converse-mcp-server 3.2.4 → 3.4.0

package/.env.example

@@ -80,9 +80,9 @@ OPENROUTER_API_KEY=your_openrouter_api_key_here
 # WARNING: Interactive policies may cause hangs in server/headless mode
 # CODEX_APPROVAL_POLICY=never
 
-# Default Codex model (default: gpt-5.6-sol)
-# Options: gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.4, gpt-5.4-mini, gpt-5.3-codex-spark
-# CODEX_MODEL=gpt-5.6-sol
+# Default Codex backend model (default: gpt-6-astra). Per-request override: models: ["codex:<model>"]
+# Options: gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5, gpt-5.3-codex-spark
+# CODEX_MODEL=gpt-6-astra
 
 # ============================================
 # Server Configuration

package/README.md

@@ -234,6 +234,7 @@ SUMMARIZATION_MODEL=gpt-5-nano        # Default: gpt-5-nano
 
 - **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-3.8-flash** (aliases: `gemini-3.8`, `flash-3.8`): Current-generation Flash with stronger long-horizon agentic performance (1M context, 65K output; thinking levels low/medium/high — no minimal)
 - **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 (1M context, 65K output)
 - **gemini-2.5-flash-lite** (alias: `flash-lite`): Lightweight fast model (1M context, 65K output)
@@ -280,7 +281,9 @@ Any other model works via its full `provider/model` slug or the `openrouter:` na
 
 ### Codex Models
 
-- **codex**: OpenAI Codex agentic coding assistant
+- **codex**: OpenAI Codex agentic coding assistant (GPT-6 Astra by default)
+  - Pick another backend per request with `codex:<model>` (e.g. `codex:sol`, `codex:gpt-5.6-terra`) or globally with `CODEX_MODEL`; backends: `gpt-6-astra` (alias `astra`), `gpt-5.6-sol` (`sol`), `gpt-5.6-terra` (`terra`), `gpt-5.6-luna` (`luna`), `gpt-5.5`, `gpt-5.3-codex-spark` (`spark`)
+  - `reasoning_effort` maps onto the tiers the chosen backend accepts (GPT-6 Astra: `low` through `max`, no `none`)
   - Thread-based sessions with persistent context
   - Direct filesystem access from working directory
   - Typical response time: 6-20 seconds (longer for complex tasks)
@@ -301,7 +304,7 @@ Reach these with the `copilot:` namespace (e.g. `copilot:gpt-5.6-terra`); uses y
 
 - **OpenAI**: `gpt-5.6-sol` (aliases: `gpt-5.6`, `gpt-5`), `gpt-5.6-terra`, `gpt-5.6-luna` (all support `reasoning_effort`)
 - **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`)
+- **Google**: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.8-flash` (aliases: `gemini-3.8`, `flash-3.8`), `gemini-3.5-flash` (alias: `gemini-flash`)
 - Any other `copilot:<id>` is forwarded to the Copilot backend verbatim
 
 ## 📚 Help & Documentation
@@ -356,6 +359,7 @@ CODEX_API_KEY=your_codex_api_key_here       # Optional if ChatGPT login availabl
 CODEX_SANDBOX_MODE=read-only                 # read-only (default), workspace-write, danger-full-access
 CODEX_SKIP_GIT_CHECK=true                    # true (default), false
 CODEX_APPROVAL_POLICY=never                  # never (default), untrusted, on-failure, on-request
+CODEX_MODEL=gpt-6-astra                      # gpt-6-astra (default), gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5
 ```
 
 ### Configuration Options
@@ -371,18 +375,42 @@ CODEX_APPROVAL_POLICY=never                  # never (default), untrusted, on-fa
 
 These must be set in your system environment or when launching Claude Code, NOT in the project .env file:
 
-| Variable                | Description                 | Default  | Example                              |
-| ----------------------- | --------------------------- | -------- | ------------------------------------ |
-| `MAX_MCP_OUTPUT_TOKENS` | Token response limit        | `25000`  | `200000`                             |
-| `MCP_TOOL_TIMEOUT`      | Tool execution timeout (ms) | `120000` | `5400000` (90 min for deep research) |
+| Variable                             | Description                                                             | Default                         | Example                     |
+| ------------------------------------ | ----------------------------------------------------------------------- | ------------------------------- | --------------------------- |
+| `MAX_MCP_OUTPUT_TOKENS`              | Token response limit                                                    | `25000`                         | `200000`                    |
+| `MCP_TOOL_TIMEOUT`                   | Wall-clock limit per tool call (ms)                                     | ~28 hours when unset            | `7200000` (2 h)             |
+| `CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT`  | Idle window (ms) — aborts a call that produces no output for this long  | 30 min (stdio) / 5 min (HTTP)   | `3600000` (60 min)          |
 
 ```bash
 # Example: Set globally before starting Claude Code
 export MAX_MCP_OUTPUT_TOKENS=200000
-export MCP_TOOL_TIMEOUT=5400000  # 90 minutes for deep research models
+export CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT=3600000  # 60 min for long silent agentic calls
 claude  # Then start Claude Code
 ```
 
+Or persist them in `~/.claude/settings.json`:
+
+```json
+{
+	"env": {
+		"MAX_MCP_OUTPUT_TOKENS": "200000",
+		"CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT": "3600000"
+	}
+}
+```
+
+**The idle timeout is usually what kills long calls.** Agentic models (Codex, Claude Agent SDK) can work silently for 30+ minutes; Converse holds one MCP request open the whole time, and Claude Code aborts it after the idle window with an error like _"failed after 30 minutes of silence. The idle timeout aborted it."_ Progress-notification heartbeats can't prevent this — Claude Code doesn't send a `progressToken` on `tools/call`, so an MCP server has no spec-compliant way to emit them ([claude-code#58687](https://github.com/anthropics/claude-code/issues/58687)). Raising the idle window is the only fix.
+
+#### Codex CLI Tool Timeout
+
+If you register Converse in OpenAI's Codex CLI, note that Codex enforces its own **hard 300-second default** per MCP tool call (`tool_timeout_sec`, undocumented — it exists only in Codex's config schema). Progress notifications don't extend it. Raise it in `~/.codex/config.toml`:
+
+```toml
+[mcp_servers.converse]
+# ... command/env ...
+tool_timeout_sec = 3600  # default 300 kills long calls at 5 minutes
+```
+
 ### Model Selection
 
 Use `"auto"` for automatic model selection, or specify exact models:
@@ -421,7 +449,7 @@ Use `"auto"` for automatic model selection, or specify exact models:
 Provider priority order (subscription-based SDK providers first, then API-key providers):
 
 1. Codex (`codex`)
-2. Gemini via Antigravity CLI (`gemini`, `gemini:flash`)
+2. Gemini via Antigravity CLI (`gemini` → Gemini 3.8 Flash, `gemini:pro`)
 3. Claude Agent SDK (`claude` → Claude Fable 5)
 4. Copilot (`copilot`)
 5. OpenAI (`gpt-5.6`)
@@ -532,6 +560,13 @@ For development setup, see the [Development](#-development) section below.
 
 - Clear cache and reinstall: `npm run clean`
 
+**Long tool calls aborted mid-run (idle/timeout errors):**
+
+- The abort almost always comes from the MCP _client_, not Converse — Converse's own limits are 30 min per provider call and 90 min per async job.
+- Claude Code: raise `CLAUDE_CODE_MCP_TOOL_IDLE_TIMEOUT` (idle window, default 30 min stdio / 5 min HTTP) and check `MCP_TOOL_TIMEOUT` (wall-clock). See [Claude Code Environment Variables](#claude-code-environment-variables-systemglobal).
+- Codex CLI: set `tool_timeout_sec` under `[mcp_servers.converse]` in `~/.codex/config.toml` — the undocumented default is 300 seconds.
+- Immune alternative: run the call with `async: true` and poll `check_status` — each poll is a fresh short request, so no client timeout applies.
+
 ### Debug Mode
 
 ```bash

package/docs/API.md

@@ -391,6 +391,7 @@ Provide models as plain name strings in the `models` array. Bare names and alias
 |-------|---------|---------|--------|-------|
 | `gemini-3.1-pro-preview` | `pro`, `gemini-pro` | 1M | 64K | Most advanced reasoning, expanded thinking levels |
 | `gemini-3.5-flash` | `gemini-3.5`, `flash-3.5` | 1M | 65K | Frontier agentic/coding at Flash speed |
+| `gemini-3.8-flash` | `gemini-3.8`, `flash-3.8` | 1M | 65K | Current-generation Flash; thinking low/medium/high only |
 | `gemini-2.5-pro` | `pro 2.5` | 1M | 65K | Deep reasoning with thinking budget |
 | `gemini-2.5-flash` | `flash` | 1M | 65K | Ultra-fast |
 | `gemini-2.5-flash-lite` | `flash-lite` | 1M | 65K | Lightweight fast model |
@@ -455,12 +456,13 @@ Any other model works via its full `provider/model` slug (e.g. `anthropic/claude
 
 **Codex** is an agentic coding assistant with direct filesystem access:
 
-- **Model**: `codex` (underlying model: GPT-5.6)
+- **Model**: `codex` (underlying model: GPT-6 Astra by default)
+- **Backend selection**: `codex:<model>` per request (e.g. `codex:astra`, `codex:sol`, `codex:gpt-5.6-terra`), or `CODEX_MODEL` globally; unknown names pass through to the CLI verbatim
 - **Thread-based sessions**: persistent conversation history via `continuation_id` in `chat` mode
 - **Direct file access**: reads files from the working directory (paths relative to `CLIENT_CWD`)
 - **Response times**: 6-20 seconds typical (complex tasks may take minutes)
 - **Authentication**: ChatGPT login OR `CODEX_API_KEY` (NOT `OPENAI_API_KEY`)
-- `reasoning_effort` and web search are not applicable — Codex manages its own execution
+- `reasoning_effort` is clamped onto the tiers the chosen backend accepts (GPT-6 Astra: `low`–`max`, no `none`; GPT-5.6: `none`–`max`); web search is not applicable — Codex manages its own execution
 
 ### Claude Agent SDK (subscription)
 
@@ -476,7 +478,7 @@ Any other model works via its full `provider/model` slug (e.g. `anthropic/claude
 
 The **Antigravity CLI** (`agy`) provides subscription-based access to Gemini models through Google OAuth:
 
-- **Models** (text-only): `gemini` (= `gemini:pro`, Gemini 3.1 Pro), `gemini:flash` (Gemini 3.5 Flash)
+- **Models** (text-only): `gemini` (= `gemini:flash`, Gemini 3.8 Flash), `gemini:pro` (Gemini 3.1 Pro)
 - **Authentication**: Google OAuth via `agy` (one-time interactive login)
 - **Setup**: install the Antigravity CLI and run `agy` once to log in
 - **Billing**: uses your Antigravity subscription/compute allowance instead of API credits
@@ -502,7 +504,7 @@ Reach these with the `copilot:` namespace (e.g. `copilot:gpt-5.6-terra`); uses y
 
 - **OpenAI**: `gpt-5.6-sol` (aliases: `gpt-5.6`, `gpt-5`), `gpt-5.6-terra`, `gpt-5.6-luna` (all accept `reasoning_effort`)
 - **Anthropic**: `claude-fable-5` (alias: `fable`), `claude-sonnet-5` (alias: `sonnet`), `claude-opus-5` (aliases: `opus`, `claude`), `claude-opus-4.8`
-- **Google**: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.5-flash` (alias: `gemini-flash`)
+- **Google**: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.8-flash` (aliases: `gemini-3.8`, `flash-3.8`), `gemini-3.5-flash` (alias: `gemini-flash`)
 - Any other `copilot:<id>` is forwarded to the Copilot backend verbatim
 
 ### Model Selection
@@ -522,7 +524,7 @@ Use `"auto"` for automatic selection, or specify exact models:
 "opus"                     // Anthropic API (-> claude-opus-5)
 "claude"                   // Claude Agent SDK (-> Claude Fable 5)
 "claude:opus"              // Claude Agent SDK (Claude Opus 5)
-"gemini"                   // Antigravity CLI (Gemini 3.1 Pro)
+"gemini"                   // Antigravity CLI (Gemini 3.8 Flash)
 "copilot:gpt-5.6-terra"    // GitHub Copilot SDK
 ```
 

package/docs/PROVIDERS.md

@@ -24,6 +24,7 @@ This guide documents all supported AI providers in the Converse MCP Server and t
 - **Supported Models**:
   - `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-3.8-flash` (aliases: `gemini-3.8`, `flash-3.8`) - Current-generation Flash with stronger long-horizon agentic performance (1M context, 65K output; thinking levels low/medium/high — no minimal)
   - `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)
   - `gemini-2.5-flash-lite` (alias: `flash-lite`) - Lightweight fast model (1M context, 65K output)
@@ -100,9 +101,11 @@ This guide documents all supported AI providers in the Converse MCP Server and t
   - `CODEX_SANDBOX_MODE` - Filesystem access control (default: read-only)
   - `CODEX_SKIP_GIT_CHECK` - Skip Git repository validation (default: true)
   - `CODEX_APPROVAL_POLICY` - Command approval behavior (default: never)
-  - `CODEX_MODEL` - Underlying model for Codex sessions (default: gpt-5.6-sol; e.g. gpt-5.6-terra, gpt-5.6-luna, gpt-5.5)
+  - `CODEX_MODEL` - Underlying model for Codex sessions (default: gpt-6-astra; e.g. gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5)
 - **Supported Models**:
-  - `codex` - OpenAI Codex agentic coding assistant (GPT-5.6 by default)
+  - `codex` - OpenAI Codex agentic coding assistant (GPT-6 Astra by default)
+  - `codex:<model>` - Same, with an explicit backend: `codex:astra`, `codex:sol`, `codex:terra`, `codex:luna`, `codex:gpt-5.5`, `codex:spark`, or any slug the Codex CLI knows
+  - `reasoning_effort` is clamped onto what the backend accepts (GPT-6 Astra: `low`–`max`, no `none`)
   - Thread-based sessions with persistent context
   - Direct filesystem access from working directory
   - Typical response time: 6-20 seconds (longer for complex tasks)
@@ -129,8 +132,8 @@ This guide documents all supported AI providers in the Converse MCP Server and t
   2. Authenticate: run `agy` once interactively and complete the Google OAuth login. This also establishes workspace trust for your home directory (the provider spawns each call in a per-call subdirectory under `~/.converse/agy-runs`).
 - **Environment Variables**: None (the provider detects the `agy` binary on PATH or at the platform install location)
 - **Supported Models** (text-only — print mode has no image input channel):
-  - `gemini` (= `gemini:pro`) - Gemini 3.1 Pro
-  - `gemini:flash` - Gemini 3.5 Flash
+  - `gemini` (= `gemini:flash`) - Gemini 3.8 Flash (default)
+  - `gemini:pro` - Gemini 3.1 Pro
   - `reasoning_effort` selects the variant: `low` → (Low), `medium` → (Medium) for Flash / (High) for Pro, `high`/`max` → (High); unset defaults to (High)
 
 **Key Features:**
@@ -216,7 +219,7 @@ agy
   - `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-5` (aliases: `opus`, `claude`), `claude-opus-4.8`
-  - Google: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.5-flash` (alias: `gemini-flash`)
+  - Google: `gemini-3.1-pro-preview` (aliases: `gemini`, `gemini-3.1-pro`), `gemini-3.8-flash` (aliases: `gemini-3.8`, `flash-3.8`), `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.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "converse-mcp-server",
-  "version": "3.2.4",
+  "version": "3.4.0",
   "description": "Converse MCP Server - Converse with other LLMs with chat and consensus tools",
   "type": "module",
   "main": "src/index.js",
@@ -93,28 +93,28 @@
     ".env.example"
   ],
   "dependencies": {
-    "@anthropic-ai/claude-agent-sdk": "^0.3.220",
-    "@anthropic-ai/sdk": "^0.115.0",
-    "@github/copilot-sdk": "^1.0.8",
-    "@google/genai": "^2.13.0",
-    "@lydell/node-pty": "1.2.0-beta.14",
-    "@mistralai/mistralai": "^2.5.0",
-    "@modelcontextprotocol/sdk": "^1.29.0",
-    "@openai/codex-sdk": "^0.145.0",
+    "@anthropic-ai/claude-agent-sdk": "^0.3.261",
+    "@anthropic-ai/sdk": "^0.123.0",
+    "@github/copilot-sdk": "^1.0.11",
+    "@google/genai": "^2.21.0",
+    "@lydell/node-pty": "1.2.0-beta.15",
+    "@mistralai/mistralai": "^2.6.4",
+    "@modelcontextprotocol/sdk": "^1.30.0",
+    "@openai/codex-sdk": "^0.153.3",
     "cors": "^2.8.6",
     "dotenv": "^17.4.2",
     "express": "^5.2.1",
     "lru-cache": "^11.5.2",
-    "nanoid": "^6.0.0",
-    "openai": "^6.49.0",
-    "p-limit": "^7.3.1",
-    "vite": "^8.1.5"
+    "nanoid": "^6.0.1",
+    "openai": "^7.10.0",
+    "p-limit": "^7.3.2",
+    "vite": "^8.2.2"
   },
   "devDependencies": {
-    "@vitest/coverage-v8": "^4.1.10",
+    "@vitest/coverage-v8": "^5.0.0",
     "cross-env": "^10.1.0",
-    "eslint": "^10.8.0",
+    "eslint": "^10.10.0",
     "rimraf": "^6.1.3",
-    "vitest": "^4.1.10"
+    "vitest": "^5.0.0"
   }
 }

package/src/config.js

@@ -280,9 +280,9 @@ const CONFIG_SCHEMA = {
     },
     CODEX_MODEL: {
       type: 'string',
-      default: 'gpt-5.6-sol',
+      default: 'gpt-6-astra',
       description:
-        'Default Codex model (e.g., gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5)',
+        'Default Codex backend model (e.g., gpt-6-astra, gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5)',
     },
 
     // Copilot configuration

package/src/providers/codex.js

@@ -17,35 +17,75 @@ import { debugLog, debugError } from '../utils/console.js';
 import { ProviderError, ErrorCodes, StopReasons } from './interface.js';
 import { normalizeExtendedPath } from '../utils/pathUtils.js';
 
-// Supported Codex models with their configurations
+/**
+ * Every Codex reasoning tier, weakest to strongest. Used to clamp a requested
+ * tier onto the set a given backend model accepts.
+ *
+ * Codex also exposes 'ultra' above 'max', but that tier turns on automatic
+ * sub-agent delegation — a change in how the run executes, not just how deep
+ * it reasons — so nothing at the tool level maps to it and it is kept off the
+ * ladder so the clamp can never select it.
+ */
+const EFFORT_LADDER = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh', 'max'];
+
+/**
+ * Backend models Codex can run, keyed by the slug passed to the CLI as
+ * --model. The reasoning tiers are the ones each model's API accepts, verified
+ * against the API's own rejection messages (gpt-6-astra: "Supported values
+ * are: 'low', 'medium', 'high', 'xhigh', and 'max'"). The SDK's
+ * ModelReasoningEffort type is the union across models, so the backend is the
+ * authority and requests are clamped per model.
+ */
+const CODEX_BACKEND_MODELS = {
+  'gpt-6-astra': {
+    aliases: ['astra', 'gpt-6', 'gpt6', 'gpt6-astra'],
+    contextWindow: 272000,
+    supportedEfforts: ['low', 'medium', 'high', 'xhigh', 'max'],
+  },
+  'gpt-5.6-sol': {
+    aliases: ['sol', 'gpt-5.6', 'gpt5.6', 'gpt-5.6-codex'],
+    contextWindow: 272000,
+    supportedEfforts: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+  },
+  'gpt-5.6-terra': {
+    aliases: ['terra'],
+    contextWindow: 272000,
+    supportedEfforts: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+  },
+  'gpt-5.6-luna': {
+    aliases: ['luna'],
+    contextWindow: 272000,
+    supportedEfforts: ['none', 'low', 'medium', 'high', 'xhigh', 'max'],
+  },
+  'gpt-5.5': {
+    aliases: ['gpt5.5'],
+    contextWindow: 272000,
+    supportedEfforts: ['low', 'medium', 'high', 'xhigh'],
+  },
+  'gpt-5.3-codex-spark': {
+    aliases: ['spark', 'codex-spark'],
+    contextWindow: 128000,
+    supportedEfforts: ['low', 'medium', 'high', 'xhigh'],
+  },
+};
+
+const DEFAULT_BACKEND_MODEL = 'gpt-6-astra';
+
+// The single user-facing model the router exposes. The backend model behind it
+// comes from CODEX_MODEL, or from a `codex:<model>` spec.
 const SUPPORTED_MODELS = {
   codex: {
     modelName: 'codex',
-    friendlyName: 'OpenAI Codex (GPT-5.6)',
-    contextWindow: 400000,
+    friendlyName: 'OpenAI Codex (GPT-6 Astra)',
+    contextWindow: CODEX_BACKEND_MODELS[DEFAULT_BACKEND_MODEL].contextWindow,
     maxOutputTokens: 128000,
     supportsStreaming: true,
     supportsImages: true, // Codex SDK 0.118+ supports images via --image (local_image input)
     supportsWebSearch: false, // Codex accesses files directly, not web
-    // Reasoning tiers this model's backend actually accepts. GPT-5.6 dropped
-    // 'minimal' and added 'none', while the SDK's ModelReasoningEffort type
-    // still advertises the pre-5.6 set — the backend is the authority, so the
-    // accepted tiers are declared per model and requests are clamped onto them.
-    supportedEfforts: ['none', 'low', 'medium', 'high', 'xhigh'],
     timeout: 1800000, // 30 minutes
     description:
-      'OpenAI Codex agentic coding assistant with local file access and tool execution (GPT-5.6)',
-    aliases: [
-      'gpt-5-codex',
-      'gpt5-codex',
-      'gpt-5.2-codex',
-      'gpt-5.3-codex',
-      'gpt5.3-codex',
-      'gpt-5.5',
-      'gpt5.5',
-      'gpt-5.6-codex',
-      'gpt5.6-codex',
-    ],
+      'OpenAI Codex agentic coding assistant with local file access and tool execution (GPT-6 Astra by default; pick another backend with codex:<model> or CODEX_MODEL)',
+    aliases: [],
   },
 };
 
@@ -195,6 +235,57 @@ function extractPromptText(input) {
     .join('\n\n');
 }
 
+/**
+ * Resolve a backend slug or alias (case-insensitive) to its catalog slug.
+ * @param {string} name
+ * @returns {string|null}
+ */
+function findBackendSlug(name) {
+  const lower = String(name || '').trim().toLowerCase();
+  if (!lower) {
+    return null;
+  }
+  if (CODEX_BACKEND_MODELS[lower]) {
+    return lower;
+  }
+  return (
+    Object.keys(CODEX_BACKEND_MODELS).find((slug) =>
+      CODEX_BACKEND_MODELS[slug].aliases.includes(lower),
+    ) || null
+  );
+}
+
+/**
+ * Catalog entry for a backend model, or null when the slug is not catalogued.
+ * @param {string} name - Backend slug or alias
+ * @returns {{ slug: string, aliases: string[], contextWindow: number, supportedEfforts: string[] }|null}
+ */
+export function getBackendModelConfig(name) {
+  const slug = findBackendSlug(name);
+  return slug ? { slug, ...CODEX_BACKEND_MODELS[slug] } : null;
+}
+
+/**
+ * Resolve the requested model spec to the backend slug passed to the CLI.
+ *
+ * `codex` uses CODEX_MODEL (default gpt-6-astra); `codex:<model>` names a
+ * backend directly, by slug or alias. Unknown names pass through verbatim so a
+ * newly released model works before it is catalogued here — the CLI rejects
+ * anything the backend does not know.
+ *
+ * @param {string} spec - Requested model, e.g. 'codex' or 'codex:sol'
+ * @param {Object} [config] - Loaded configuration
+ * @returns {string} Backend slug for the SDK's `model` option
+ */
+export function resolveBackendModel(spec, config) {
+  const raw = String(spec || '').trim();
+  const requested = raw.toLowerCase().startsWith('codex:')
+    ? raw.slice('codex:'.length).trim()
+    : '';
+  const name = requested || config?.providers?.codexmodel || DEFAULT_BACKEND_MODEL;
+  return findBackendSlug(name) || name;
+}
+
 /**
  * Get thread ID from continuation metadata.
  * Codex thread IDs are stored per call-plan in `providerThreads`, keyed by a
@@ -216,12 +307,6 @@ async function getThreadIdFromContinuation(
   }
 }
 
-/**
- * Every Codex reasoning tier, weakest to strongest. Used to clamp a requested
- * tier onto the set a given model actually accepts.
- */
-const EFFORT_LADDER = ['none', 'minimal', 'low', 'medium', 'high', 'xhigh'];
-
 /**
  * Tool-level reasoning_effort values translated to their Codex equivalent.
  * Tool enum: 'none' | 'minimal' | 'low' | 'medium' | 'high' | 'max'
@@ -232,7 +317,7 @@ const EFFORT_ALIASES = {
   low: 'low',
   medium: 'medium',
   high: 'high',
-  max: 'xhigh',
+  max: 'max',
 };
 
 /**
@@ -403,8 +488,9 @@ export const codexProvider = {
       const approvalPolicy = config.providers?.codexapprovalpolicy || 'never';
 
       // Create or resume thread
+      const backendModel = resolveBackendModel(model, config);
       const threadOptions = {
-        model: config.providers?.codexmodel,
+        model: backendModel,
         workingDirectory,
         sandboxMode,
         skipGitRepoCheck,
@@ -413,13 +499,12 @@ export const codexProvider = {
 
       if (reasoning_effort) {
         const supportedEfforts =
-          findModelConfig(model)?.supportedEfforts ||
-          SUPPORTED_MODELS.codex.supportedEfforts;
+          getBackendModelConfig(backendModel)?.supportedEfforts || EFFORT_LADDER;
         const mappedEffort = mapReasoningEffort(reasoning_effort, supportedEfforts);
         threadOptions.modelReasoningEffort = mappedEffort;
         if (mappedEffort !== EFFORT_ALIASES[reasoning_effort]) {
           debugLog(
-            `[Codex] reasoning_effort "${reasoning_effort}" not supported by ${model} — using "${mappedEffort}"`,
+            `[Codex] reasoning_effort "${reasoning_effort}" not supported by ${backendModel} — using "${mappedEffort}"`,
           );
         }
       }
@@ -470,6 +555,7 @@ export const codexProvider = {
         metadata: {
           provider: 'codex',
           model,
+          backendModel,
           threadId: threadIdFromStream || thread.id,
           usage: usage
             ? {

package/src/providers/gemini-cli.js

@@ -20,7 +20,9 @@
  *
  * The provider registry key remains 'gemini-cli' and the user-facing alias
  * remains 'gemini' for routing/normalization stability. Only three user-facing
- * model names are exposed: gemini (= gemini:pro), gemini:pro, gemini:flash.
+ * model names are exposed: gemini (= gemini:flash), gemini:flash, gemini:pro.
+ * Flash is the default: Gemini 3.8 Flash is the current-generation model agy
+ * lists first, while 3.1 Pro remains the only Pro tier Antigravity offers.
  */
 
 import { existsSync, mkdirSync, writeFileSync, rmSync } from 'node:fs';
@@ -57,7 +59,7 @@ const PTY_COLS = 1000;
 const SUPPORTED_MODELS = {
   gemini: {
     modelName: 'gemini',
-    friendlyName: 'Gemini 3.1 Pro (via Antigravity CLI)',
+    friendlyName: 'Gemini 3.8 Flash (via Antigravity CLI)',
     contextWindow: 1048576,
     maxOutputTokens: 65536,
     supportsStreaming: true,
@@ -66,14 +68,14 @@ const SUPPORTED_MODELS = {
     supportsThinking: true,
     timeout: DEFAULT_TIMEOUT_MS,
     description:
-      'Gemini 3.1 Pro via Antigravity CLI (agy) - requires Antigravity Google OAuth login',
+      'Gemini 3.8 Flash via Antigravity CLI (agy) - requires Antigravity Google OAuth login',
     aliases: ['gemini-cli'],
     // agy display-name base; reasoning_effort selects the parenthesized variant
-    agyModelBase: 'Gemini 3.1 Pro',
+    agyModelBase: 'Gemini 3.8 Flash',
   },
-  'gemini:pro': {
-    modelName: 'gemini:pro',
-    friendlyName: 'Gemini 3.1 Pro (via Antigravity CLI)',
+  'gemini:flash': {
+    modelName: 'gemini:flash',
+    friendlyName: 'Gemini 3.8 Flash (via Antigravity CLI)',
     contextWindow: 1048576,
     maxOutputTokens: 65536,
     supportsStreaming: true,
@@ -82,13 +84,13 @@ const SUPPORTED_MODELS = {
     supportsThinking: true,
     timeout: DEFAULT_TIMEOUT_MS,
     description:
-      'Gemini 3.1 Pro via Antigravity CLI (agy) - explicit alias of `gemini`',
-    aliases: [],
-    agyModelBase: 'Gemini 3.1 Pro',
+      'Gemini 3.8 Flash via Antigravity CLI (agy) - explicit alias of `gemini`',
+    aliases: ['flash'],
+    agyModelBase: 'Gemini 3.8 Flash',
   },
-  'gemini:flash': {
-    modelName: 'gemini:flash',
-    friendlyName: 'Gemini 3.5 Flash (via Antigravity CLI)',
+  'gemini:pro': {
+    modelName: 'gemini:pro',
+    friendlyName: 'Gemini 3.1 Pro (via Antigravity CLI)',
     contextWindow: 1048576,
     maxOutputTokens: 65536,
     supportsStreaming: true,
@@ -97,9 +99,9 @@ const SUPPORTED_MODELS = {
     supportsThinking: true,
     timeout: DEFAULT_TIMEOUT_MS,
     description:
-      'Gemini 3.5 Flash via Antigravity CLI (agy) - requires Antigravity Google OAuth login',
-    aliases: ['flash'],
-    agyModelBase: 'Gemini 3.5 Flash',
+      'Gemini 3.1 Pro via Antigravity CLI (agy) - requires Antigravity Google OAuth login',
+    aliases: ['pro'],
+    agyModelBase: 'Gemini 3.1 Pro',
   },
 };
 
@@ -175,7 +177,7 @@ export function findAgyBinary() {
 /**
  * Map a reasoning_effort value to the agy parenthesized variant suffix.
  * Flash supports Low/Medium/High; Pro supports Low/High (no Medium).
- * @param {string} base - agy model base ('Gemini 3.5 Flash' / 'Gemini 3.1 Pro')
+ * @param {string} base - agy model base ('Gemini 3.8 Flash' / 'Gemini 3.1 Pro')
  * @param {string} [reasoningEffort]
  * @returns {string} e.g. '(Low)', '(Medium)', '(High)'
  */
@@ -207,7 +209,7 @@ function effortSuffix(base, reasoningEffort) {
  * verbatim so power users aren't blocked.
  * @param {string} model - e.g. 'gemini', 'gemini:flash', or a full agy name
  * @param {string} [reasoningEffort]
- * @returns {string} agy --model value, e.g. 'Gemini 3.1 Pro (High)'
+ * @returns {string} agy --model value, e.g. 'Gemini 3.8 Flash (High)'
  */
 export function resolveAgyModel(model, reasoningEffort) {
   const raw = typeof model === 'string' ? model.trim() : '';
@@ -230,11 +232,11 @@ export function resolveAgyModel(model, reasoningEffort) {
     !nameLower ||
     nameLower === 'gemini' ||
     nameLower === 'gemini-cli' ||
-    nameLower === 'pro'
+    nameLower === 'flash'
   ) {
     base = SUPPORTED_MODELS.gemini.agyModelBase;
-  } else if (nameLower === 'flash') {
-    base = SUPPORTED_MODELS['gemini:flash'].agyModelBase;
+  } else if (nameLower === 'pro') {
+    base = SUPPORTED_MODELS['gemini:pro'].agyModelBase;
   } else {
     // Unknown suffix: pass through verbatim (power-user agy display name)
     return raw;
@@ -762,19 +764,16 @@ export const geminiCliProvider = {
 
     const name = modelName.toLowerCase().trim();
 
-    // Full agy display-name passthrough → matching base config.
-    if (/gemini 3\.5 flash/i.test(modelName)) {
+    // Full agy display-name passthrough → matching tier config. Any 3.x Flash
+    // (agy also lists 3.6/3.7) shares the Flash config; any 3.x Pro the Pro one.
+    if (/gemini 3\.\d+ flash/i.test(modelName)) {
       return SUPPORTED_MODELS['gemini:flash'];
     }
-    if (/gemini 3\.1 pro/i.test(modelName)) {
-      return SUPPORTED_MODELS.gemini;
-    }
-
-    if (name === 'pro') {
+    if (/gemini 3\.\d+ pro/i.test(modelName)) {
       return SUPPORTED_MODELS['gemini:pro'];
     }
 
-    // Exact key match (gemini, gemini:pro, gemini:flash)
+    // Exact key match (gemini, gemini:flash, gemini:pro)
     if (SUPPORTED_MODELS[name]) {
       return SUPPORTED_MODELS[name];
     }

package/src/providers/google.js

@@ -127,6 +127,32 @@ const SUPPORTED_MODELS = {
       '3.5-flash',
     ],
   },
+  'gemini-3.8-flash': {
+    modelName: 'gemini-3.8-flash',
+    friendlyName: 'Gemini (Flash 3.8)',
+    contextWindow: 1048576, // 1M tokens
+    maxOutputTokens: 65536,
+    supportsStreaming: true,
+    supportsImages: true,
+    supportsThinking: true,
+    supportsWebSearch: true,
+    thinkingMode: 'level',
+    // 3.8 Flash rejects thinkingLevel "minimal" outright (API error), unlike 3.5.
+    thinkingLevels: ['low', 'medium', 'high'],
+    timeout: 900000,
+    description:
+      'Gemini 3.8 Flash - Current-generation Flash with stronger long-horizon agentic performance (1M context)',
+    aliases: [
+      'gemini-3.8',
+      'gemini3.8',
+      'gemini-3.8-flash-latest',
+      'flash-3.8',
+      'flash3.8',
+      'gemini-flash-3.8',
+      'gemini flash 3.8',
+      '3.8-flash',
+    ],
+  },
 };
 
 // Thinking mode budget percentages
@@ -514,10 +540,17 @@ export const googleProvider = {
           };
           thinkingLevel = levelMap[reasoning_effort] || 'high';
           if (!modelConfig.thinkingLevels.includes(thinkingLevel)) {
+            // Clamp to the nearest supported level: a request below the
+            // model's floor (e.g. minimal on 3.8 Flash) takes the lowest
+            // level, anything else the highest.
+            const rank = ['minimal', 'low', 'medium', 'high'];
+            const [lowest] = modelConfig.thinkingLevels;
             thinkingLevel =
-              modelConfig.thinkingLevels[
-                modelConfig.thinkingLevels.length - 1
-              ];
+              rank.indexOf(thinkingLevel) < rank.indexOf(lowest)
+                ? lowest
+                : modelConfig.thinkingLevels[
+                  modelConfig.thinkingLevels.length - 1
+                ];
           }
         } else {
           // Binary levels only (Gemini 3.0 Pro: low/high)

package/src/utils/modelRouting.js

@@ -138,9 +138,10 @@ function classifyModelSpec(spec, providers) {
   }
 
   // Other explicit namespaces pass the spec through unchanged; the target
-  // provider strips its own prefix (preserves current copilot/claude/gemini-cli
+  // provider strips its own prefix (preserves current codex/copilot/claude/gemini-cli
   // behavior).
   if (
+    lower.startsWith('codex:') ||
     lower.startsWith('copilot:') ||
     lower.startsWith('claude:') ||
     lower.startsWith('gemini:')
@@ -308,8 +309,9 @@ export function mapModelToProvider(model, providers) {
     return 'openai';
   }
 
-  // Check Codex (exact match only - don't route "gpt-5-codex" etc to Codex provider)
-  if (modelLower === 'codex') {
+  // Check Codex: bare `codex` or the `codex:<backend>` namespace (e.g. codex:astra,
+  // codex:gpt-5.6-sol). Bare "gpt-*" names never route here — they are OpenAI API models.
+  if (modelLower === 'codex' || modelLower.startsWith('codex:')) {
     return 'codex';
   }