Support GPT-5.6 provider compatibility and version 24.0.0 release

CHANGELOG.md

@@ -4,6 +4,22 @@ All notable changes to WebBrain are documented in this file.
 
 This changelog was generated from the repository Git history and release tags. Versions without a Git tag are inferred from version-bump commits and the current `package.json` / browser manifest versions.
 
+## [24.0.0] - 2026-07-16
+
+### Added
+- First-class **GPT-5.6** support for official OpenAI: default model `gpt-5.6-terra`, UI suggestions for Terra / Sol / Luna, and a **1,050,000**-token context window for the `gpt-5.6*` family.
+- Official OpenAI GPT-5.6 routes through the **Responses API** (`/v1/responses`) with encrypted reasoning replay (`response_items`), function-call conversion, streaming, and usage normalization. Older models and custom/proxy base URLs stay on Chat Completions.
+- **Advanced model compatibility** settings on every provider card: preset (Auto / OpenAI / Qwen / DeepSeek / OpenRouter / Custom), reasoning effort, system vs developer role, max-token field, and safe custom request-body JSON.
+- Shared `provider-compatibility.js` layer (Chrome + Firefox) that maps roles, token fields, reasoning knobs, and merges protected extra body fields for OpenAI-compatible, Azure OpenAI, and llama.cpp request builders.
+- Documented the release in `WHATS_NEW_IN_V24.md` at the repository root.
+
+### Changed
+- OpenAI-compatible request building is centralized so compatibility presets apply consistently without mutating stored chat history.
+- OpenAI default cost metadata migrates safely when upgrading from earlier shipped defaults.
+
+### Tests
+- Added / retained coverage for Responses routing, reasoning replay, streaming, provider-compatibility merges, GPT-5.6 context windows, and default-model migration.
+
 ## [23.3.6] - 2026-07-15
 
 ### Added

WHATS_NEW_IN_V24.md

@@ -0,0 +1,184 @@
+# What’s New in WebBrain v24
+
+v24 is a **provider / model-compatibility major release**. It does not redesign the agent loop, tools, or browser automation surface. It changes how WebBrain talks to LLMs—especially official OpenAI GPT-5.6 and other OpenAI-compatible endpoints with different wire formats.
+
+---
+
+## TL;DR
+
+| Area | What changed |
+|------|----------------|
+| **Official OpenAI GPT-5.6** | First-class model family (default **Terra**), 1.05M context, **Responses API** on `api.openai.com` |
+| **Older / proxy endpoints** | Still use **Chat Completions**; no forced `/v1/responses` on proxies or non-OpenAI hosts |
+| **Configurable compatibility** | Per-provider settings for reasoning, system role, max-token field, and safe extra JSON body |
+| **Who is affected at runtime** | OpenAI-compatible providers, Azure OpenAI, llama.cpp |
+| **Who is not** | Native Anthropic and AWS Bedrock request builders (UI panel may show; runtime path does not apply compat) |
+
+---
+
+## Did main already support GPT-5.6?
+
+**Mostly yes, as a generic GPT-5 Chat Completions model—not as a special-cased product.**
+
+On earlier 23.x lines:
+
+- Any model id matching the “new OpenAI contract” (including names starting with `gpt-5`) already used `max_completion_tokens` instead of `max_tokens`.
+- You could type a GPT-5.6 model id and call Chat Completions if the API accepted it.
+- There was **no** first-class 5.6 default/suggestions list, **no** dedicated 5.6 context-window rule, and (until later 23.3.x work on `main`) **no** official Responses API path.
+
+v24 (and the 23.3.x Responses work merged into this release) makes GPT-5.6 **first-class**:
+
+- Default OpenAI model: `gpt-5.6-terra`
+- Suggestions: Terra, Sol, Luna, base `gpt-5.6`, plus older GPT-5.x ids
+- Context window for `gpt-5.6*`: **1,050,000** tokens (same class as GPT-5.5-pro)
+- Official OpenAI only: route GPT-5.6 family traffic to **`POST /v1/responses`**
+
+### Responses API gate (intentionally narrow)
+
+Responses is used only when **all** of these hold:
+
+1. Provider is OpenAI (`providerName === 'openai'`)
+2. Base URL is official `https://api.openai.com/v1`
+3. Model matches the GPT-5.6 family: `gpt-5.6`, `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna`, including optional dated suffixes (e.g. `gpt-5.6-terra-2026-07-15`)
+
+**Not** routed to Responses:
+
+- GPT-5.5 and older
+- OpenRouter / proxies / custom base URLs (even if the model string contains `gpt-5.6`)
+- Local OpenAI-compatible servers (Ollama, LM Studio, vLLM, etc.)
+
+### What the Responses path does
+
+- Builds Responses-shaped bodies: `input`, `max_output_tokens`, Responses tool format, `store: false`
+- Includes `reasoning.encrypted_content` and defaults reasoning effort to **medium** (with compatibility / extraBody overrides)
+- Maps chat history for tools: assistant `tool_calls` → `function_call`, tool results → `function_call_output`
+- Replays prior **`response_items`** so encrypted reasoning state survives multi-step tool loops
+- Streams via Responses SSE events and normalizes usage back to `prompt_tokens` / `completion_tokens`
+- Returns `responseItems` for the agent to attach on the next turn
+
+Chrome uses `fetchWithFallback`; Firefox uses `fetchWithTimeout`—same provider logic, platform-appropriate fetch helpers.
+
+---
+
+## Configurable compatibility (“advanced model compatibility”)
+
+**Not a new agent mode.** Per-provider knobs that change **how the HTTP request is shaped** so different servers accept the same agent conversation.
+
+### Why it exists
+
+“OpenAI-compatible” is not one wire format. Different stacks want different fields for the same idea:
+
+| Idea | Examples of wire differences |
+|------|------------------------------|
+| Token limit | `max_tokens` vs `max_completion_tokens` vs Responses `max_output_tokens` |
+| System prompt role | `system` vs `developer` |
+| Reasoning / thinking | OpenAI `reasoning` / `reasoning_effort`, OpenRouter `reasoning: { effort }`, Qwen `chat_template_kwargs.enable_thinking`, DeepSeek `chat_template_kwargs.thinking` |
+| Extra server flags | Template kwargs, repetition penalty, etc. |
+
+Without config, WebBrain either hard-codes one convention or fails on some endpoints. v24 makes those choices **settings**.
+
+### Settings UI
+
+On each provider card → **Advanced model compatibility**:
+
+| Control | Effect on the request |
+|---------|------------------------|
+| **Preset** (`auto` / OpenAI / Qwen / DeepSeek / OpenRouter / Custom) | Which family of field names to use for reasoning, etc. `auto` infers from provider/model name. |
+| **Reasoning effort** (`auto` / off / minimal … max) | Turns thinking on/off or sets effort; mapped per preset. |
+| **System prompt role** (`auto` / system / developer) | Optionally rewrites `system` → `developer` when building the request (**stored history is not mutated**). |
+| **Max tokens field** (`auto` / `max_tokens` / `max_completion_tokens`) | Which token-limit key is sent. |
+| **Custom request body JSON** | Extra object merged after the preset (e.g. `{"chat_template_kwargs":{"enable_thinking":true}}`). |
+
+**Protected / reserved keys** cannot be overridden via custom JSON: `model`, `messages`, `input`, `instructions`, `tools`, `tool_choice`, `stream`, `max_tokens`, `max_completion_tokens`, `max_output_tokens`. Unsafe prototype keys (`__proto__`, `prototype`, `constructor`) are blocked.
+
+There is a **Reset to automatic** control and live validation for the JSON field.
+
+### Auto presets
+
+| Auto detection | When |
+|----------------|------|
+| **openrouter** | `providerName === 'openrouter'` |
+| **deepseek** | provider is deepseek *or* model name contains `deepseek` |
+| **qwen** | model name contains `qwen` |
+| **openai** | official `api.openai.com` |
+| **standard** | everyone else — no automatic reasoning fields unless you set knobs / custom JSON |
+
+So for Grok, Groq, Mistral, Ollama, etc., **auto is mostly a no-op** until you set reasoning / role / token field / custom body yourself (or force a preset like Qwen when serving Qwen).
+
+### Example
+
+Same chat (system rules + user “hello” + tools):
+
+- **Legacy / default:** `{ messages, max_tokens, tools, … }`
+- **Qwen-style + high reasoning:** messages (maybe `developer` role), `max_completion_tokens`, plus `chat_template_kwargs: { enable_thinking: true, … }`
+- **Official GPT-5.6:** Responses API path; still honors reasoning/role compatibility where applicable
+
+---
+
+## Which providers get this?
+
+### UI
+
+The compatibility panel is rendered on **every provider card** in Settings.
+
+### Runtime (actually applied)
+
+Only implementations that call the shared helpers (`_mapMessages`, `_addConfiguredMaxTokens`, `_mergeConfiguredRequestBody`):
+
+| Implementation type | Providers |
+|---------------------|-----------|
+| **`openai` (OpenAI-compatible)** | Almost the entire cloud + local list (see below) |
+| **`azure_openai`** | Azure OpenAI |
+| **`llamacpp`** | Local llama.cpp server |
+
+**Not wired into the native request builders:** Anthropic Claude and AWS Bedrock. Changing the panel for those cards does not reshape their native API calls on this release.
+
+### OpenAI-compatible provider ids (examples)
+
+**Local / self-hosted:** WebBrain Cloud, Ollama, LM Studio, Jan, vLLM, SGLang, LocalAI, llama.cpp  
+
+**Cloud (OpenAI-compatible HTTP):** OpenAI, Gemini (compat endpoint), Cloudflare, Mistral, DeepSeek, xAI, NVIDIA, Groq, MiniMax, Alibaba/Qwen, Together, OpenRouter, Hugging Face, Fireworks, Azure OpenAI  
+
+Custom OpenAI base-URL providers you add also get the same path.
+
+---
+
+## Implementation map (for contributors)
+
+| Piece | Role |
+|-------|------|
+| `src/*/src/providers/provider-compatibility.js` | Normalize compat config, detect Responses eligibility, map messages, max-token field, reasoning body fragments, safe extraBody merge/validation |
+| `src/*/src/providers/base.js` | `_mapMessages`, `_addConfiguredMaxTokens`, `_mergeConfiguredRequestBody` |
+| `src/*/src/providers/openai.js` | Chat Completions + Responses routing, streaming, `response_items` replay |
+| `src/*/src/providers/azure-openai.js`, `llamacpp.js` | Use shared body merge helpers |
+| `src/*/src/providers/manager.js` | OpenAI default model constant / migration |
+| `src/*/src/providers/context-windows.js` | 1.05M for `gpt-5.6*` |
+| `src/*/src/ui/settings.js` + `settings.html` | Advanced compatibility UI |
+| `test/run.js` | Responses + compatibility regression coverage |
+
+Chrome and Firefox stay mirrored except for platform fetch helpers.
+
+---
+
+## What v24 is *not*
+
+- Not a new set of browser tools or a change to Ask vs Act safety defaults  
+- Not “enable GPT-5.6 by allowlist only”—you could already type model ids; v24 is about **correct wire format + usability**  
+- Not full native Anthropic/Bedrock compatibility knobs  
+- Not a guarantee that every third-party “OpenAI-compatible” server implements every optional field—you still configure per endpoint
+
+---
+
+## Upgrade notes
+
+- **New installs:** OpenAI defaults to `gpt-5.6-terra`.
+- **Existing installs:** Manager migration promotes untouched prior OpenAI defaults (e.g. older default model / cost metadata) to the new shipped default when safe; custom model choices you saved are preserved.
+- **Proxies:** If you point `baseUrl` at a proxy that only speaks Chat Completions, keep using GPT-5.6 model ids—WebBrain will **not** call `/v1/responses` for non-official hosts.
+- **Reasoning:** On official Responses, default effort is medium; use Advanced compatibility or extra body to tune. Prefer dedicated settings over stuffing reserved fields into custom JSON.
+
+---
+
+## Version
+
+- Extension / package version: **24.0.0**
+- See [CHANGELOG.md](./CHANGELOG.md) for the formal release entry.

package.json

@@ -1,6 +1,6 @@
 {
   "name": "webbrain",
-  "version": "23.3.10",
+  "version": "24.0.0",
   "description": "Open-source AI browser agent — chat with pages, automate tasks, multi-provider LLM support.",
   "private": true,
   "type": "module",

src/chrome/ARCHITECTURE.md

@@ -1,6 +1,6 @@
 # WebBrain Chrome/Edge Extension — Architecture
 
-> Version 23.3.10 · Manifest V3 · Service Worker background
+> Version 24.0.0 · Manifest V3 · Service Worker background
 
 ## High-Level Overview
 

src/chrome/manifest.json

@@ -1,7 +1,7 @@
 {
   "manifest_version": 3,
   "name": "WebBrain",
-  "version": "23.3.10",
+  "version": "24.0.0",
   "description": "Open-source AI browser agent — chat with pages, automate tasks, multi-provider LLM support.",
   "permissions": [
     "sidePanel",

src/chrome/src/providers/azure-openai.js

@@ -63,7 +63,7 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
 
   _addMaxTokens(body, options) {
     // Azure OpenAI follows the legacy OpenAI contract here.
-    body.max_tokens = options.maxTokens ?? 4096;
+    this._addConfiguredMaxTokens(body, options, 'max_tokens');
   }
 
   _addTemperature(body, options) {
@@ -89,17 +89,21 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
     body.stream_options = { ...streamOptions, include_usage: true };
   }
 
-  async chat(messages, options = {}) {
-    const body = { messages, stream: false };
+  _buildRequestBody(messages, options = {}, stream = false) {
+    let body = { messages: this._mapMessages(messages), stream };
     this._addTemperature(body, options);
     this._addMaxTokens(body, options);
     if (this._shouldSendTools(messages, options)) {
       body.tools = options.tools;
       body.tool_choice = options.toolChoice || 'auto';
     }
-    if (options.extraBody && typeof options.extraBody === 'object') {
-      Object.assign(body, options.extraBody);
-    }
+    body = this._mergeConfiguredRequestBody(body, options);
+    if (stream) this._addStreamUsageOptions(body);
+    return body;
+  }
+
+  async chat(messages, options = {}) {
+    const body = this._buildRequestBody(messages, options, false);
 
     const url = this._chatUrl();
     let res;
@@ -129,17 +133,7 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
   }
 
   async *chatStream(messages, options = {}) {
-    const body = { messages, stream: true };
-    this._addTemperature(body, options);
-    this._addMaxTokens(body, options);
-    if (this._shouldSendTools(messages, options)) {
-      body.tools = options.tools;
-      body.tool_choice = options.toolChoice || 'auto';
-    }
-    if (options.extraBody && typeof options.extraBody === 'object') {
-      Object.assign(body, options.extraBody);
-    }
-    this._addStreamUsageOptions(body);
+    const body = this._buildRequestBody(messages, options, true);
 
     const url = this._chatUrl();
     let res;
@@ -185,4 +179,3 @@ export class AzureOpenAIProvider extends BaseLLMProvider {
     yield { type: 'done', content: '' };
   }
 }
-

src/chrome/src/providers/base.js

@@ -1,4 +1,9 @@
 import { inferContextWindow } from './context-windows.js';
+import {
+  addConfiguredMaxTokens,
+  mapProviderMessages,
+  mergeProviderRequestBody,
+} from './provider-compatibility.js';
 
 /**
  * Base LLM Provider — all providers implement this interface.
@@ -83,6 +88,18 @@ export class BaseLLMProvider {
     return !!this.config.useCompactPrompt;
   }
 
+  _mapMessages(messages) {
+    return mapProviderMessages(messages, this.config);
+  }
+
+  _addConfiguredMaxTokens(body, options, fallback = 'max_tokens') {
+    return addConfiguredMaxTokens(body, options.maxTokens ?? 4096, this.config, fallback);
+  }
+
+  _mergeConfiguredRequestBody(body, options = {}) {
+    return mergeProviderRequestBody(body, this.config, options.extraBody);
+  }
+
   /**
    * Prompt tier for this provider: 'compact' | 'mid' | 'full'. Drives both
    * which ACT system prompt and which tool set the agent uses.

src/chrome/src/providers/llamacpp.js

@@ -38,22 +38,23 @@ export class LlamaCppProvider extends BaseLLMProvider {
     return !!this.config.useCompactPrompt;
   }
 
-  async chat(messages, options = {}) {
+  _buildRequestBody(messages, options = {}, stream = false) {
     const body = {
-      messages,
+      messages: this._mapMessages(messages),
       temperature: options.temperature ?? 0.7,
-      max_tokens: options.maxTokens ?? 4096,
-      stream: false,
+      stream,
     };
-
-    if (this.model) {
-      body.model = this.model;
-    }
-
+    this._addConfiguredMaxTokens(body, options, 'max_tokens');
+    if (this.model) body.model = this.model;
     if (options.tools && options.tools.length > 0) {
       body.tools = options.tools;
       body.tool_choice = options.toolChoice || 'auto';
     }
+    return this._mergeConfiguredRequestBody(body, options);
+  }
+
+  async chat(messages, options = {}) {
+    const body = this._buildRequestBody(messages, options, false);
 
     const url = `${this.baseUrl}/v1/chat/completions`;
     let res;
@@ -90,21 +91,7 @@ export class LlamaCppProvider extends BaseLLMProvider {
   }
 
   async *chatStream(messages, options = {}) {
-    const body = {
-      messages,
-      temperature: options.temperature ?? 0.7,
-      max_tokens: options.maxTokens ?? 4096,
-      stream: true,
-    };
-
-    if (this.model) {
-      body.model = this.model;
-    }
-
-    if (options.tools && options.tools.length > 0) {
-      body.tools = options.tools;
-      body.tool_choice = options.toolChoice || 'auto';
-    }
+    const body = this._buildRequestBody(messages, options, true);
 
     const streamUrl = `${this.baseUrl}/v1/chat/completions`;
     let res;