this.toolCallClose);
+                if (closeIdx !== -1) {
+                    const body = this.toolCallBody.slice(0, closeIdx);
+                    const remainder = this.toolCallBody.slice(closeIdx + this.toolCallClose.length);
+
+                    if (body.trim()) {
+                        try {
+                            const normalizedJson = body.trim()
+                                .replace(/'/g, '"')
+                                .replace(/True/g, 'true')
+                                .replace(/False/g, 'false')
+                                .replace(/None/g, 'null');
+                            const parsed = JSON.parse(normalizedJson);
+                            const calls = Array.isArray(parsed) ? parsed : [parsed];
+                            const validatedCalls: LLMToolCall[] = [];
+                            for (const call of calls) {
+                                if (call && typeof call === 'object' && call.name) {
+                                    validatedCalls.push({
+                                        id: call.id || `recovered_${Date.now()}_${Math.random().toString(36).slice(2)}`,
+                                        type: 'function',
+                                        function: {
+                                            name: call.name,
+                                            arguments: typeof call.arguments === 'string'
+                                                ? call.arguments
+                                                : JSON.stringify(call.arguments ?? {}),
+                                        }
+                                    });
+                                }
+                            }
+                            if (validatedCalls.length > 0) {
+                                this.callback({ type: 'tool_call', calls: validatedCalls });
+                            }
+                        } catch {
+                            // ignore
+                        }
+                    }
+
+                    this.inToolCallTag = false;
+                    this.toolCallBody = '';
+                    this.toolCallClose = '';
+                    if (remainder) this.push(remainder);
+                }
+                return;
+            }
+
+            if (this.inProgressTag) {
+                this.progressBody += token.slice(pos);
+                const closeIdx = this.progressBody.indexOf('</progress>');
+                if (closeIdx !== -1) {
+                    const body = this.progressBody.slice(0, closeIdx);
+                    const remainder = this.progressBody.slice(closeIdx + '</progress>'.length);
+                    if (body) {
+                        this.callback({ type: 'progress', content: body });
+                    }
+                    this.inProgressTag = false;
+                    this.progressBody = '';
+                    if (remainder) this.push(remainder);
+                }
+                return;
+            }
+
+            if (this.tagBuffer.length > 0) {
+                const ch = token[pos]!;
+                pos++;
+                this.tagBuffer += ch;
+                if (this.matchesStructuralOpenerPrefix(this.tagBuffer)) {
+                    if (this.tagBuffer === '<progress>') {
+                        this.inProgressTag = true;
+                        this.progressBody = '';
+                        this.tagBuffer = '';
+                    } else if (this.tagBuffer === '<tool_call|>') {
+                        this.inToolCallTag = true;
+                        this.toolCallBody = '';
+                        this.toolCallClose = '<|tool_response>';
+                        this.tagBuffer = '';
+                    } else if (this.tagBuffer === '<|tool_response>') {
+                        this.tagBuffer = '';
+                    } else if (this.tagBuffer === '<|channel>thought') {
+                        this.inGemmaThought = true;
+                        this.gemmaThoughtBody = '';
+                        this.gemmaThoughtClose = '<channel|>';
+                        this.tagBuffer = '';
+                    } else if (this.tagBuffer === '<|thought') {
+                        this.inGemmaThought = true;
+                        this.gemmaThoughtBody = '';
+                        this.gemmaThoughtClose = '|>';
+                        this.tagBuffer = '';
+                    }
+                } else {
+                    this.emitText(this.tagBuffer);
+                    this.tagBuffer = '';
+                }
+                continue;
+            }
+
+            const ltIdx = token.indexOf('<', pos);
+            if (ltIdx === -1) {
+                this.emitText(token.slice(pos));
+                return;
+            }
+
+            if (ltIdx > pos) {
+                this.emitText(token.slice(pos, ltIdx));
+            }
+            this.tagBuffer = '<';
+            pos = ltIdx + 1;
+        }
+    }
+
+    private emitText(text: string): void {
+        if (!text) return;
+        this.content += text;
+        this.callback({ type: 'text', content: text });
+    }
+
+    private emitReasoning(content: string): void {
+        if (!content) return;
+        this.reasoning += content;
+        this.callback({ type: 'thinking', content });
+    }
+
+    private matchesStructuralOpenerPrefix(candidate: string): boolean {
+        if ('<progress>'.startsWith(candidate)) return true;
+        if ('<tool_call|>'.startsWith(candidate)) return true;
+        if ('<|tool_response>'.startsWith(candidate)) return true;
+        return GEMMA_THOUGHT_OPENERS.some(opener => opener.startsWith(candidate));
+    }
+
+    /** Feed native reasoning tokens from the provider */
+    pushReasoning(content: string): void {
+        this.emitReasoning(content);
+    }
+
+    /** Feed structured tool calls from the provider API response */
+    pushToolCalls(calls: LLMToolCall[]): void {
+        this.callback({ type: 'tool_call', calls });
+    }
+
+    flush(): void {
+        if (this.tagBuffer) {
+            this.emitText(this.tagBuffer);
+            this.tagBuffer = '';
+        }
+        if (this.inGemmaThought) {
+            this.emitReasoning(normalizeGemmaThought(this.gemmaThoughtBody));
+            this.inGemmaThought = false;
+            this.gemmaThoughtBody = '';
+            this.gemmaThoughtClose = '';
+        }
+        if (this.inProgressTag) {
+            if (this.progressBody) {
+                this.emitText('<progress>' + this.progressBody);
+            }
+            this.inProgressTag = false;
+            this.progressBody = '';
+        }
+        if (this.inToolCallTag) {
+            this.inToolCallTag = false;
+            this.toolCallBody = '';
+            this.toolCallClose = '';
+        }
+    }
+
+    getCleanContent(): string {
+        return this.content;
+    }
+
+    getReasoning(): string | undefined {
+        return this.reasoning || undefined;
+    }
+}
+
+// ============================================================================
+// Interleaved Reasoning Decoder
+// ============================================================================
+
+/**
+ * Decoder for models that emit interleaved reasoning tags in text.
+ * Parses <think>...</think> and <progress>...</progress> tags from the
+ * raw token stream and emits typed events for each.
+ *
+ * Handles streaming where tags may be split across chunks.
+ */
+export class InterleavedReasoningDecoder implements StreamDecoder {
+    private buffer = '';
+    private content = '';
+    private reasoning = '';
+    private readonly callback: DecoderCallback;
+    private inThink = false;
+    private inProgress = false;
+
+    constructor(callback: DecoderCallback) {
+        this.callback = callback;
+    }
+
+    push(token: string): void {
+        this.buffer += token;
+        this.processBuffer();
+    }
+
+    flush(): void {
+        // Emit any remaining buffer content as text
+        if (this.buffer.length > 0) {
+            if (this.inThink) {
+                this.reasoning += this.buffer;
+                this.callback({ type: 'thinking', content: this.buffer });
+            } else if (this.inProgress) {
+                this.callback({ type: 'progress', content: this.buffer });
+            } else {
+                this.content += this.buffer;
+                this.callback({ type: 'text', content: this.buffer });
+            }
+            this.buffer = '';
+        }
+    }
+
+    getCleanContent(): string {
+        return this.content;
+    }
+
+    getReasoning(): string | undefined {
+        return this.reasoning || undefined;
+    }
+
+    private processBuffer(): void {
+        let safety = 0;
+        while (this.buffer.length > 0 && safety++ < 200) {
+            if (this.inThink) {
+                const closeIdx = this.buffer.indexOf('</think>');
+                if (closeIdx === -1) {
+                    // Might have partial closing tag at end
+                    if (this.buffer.endsWith('<') || this.buffer.endsWith('</') ||
+                        this.buffer.endsWith('</t') || this.buffer.endsWith('</th') ||
+                        this.buffer.endsWith('</thi') || this.buffer.endsWith('</thin') ||
+                        this.buffer.endsWith('</think')) {
+                        return; // Wait for more data
+                    }
+                    this.reasoning += this.buffer;
+                    this.callback({ type: 'thinking', content: this.buffer });
+                    this.buffer = '';
+                    return;
+                }
+                const thinkContent = this.buffer.slice(0, closeIdx);
+                if (thinkContent) {
+                    this.reasoning += thinkContent;
+                    this.callback({ type: 'thinking', content: thinkContent });
+                }
+                this.buffer = this.buffer.slice(closeIdx + 8); // '</think>'.length
+                this.inThink = false;
+                continue;
+            }
+
+            if (this.inProgress) {
+                const closeIdx = this.buffer.indexOf('</progress>');
+                if (closeIdx === -1) {
+                    if (this.couldBePartialTag(this.buffer, '</progress>')) return;
+                    this.callback({ type: 'progress', content: this.buffer });
+                    this.buffer = '';
+                    return;
+                }
+                const progressContent = this.buffer.slice(0, closeIdx);
+                if (progressContent) {
+                    this.callback({ type: 'progress', content: progressContent });
+                }
+                this.buffer = this.buffer.slice(closeIdx + 11); // '</progress>'.length
+                this.inProgress = false;
+                continue;
+            }
+
+            // Look for opening tags
+            const thinkIdx = this.buffer.indexOf('<think>');
+            const progressIdx = this.buffer.indexOf('<progress>');
+
+            // Find earliest tag
+            const nextTag = this.findEarliest(thinkIdx, progressIdx);
+
+            if (nextTag === -1) {
+                // No complete opening tags — check for partial tag at end
+                const lastAngle = this.buffer.lastIndexOf('<');
+                if (lastAngle >= 0 && lastAngle > this.buffer.length - 12) {
+                    // Potential partial tag — emit text before it, keep the rest
+                    const textBefore = this.buffer.slice(0, lastAngle);
+                    if (textBefore) {
+                        this.content += textBefore;
+                        this.callback({ type: 'text', content: textBefore });
+                    }
+                    this.buffer = this.buffer.slice(lastAngle);
+                    return;
+                }
+                // No partial tags — emit all as text
+                this.content += this.buffer;
+                this.callback({ type: 'text', content: this.buffer });
+                this.buffer = '';
+                return;
+            }
+
+            // Emit text before the tag
+            const textBefore = this.buffer.slice(0, nextTag);
+            if (textBefore) {
+                this.content += textBefore;
+                this.callback({ type: 'text', content: textBefore });
+            }
+
+            if (nextTag === thinkIdx) {
+                this.buffer = this.buffer.slice(nextTag + 7); // '<think>'.length
+                this.inThink = true;
+            } else {
+                this.buffer = this.buffer.slice(nextTag + 10); // '<progress>'.length
+                this.inProgress = true;
+            }
+        }
+    }
+
+    private findEarliest(a: number, b: number): number {
+        if (a === -1) return b;
+        if (b === -1) return a;
+        return Math.min(a, b);
+    }
+
+    private couldBePartialTag(buffer: string, tag: string): boolean {
+        for (let i = 1; i < tag.length; i++) {
+            if (buffer.endsWith(tag.slice(0, i))) return true;
+        }
+        return false;
+    }
+}
+
+// ============================================================================
+// Pluggable Decoder Registry
+// ============================================================================
+
+export interface DecoderOptions {
+    /** Known tool names for text-based tool call recovery */
+    knownToolNames?: Set<string>;
+}
+
+/**
+ * Factory function that creates a StreamDecoder instance.
+ * External code registers these via `registerDecoder()`.
+ */
+export type DecoderFactory = (callback: DecoderCallback, options?: DecoderOptions) => StreamDecoder;
+
+/** Internal registry of decoder factories, keyed by decoder type name */
+const decoderRegistry = new Map<string, DecoderFactory>();
+
+/**
+ * Register a custom stream decoder type.
+ * Once registered, it can be used via `createDecoder(name, ...)` or
+ * by passing `decoderType: name` in ChatOptions.
+ *
+ * @example
+ * ```typescript
+ * import { registerDecoder } from 'universal-llm-client';
+ *
+ * registerDecoder('my-decoder', (callback, options) => {
+ *   return new MyCustomDecoder(callback, options);
+ * });
+ * ```
+ */
+export function registerDecoder(type: string, factory: DecoderFactory): void {
+    decoderRegistry.set(type, factory);
+}
+
+/**
+ * Get all registered decoder type names.
+ */
+export function getRegisteredDecoders(): string[] {
+    return Array.from(decoderRegistry.keys());
+}
+
+// Pre-register built-in decoders
+registerDecoder('passthrough', (cb) => new PassthroughDecoder(cb));
+registerDecoder('standard-chat', (cb) => new StandardChatDecoder(cb));
+registerDecoder('interleaved-reasoning', (cb) => new InterleavedReasoningDecoder(cb));
+
+/**
+ * Create a stream decoder by type name.
+ * Looks up the decoder in the registry (built-in + custom).
+ *
+ * @throws Error if the decoder type is not registered
+ */
+export function createDecoder(
+    type: DecoderType | string,
+    callback: DecoderCallback,
+    options?: DecoderOptions,
+): StreamDecoder {
+    const factory = decoderRegistry.get(type);
+    if (!factory) {
+        const available = Array.from(decoderRegistry.keys()).join(', ');
+        throw new Error(`Unknown decoder type: "${type}". Available: ${available}`);
+    }
+    return factory(callback, options);
+}
+