converse-mcp-server 2.29.2 → 3.0.1

package/docs/PROVIDERS.md

@@ -22,18 +22,22 @@ This guide documents all supported AI providers in the Converse MCP Server and t
 - **Get Key**: [makersuite.google.com/app/apikey](https://makersuite.google.com/app/apikey)
 - **Environment Variable**: `GOOGLE_API_KEY`
 - **Supported Models**:
-  - `gemini-3-pro-preview` (alias: `pro`) - Enhanced reasoning with thinking levels (1M context, 64K output)
+  - `gemini-3.1-pro-preview` (aliases: `pro`, `gemini-pro`) - Most advanced reasoning with expanded thinking levels (1M context, 64K output)
+  - `gemini-3.5-flash` (aliases: `gemini-3.5`, `flash-3.5`) - Frontier-level agentic and coding performance at Flash speed (1M context, 65K output)
   - `gemini-2.5-pro` (alias: `pro 2.5`) - Deep reasoning with thinking budget (1M context, 65K output)
   - `gemini-2.5-flash` (alias: `flash`) - Ultra-fast model with thinking budget (1M context, 65K output)
-- **Note**: The short model name `gemini` (and `gemini:flash` / `gemini:pro`) routes to the **Antigravity CLI** (`agy`, OAuth-based access). For Google API access, use specific model names like `gemini-2.5-pro` or `gemini-2.5-flash` (bare `gemini-pro`/`gemini-flash` also route to the Google API).
+  - `gemini-2.5-flash-lite` (alias: `flash-lite`) - Lightweight fast model (1M context, 65K output)
+- **Note**: The short model name `gemini` (and `gemini:flash` / `gemini:pro`) routes to the **Antigravity CLI** (`agy`, OAuth-based access). For Google API access, use specific model names like `gemini-3.1-pro-preview` or `gemini-2.5-flash` (bare `gemini-pro`/`gemini-flash` also route to the Google API).
 
 ### X.AI (Grok)
 - **API Key Format**: `xai-...` (starts with `xai-`)
 - **Get Key**: [console.x.ai](https://console.x.ai/)
 - **Environment Variable**: `XAI_API_KEY`
 - **Supported Models**:
-  - `grok-4-0709` - Latest with image support and web search
-  - `grok-code-fast-1` - Speedy and economical reasoning model that excels at agentic coding
+  - `grok-4.5` (default; aliases: `grok`, `grok-4.5-latest`, `grok-build-latest`) - Flagship model with image input, reasoning content, and native web/X search (500K context)
+- **Reasoning**: `reasoning_effort` maps to Grok's `low`/`medium`/`high`. Grok 4.5 always reasons and cannot be turned off, so `none`/`minimal`/`low` clamp to `low`, `medium` stays `medium`, and `high`/`max` clamp to `high`.
+- **Web search**: Automatic — native web/X search (Agent Tools) is attached on every Grok 4.5 request; the model decides per-request whether to search, and any citations are returned in metadata.
+- **Retired IDs**: Older Grok identifiers (e.g. `grok-4-0709`, `grok-code-fast-1`) still pass through as explicit model strings, but xAI does not surface a retirement error for them — it silently remaps them upstream to a current model (HTTP 200). Use `grok-4.5` for predictable results.
 
 ### Anthropic (Claude)
 - **API Key Format**: `sk-ant-...` (starts with `sk-ant-`)
@@ -53,39 +57,40 @@ This guide documents all supported AI providers in the Converse MCP Server and t
 - **Get Key**: [console.mistral.ai](https://console.mistral.ai/)
 - **Environment Variable**: `MISTRAL_API_KEY`
 - **Supported Models**:
-  - `mistral-large-latest` - Most capable model
-  - `mistral-medium-latest` - Balanced performance
-  - `mistral-small-latest` - Fast and efficient
-  - `open-mistral-7b`, `open-mixtral-8x7b`, `open-mixtral-8x22b` - Open-source models
+  - `mistral-medium-3-5` (default; aliases: `mistral`, `mistral-medium`, `mistral-medium-latest`) - Frontier-class multimodal model with adjustable reasoning (256K context)
+  - `mistral-small-2603` (aliases: `mistral-small`, `mistral-small-latest`) - Hybrid multimodal model unifying instruct, reasoning, and coding (256K context)
+  - `mistral-large-2512` (aliases: `mistral-large`, `mistral-large-latest`) - Open-weight MoE flagship, image-capable (256K context)
+- **Reasoning**: Mistral exposes only `high` and `none`. On the reasoning-capable models (`mistral-medium-3-5`, `mistral-small-2603`), every enabled `reasoning_effort` level maps to `high` and only `none` disables thinking. `mistral-large-2512` has no adjustable reasoning, so `reasoning_effort` is never sent for it.
+- **Images**: `mistral-medium-3-5` and `mistral-small-2603` accept image input; `mistral-large-2512` is text-only.
 
 ### DeepSeek
 - **API Key Format**: 32+ character string
 - **Get Key**: [platform.deepseek.com](https://platform.deepseek.com/)
 - **Environment Variable**: `DEEPSEEK_API_KEY`
 - **Supported Models**:
-  - `deepseek-chat` - Advanced conversational model
-  - `deepseek-coder` - Specialized for code generation
+  - `deepseek-v4-pro` (default; aliases: `deepseek`, `deepseek-pro`) - Flagship MoE model with thinking mode (1M context, 384K max output, text-only)
+  - `deepseek-v4-flash` (alias: `deepseek-flash`) - Faster, lower-cost V4 tier with thinking mode (1M context, 384K max output, text-only)
+- **Reasoning**: DeepSeek V4 exposes thinking mode via a `thinking` toggle plus a `reasoning_effort` of `high` or `max`. `none` disables thinking; `minimal`/`low`/`medium`/`high` enable thinking at `high`; `max` enables thinking at `max`.
 
 ### OpenRouter
 - **API Key Format**: `sk-or-...` (starts with `sk-or-`)
 - **Get Key**: [openrouter.ai/keys](https://openrouter.ai/keys)
-- **Environment Variables**: 
+- **Environment Variables**:
   - `OPENROUTER_API_KEY` - Your API key
-  - `OPENROUTER_REFERER` - Required referer URL (e.g., your GitHub repo)
-  - `OPENROUTER_TITLE` - Optional title for request tracking
-  - `OPENROUTER_DYNAMIC_MODELS` - Enable dynamic model discovery (default: false, required for `provider/model` format)
-- **Static Models**: Pre-configured models available without dynamic discovery
-  - `qwen/qwen3-235b-a22b-thinking-2507` - Qwen3 235B with thinking capabilities
-  - `qwen/qwen3-coder` - Qwen3 specialized for coding
-  - `moonshotai/kimi-k2` - Kimi K2 with 200K context window
-  - `openrouter/auto` - Auto-selects best model using NotDiamond routing
-- **Dynamic Models**: Requires `OPENROUTER_DYNAMIC_MODELS=true` to use any model in `provider/model` format
-  - `anthropic/claude-3.5-sonnet`
-  - `openai/gpt-4-turbo`
-  - `google/gemini-pro`
-  - `mistralai/mistral-large`
-  - `meta-llama/llama-3.1-405b-instruct`
-  - And many more - see [openrouter.ai/models](https://openrouter.ai/models)
+  - `OPENROUTER_REFERER` - Optional referer URL (e.g., your GitHub repo) for OpenRouter ranking credit; omitting it is valid
+  - `OPENROUTER_TITLE` - Optional title for OpenRouter ranking credit
+- **Curated Models** (default `z-ai/glm-5.2`):
+  - `z-ai/glm-5.2` (aliases: `glm`, `glm-5.2`) - Large-scale reasoning model, text-only (1M context)
+  - `deepseek/deepseek-v4-pro` - DeepSeek V4 Pro reasoning model, text-only (1M context, 384K output)
+  - `deepseek/deepseek-v4-flash` - Faster, lower-cost DeepSeek V4 tier, text-only (1M context, 384K output)
+  - `qwen/qwen3.7-max` (alias: `qwen3.7-max`) - Flagship Qwen, text-only (1M context)
+  - `qwen/qwen3.7-plus` (alias: `qwen3.7-plus`) - Image-capable Qwen (1M context)
+  - `moonshotai/kimi-k2.7-code` (alias: `kimi-k2.7-code`) - Coding model, image-capable, reasoning always on (256K context)
+  - `moonshotai/kimi-k2.6` (alias: `kimi-k2.6`) - Image-capable general model (256K context)
+  - `openrouter/auto` - Auto-selects the best model for your prompt
+- **Any other model**: Use the full `provider/model` slug directly (e.g. `anthropic/claude-sonnet-5`, `meta-llama/llama-3.1-405b-instruct`) or the `openrouter:` namespace (e.g. `openrouter:z-ai/glm-5.2`). No extra configuration is needed — slugs route to OpenRouter as-is. A slug absent from OpenRouter's live catalog fails before inference with `MODEL_NOT_FOUND`. See [openrouter.ai/models](https://openrouter.ai/models) for the full catalog.
+- **Reasoning**: Per-model. `z-ai/glm-5.2` and the `deepseek/deepseek-v4-*` slugs are effort-tiered (`max` → `xhigh`, other enabled levels → `high`, `none` disables); `qwen/qwen3.7-*` and `moonshotai/kimi-k2.6` are enable/disable only (`none` disables, any other level enables); `moonshotai/kimi-k2.7-code` always reasons and cannot be disabled; `openrouter/auto` lets the router choose.
+- **Web search (opt-in, adds cost)**: OpenRouter web search is off by default because it incurs a real per-request charge. Enable it explicitly by appending `:online` to a slug (e.g. `z-ai/glm-5.2:online` or `openrouter:qwen/qwen3.7-max:online`). When enabled, `annotations[].url_citation` citations are captured into metadata. Ordinary requests never attach a web-search plugin.
 
 ### Codex
 - **API Key Format**: Optional (uses ChatGPT login by default)
@@ -133,8 +138,6 @@ This guide documents all supported AI providers in the Converse MCP Server and t
 - **Subscription Access**: Leverages the Antigravity weekly compute allowance instead of pay-per-API-call
 - **One-shot responses**: The provider shells out to `agy -p` under a pseudo-terminal and returns the full response in a single chunk (no token-level streaming; ~7s minimum per call, ~30-60s for very large prompts)
 
-> Note: This replaces the previous `@google/gemini-cli` (`ai-sdk-provider-gemini-cli`) integration, whose OAuth access Google sunsets on 2026-06-18.
-
 **Authentication Setup:**
 ```bash
 # Install the Antigravity CLI (agy)
@@ -155,18 +158,19 @@ agy
   "name": "chat",
   "arguments": {
     "prompt": "Explain async/await in JavaScript",
-    "model": "gemini"
+    "models": ["gemini"]
   }
 }
 ```
 
-*Consensus Tool:*
+*Consensus mode:*
 ```json
 {
-  "name": "consensus",
+  "name": "chat",
   "arguments": {
     "prompt": "Should we use TypeScript for this component?",
-    "models": ["gemini", "gpt-5", "claude-sonnet-4-6"]
+    "models": ["gemini", "gpt-5.6", "claude"],
+    "mode": "consensus"
   }
 }
 ```
@@ -197,13 +201,30 @@ agy
 - **Subscription Access**: Uses your Claude subscription instead of API credits
 - **Local File Access**: Reads files directly from the working directory
 - **Image Support**: Via the SDK's streaming input mode
-- **SDK-Managed Parameters**: `temperature`, `use_websearch`, and `reasoning_effort` are managed internally (ignored if specified)
+- **SDK-Managed Parameters**: `reasoning_effort` and other sampling parameters are managed internally by the SDK
 
 **Differences from Anthropic API Provider:**
 - **Authentication**: Claude Code login vs `ANTHROPIC_API_KEY`
 - **Billing**: Claude subscription vs pay-per-use API
 - **Model Routing**: `claude` and `claude:*` → SDK provider; specific names (e.g., `claude-fable-5`, `opus`, `sonnet`) → API provider
 
+### GitHub Copilot SDK
+- **Authentication**: GitHub Copilot subscription via the Copilot CLI (`gh auth login` with an active Copilot subscription) — no API key needed
+- **Setup Required**: Authenticate the GitHub CLI and ensure your account has an active Copilot subscription
+- **Environment Variables**: None
+- **Supported Models** (reach them with the `copilot:` namespace, e.g. `copilot:gpt-5.6-terra`):
+  - `copilot` - Uses Copilot's default or env-configured model
+  - OpenAI: `gpt-5.6-sol` (aliases: bare `gpt-5.6`, `gpt-5`), `gpt-5.6-terra` (recommended balanced tier), `gpt-5.6-luna`
+  - Anthropic: `claude-fable-5` (alias: `fable`), `claude-sonnet-5` (alias: `sonnet`), `claude-opus-4.8` (aliases: `opus`, `claude`)
+  - Google: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.5-flash` (alias: `gemini-flash`)
+- **Reasoning**: The `gpt-5.6-sol`/`terra`/`luna` tiers accept `reasoning_effort`.
+- **Explicit pass-through**: Any other `copilot:<id>` model string is forwarded to the Copilot backend verbatim, so IDs outside the curated list still work while the backend accepts them.
+
+**Key Features:**
+- **Subscription Access**: Uses your GitHub Copilot subscription instead of API credits
+- **Multiple Backends**: OpenAI, Anthropic, and Google models through a single subscription
+- **Model IDs**: Lowercase dot-versioned (e.g. `claude-opus-4.8`, `gemini-3.1-pro-preview`)
+
 ## Configuration Examples
 
 ### Basic Configuration (.env file)
@@ -214,12 +235,10 @@ ANTHROPIC_API_KEY=sk-ant-your_key_here
 MISTRAL_API_KEY=your_mistral_key_here
 DEEPSEEK_API_KEY=your_deepseek_key_here
 
-# OpenRouter requires both API key and referer
+# OpenRouter needs only an API key; any provider/model slug works directly
 OPENROUTER_API_KEY=sk-or-your_key_here
+# Optional: referer and title for OpenRouter ranking credit
 OPENROUTER_REFERER=https://github.com/YourUsername/YourApp
-# Optional: Enable dynamic model discovery to use any OpenRouter model
-OPENROUTER_DYNAMIC_MODELS=true
-# Optional: Add title for request tracking
 OPENROUTER_TITLE=Converse
 
 # Codex - Optional API key (uses ChatGPT login by default)
@@ -243,7 +262,6 @@ CODEX_APPROVAL_POLICY=never                  # never (default), untrusted, on-fa
         "DEEPSEEK_API_KEY": "your_key_here",
         "OPENROUTER_API_KEY": "your_key_here",
         "OPENROUTER_REFERER": "https://github.com/YourUsername/YourApp",
-        "OPENROUTER_DYNAMIC_MODELS": "true",
         "OPENROUTER_TITLE": "Converse"
       }
     }
@@ -257,20 +275,26 @@ CODEX_APPROVAL_POLICY=never                  # never (default), untrusted, on-fa
 All providers support streaming responses for real-time output.
 
 ### Image Support
-- **Full Support**: OpenAI, Google, X.AI (Grok-4), Anthropic (Claude-4 series, Claude-3-Opus)
-- **Via OpenRouter**: Depends on the underlying model
-- **No Support**: DeepSeek, Mistral (except Large), Codex
+- **Full Support**: OpenAI, Google, X.AI (Grok 4.5), Anthropic (Claude-4 series, Claude-3-Opus)
+- **Mistral**: `mistral-medium-3-5` and `mistral-small-2603` accept images; `mistral-large-2512` is text-only
+- **Via OpenRouter**: Depends on the model — `qwen/qwen3.7-plus`, `moonshotai/kimi-k2.7-code`, and `moonshotai/kimi-k2.6` accept images; `z-ai/glm-5.2` and the `deepseek/deepseek-v4-*` slugs are text-only
+- **No Support**: DeepSeek (native), Codex
 
 ### Web Search
-- **Native Support**: OpenAI, Google, X.AI (Grok-4)
-- **No Support**: Anthropic, Mistral, DeepSeek, OpenRouter, Codex
+- **Automatic where supported**: OpenAI, Google, and X.AI (Grok 4.5, via Agent Tools) attach web search on every request for capable models; the model decides whether to use it
+- **OpenRouter (opt-in, adds cost)**: Off by default; append `:online` to a slug to enable it per request (real per-request charge), with citations captured into metadata
+- **No Support**: Anthropic, Mistral, DeepSeek, Codex
 
 ### Thinking/Reasoning Modes
-- **OpenAI**: O3 series models support `reasoning_effort` parameter
+- **OpenAI**: GPT-5 family and O3 series models support the `reasoning_effort` parameter (GPT-5.6 accepts `none` through `max`, mapping `minimal` to `low`; GPT-5 Pro is fixed at `high`)
 - **Google**:
   - Gemini 3.0 Pro: Thinking levels (low/high) via `reasoning_effort` - always enabled
   - Gemini 2.5 Pro/Flash: Thinking budget (token-based) via `reasoning_effort`
 - **Anthropic**: Claude Fable 5, Opus 4.6+, and Sonnet 4.6 use adaptive thinking (depth controlled by `reasoning_effort` via Anthropic's `effort` parameter); older Claude 4 models use budget-based extended thinking
+- **X.AI**: Grok 4.5 maps `reasoning_effort` to `low`/`medium`/`high` and always reasons (cannot be disabled)
+- **Mistral**: `mistral-medium-3-5` and `mistral-small-2603` map `reasoning_effort` to `high` (enabled) or `none` (disabled); `mistral-large-2512` has no adjustable reasoning
+- **DeepSeek**: V4 models use thinking mode via `reasoning_effort` (`none` disables; enabled levels use `high`, `max` uses `max`)
+- **OpenRouter**: Reasoning is per-model (effort-tiered, enable/disable-only, mandatory, or router-chosen — see the OpenRouter section)
 - **Codex**: Thread-based agentic reasoning with persistent context
 - **Others**: Standard inference only
 
@@ -280,7 +304,7 @@ All providers support streaming responses for real-time output.
 
 ## Model Selection in Tools
 
-When using the chat or consensus tools, specify models using their identifiers:
+When using the chat tool in any mode, specify models using their identifiers:
 
 ### Model Routing Logic
 
@@ -307,31 +331,34 @@ When using the chat or consensus tools, specify models using their identifiers:
 4. **OpenRouter Auto**: Special aliases route to OpenRouter's auto-selection:
    - "openrouter/auto", "openrouter auto", "auto router", "auto-router"
 
-```javascript
-// Chat tool examples
-{
-  "model": "gpt-5",               // OpenAI (keyword match)
-  "model": "fable",               // Anthropic (keyword match -> claude-fable-5)
-  "model": "opus",                // Anthropic (keyword match -> claude-opus-4-8)
-  "model": "sonnet",              // Anthropic (keyword match -> claude-sonnet-4-6)
-  "model": "claude",              // Claude Agent SDK (defaults to Claude Fable 5)
-  "model": "claude:opus",         // Claude Agent SDK (Claude Opus 4.8)
-  "model": "gemini-2.5-pro",      // Google (keyword match)
-  "model": "grok-4",              // X.AI (keyword match)
-  "model": "mistral-large",       // Mistral (keyword match)
-  "model": "deepseek-chat",       // DeepSeek (keyword match)
-  "model": "anthropic/claude-sonnet-4",   // OpenRouter (slash format, not in Anthropic)
-  "model": "qwen/qwen3-coder",            // OpenRouter (static model)
-  "model": "openrouter/auto"              // OpenRouter auto-selection
-}
+The `models` array always holds plain model-name strings. Each string routes as follows:
+
+```text
+"gpt-5.6"                  // OpenAI (keyword match)
+"fable"                    // Anthropic (keyword match -> claude-fable-5)
+"opus"                     // Anthropic (keyword match -> claude-opus-4-8)
+"sonnet"                   // Anthropic (keyword match -> claude-sonnet-4-6)
+"claude"                   // Claude Agent SDK (defaults to Claude Fable 5)
+"claude:opus"              // Claude Agent SDK (Claude Opus 4.8)
+"gemini-2.5-pro"           // Google (keyword match)
+"grok-4.5"                 // X.AI (keyword match)
+"mistral-large"            // Mistral (alias -> mistral-large-2512)
+"deepseek"                 // DeepSeek (alias -> deepseek-v4-pro)
+"z-ai/glm-5.2"             // OpenRouter (curated slug)
+"z-ai/glm-5.2:online"      // OpenRouter with web search opt-in
+"anthropic/claude-sonnet-5" // OpenRouter (any full slug routes as-is)
+"openrouter/auto"          // OpenRouter auto-selection
+```
 
-// Consensus tool with multiple providers
+```json
+// Consensus mode with multiple providers
 {
-  "models": [
-    {"model": "o3"},
-    {"model": "claude-sonnet-4-6"},
-    {"model": "gemini-2.5-pro"}
-  ]
+  "name": "chat",
+  "arguments": {
+    "prompt": "Which database fits our workload?",
+    "models": ["gpt-5.6", "claude", "gemini-2.5-pro"],
+    "mode": "consensus"
+  }
 }
 ```
 
@@ -353,14 +380,12 @@ When using the chat or consensus tools, specify models using their identifiers:
 - OpenRouter provides unified rate limiting across providers
 - Consider using multiple providers for better availability
 
-### OpenRouter Compliance
-- The `OPENROUTER_REFERER` header is **required**
-- Use your application URL or GitHub repository
-- This helps OpenRouter track usage for compliance
-
-### OpenRouter Dynamic Models
-- Enable with `OPENROUTER_DYNAMIC_MODELS=true`
-- First request to a new model may be slower (fetches capabilities)
-- Model capabilities are cached for 24 hours
-- Use any model from [openrouter.ai/models](https://openrouter.ai/models)
-- Models must use `provider/model` format (e.g., `meta-llama/llama-3.2-90b`)
+### OpenRouter Attribution
+- `OPENROUTER_REFERER` and `OPENROUTER_TITLE` are optional — set them to your application URL / name for OpenRouter ranking credit
+- Omitting them is valid and never blocks a request
+
+### OpenRouter Model Selection
+- The 8 curated slugs work out of the box; the default is `z-ai/glm-5.2`
+- Any other model works via its full `provider/model` slug or the `openrouter:` namespace — no extra configuration
+- The first request to a non-curated slug fetches its capabilities from OpenRouter's live catalog; a slug absent from that catalog fails before inference with `MODEL_NOT_FOUND`
+- Append `:online` to a slug to enable web search for that request (adds a real per-request cost)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "converse-mcp-server",
-  "version": "2.29.2",
+  "version": "3.0.1",
   "description": "Converse MCP Server - Converse with other LLMs with chat and consensus tools",
   "type": "module",
   "main": "src/index.js",

package/src/async/asyncJobStore.js

@@ -167,9 +167,9 @@ class LRUAsyncJobStore extends AsyncJobStoreInterface {
   async create(tool, options = {}) {
     try {
       // Validate parameters
-      if (!tool || !['chat', 'consensus', 'conversation'].includes(tool)) {
+      if (!tool || !['chat'].includes(tool)) {
         throw new AsyncJobStoreError(
-          'Invalid tool: must be "chat", "consensus", or "conversation"',
+          'Invalid tool: must be "chat"',
           'INVALID_TOOL',
         );
       }

package/src/async/jobRunner.js

@@ -81,7 +81,8 @@ export class JobRunner extends EventEmitter {
    * Submit a job for background execution
    * @param {object} jobSpec - Job specification
    * @param {string} jobSpec.sessionId - Session identifier
-   * @param {string} jobSpec.tool - Tool name ('chat' | 'consensus')
+   * @param {string} jobSpec.tool - Tool name (always 'chat')
+   * @param {string} jobSpec.mode - Chat mode ('chat' | 'consensus' | 'roundtable')
    * @param {Function} runFunction - Function to execute (async)
    * @param {object} options - Execution options
    * @param {number} options.timeout - Job timeout in ms
@@ -112,6 +113,7 @@ export class JobRunner extends EventEmitter {
       // Create job in store
       const jobId = await this.asyncJobStore.create(jobSpec.tool, {
         sessionId: jobSpec.sessionId,
+        mode: jobSpec.mode,
         timeout: options.timeout || this.defaultTimeout,
         priority: options.priority || false,
         ...jobSpec.options,
@@ -336,6 +338,7 @@ export class JobRunner extends EventEmitter {
               jobId,
               status: JOB_STATUS.COMPLETED,
               tool: finalJobState.tool,
+              mode: finalJobState.mode,
               result,
               completedAt: Date.now(),
               createdAt: finalJobState.createdAt,
@@ -396,6 +399,7 @@ export class JobRunner extends EventEmitter {
                 jobId,
                 status: JOB_STATUS.CANCELLED,
                 tool: cancelledJobState.tool,
+                mode: cancelledJobState.mode,
                 result: partialResult,
                 cancelledAt: Date.now(),
                 createdAt: cancelledJobState.createdAt,
@@ -446,6 +450,7 @@ export class JobRunner extends EventEmitter {
               jobId,
               status: JOB_STATUS.FAILED,
               tool: failedJobState.tool,
+              mode: failedJobState.mode,
               error: executionError.message || executionError,
               failedAt: Date.now(),
               createdAt: failedJobState.createdAt,

package/src/async/providerStreamNormalizer.js

@@ -24,6 +24,7 @@ const EVENT_TYPES = {
   END: 'end',
   ERROR: 'error',
   REASONING_SUMMARY: 'reasoning_summary',
+  REASONING: 'reasoning',
 };
 
 /**
@@ -174,6 +175,7 @@ class ProviderStreamNormalizer {
 
     let hasStarted = false;
     let accumulatedContent = '';
+    let accumulatedReasoning = '';
     let accumulatedUsage = null;
     let finishReason = null;
     const searchMetadata = {};
@@ -194,6 +196,13 @@ class ProviderStreamNormalizer {
           continue;
         }
 
+        // Handle reasoning/thinking events (kept separate from answer text)
+        if (event.type === 'thinking') {
+          accumulatedReasoning += event.content || '';
+          yield this.createReasoningEvent(event.content || '', provider, model);
+          continue;
+        }
+
         // Handle usage events (with XAI-specific search metadata)
         if (event.type === 'usage') {
           accumulatedUsage = event.usage;
@@ -212,6 +221,7 @@ class ProviderStreamNormalizer {
           const endMetadata = {
             ...event.metadata,
             ...searchMetadata,
+            ...(accumulatedReasoning && { reasoning: accumulatedReasoning }),
           };
 
           yield this.createEndEvent(
@@ -424,6 +434,7 @@ class ProviderStreamNormalizer {
 
     let hasStarted = false;
     let accumulatedContent = '';
+    let accumulatedReasoning = '';
     let accumulatedUsage = null;
     let finishReason = null;
 
@@ -443,6 +454,13 @@ class ProviderStreamNormalizer {
           continue;
         }
 
+        // Handle reasoning/thinking events (kept separate from answer text)
+        if (event.type === 'thinking') {
+          accumulatedReasoning += event.content || '';
+          yield this.createReasoningEvent(event.content || '', provider, model);
+          continue;
+        }
+
         // Handle usage events
         if (event.type === 'usage') {
           accumulatedUsage = event.usage;
@@ -453,13 +471,17 @@ class ProviderStreamNormalizer {
         // Handle end event
         if (event.type === 'end') {
           finishReason = event.stop_reason || event.metadata?.finish_reason;
+          const endMetadata = {
+            ...event.metadata,
+            ...(accumulatedReasoning && { reasoning: accumulatedReasoning }),
+          };
           yield this.createEndEvent(
             {
               content: event.content || accumulatedContent,
               stopReason: finishReason,
               usage: event.metadata?.usage || accumulatedUsage,
               responseTime: Date.now() - startTime,
-              metadata: event.metadata,
+              metadata: endMetadata,
             },
             provider,
             model,
@@ -494,6 +516,7 @@ class ProviderStreamNormalizer {
 
     let hasStarted = false;
     let accumulatedContent = '';
+    let accumulatedReasoning = '';
     let accumulatedUsage = null;
     let finishReason = null;
     const reasoningMetadata = {};
@@ -516,6 +539,13 @@ class ProviderStreamNormalizer {
           continue;
         }
 
+        // Handle reasoning/thinking events (kept separate from answer text)
+        if (event.type === 'thinking') {
+          accumulatedReasoning += event.content || '';
+          yield this.createReasoningEvent(event.content || '', provider, model);
+          continue;
+        }
+
         // Handle usage events (with DeepSeek-specific reasoning tokens)
         if (event.type === 'usage') {
           accumulatedUsage = event.usage;
@@ -532,6 +562,7 @@ class ProviderStreamNormalizer {
           const endMetadata = {
             ...event.metadata,
             ...reasoningMetadata,
+            ...(accumulatedReasoning && { reasoning: accumulatedReasoning }),
           };
 
           yield this.createEndEvent(
@@ -575,6 +606,7 @@ class ProviderStreamNormalizer {
 
     let hasStarted = false;
     let accumulatedContent = '';
+    let accumulatedReasoning = '';
     let accumulatedUsage = null;
     let finishReason = null;
     let routingMetadata = {};
@@ -595,6 +627,13 @@ class ProviderStreamNormalizer {
           continue;
         }
 
+        // Handle reasoning/thinking events (kept separate from answer text)
+        if (event.type === 'thinking') {
+          accumulatedReasoning += event.content || '';
+          yield this.createReasoningEvent(event.content || '', provider, model);
+          continue;
+        }
+
         // Handle usage events (with OpenRouter-specific routing metadata)
         if (event.type === 'usage') {
           accumulatedUsage = event.usage;
@@ -611,6 +650,7 @@ class ProviderStreamNormalizer {
           const endMetadata = {
             ...event.metadata,
             ...routingMetadata,
+            ...(accumulatedReasoning && { reasoning: accumulatedReasoning }),
           };
 
           yield this.createEndEvent(
@@ -771,6 +811,24 @@ class ProviderStreamNormalizer {
     };
   }
 
+  /**
+   * Create standardized streamed-reasoning event. Distinct from a delta so
+   * downstream consumers keep reasoning separate from the visible answer.
+   */
+  createReasoningEvent(content, provider, model) {
+    return {
+      type: EVENT_TYPES.REASONING,
+      provider,
+      model,
+      timestamp: Date.now(),
+      data: {
+        content: content || '',
+        role: 'assistant',
+        isReasoning: true,
+      },
+    };
+  }
+
   /**
    * Create standardized reasoning summary event
    */

package/src/config.js

@@ -222,13 +222,14 @@ const CONFIG_SCHEMA = {
     OPENROUTER_TITLE: {
       type: 'string',
       required: false,
-      description: 'OpenRouter X-Title header for request tracking',
+      description:
+        'Optional OpenRouter attribution title, sent as the canonical X-OpenRouter-Title header',
     },
     OPENROUTER_DYNAMIC_MODELS: {
       type: 'boolean',
       default: false,
       description:
-        'Enable dynamic model discovery via OpenRouter endpoints API',
+        'Accepted but unnecessary — full OpenRouter provider/model slugs always resolve without it',
     },
 
     // Google Vertex AI configuration
@@ -294,7 +295,7 @@ const CONFIG_SCHEMA = {
       type: 'string',
       required: false,
       description:
-        'Default model for Copilot SDK sessions (e.g., gpt-5, claude-sonnet-4.5)',
+        'Default model for Copilot SDK sessions (e.g., gpt-5.6-sol, claude-sonnet-5)',
     },
     COPILOT_CLI_PATH: {
       type: 'string',