@qduc/term2 0.31.1 → 0.32.1

package/README.md

@@ -1,418 +1,100 @@
 # term2
 
 [![npm version](https://img.shields.io/npm/v/@qduc/term2.svg?style=flat-square)](https://www.npmjs.com/package/@qduc/term2)
-[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg?style=flat-square)](https://opensource.org/licenses/MIT)
-[![Node.js](https://img.shields.io/badge/node-%3E%3D20-brightgreen.svg?style=flat-square)](https://nodejs.org)
-[![TypeScript](https://img.shields.io/badge/TypeScript-5.x-blue.svg?style=flat-square)](https://www.typescriptlang.org/)
+[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg?style=flat-square)](LICENSE)
 
-term2 is an open-source terminal AI assistant and autonomous agent runtime built for high-velocity software engineering and system administration.
+**term2 is an open-source AI assistant for coding and terminal work.**
 
-Combining a fast React (Ink) terminal interface with an application-owned run loop, term2 provides fine-grained execution safety, native multi-agent delegation, provider-neutral context compaction, and seamless multi-provider support—giving developers complete agency over their workflow without vendor lock-in or subscription gates.
+Read a codebase, edit files, and run commands locally or over SSH. Choose your model, delegate work to subagents, and return to saved conversations.
 
----
+![term2 tracing, fixing, and testing a real bug from its own history](docs/demo/demo.gif)
 
-## Core Capabilities
+[Website & documentation](https://qduc.github.io/term2/) · [First run](https://qduc.github.io/term2/getting-started/first-run/) · [Changelog](CHANGELOG.md)
 
-- **Open Source and Local First**: MIT licensed. No mandatory subscriptions, telemetry lock-in, or proprietary walled gardens.
-- **Universal Provider Support**: Native integration with OpenAI (WebSocket Responses and Chat Completions), ChatGPT/Codex OAuth (browser PKCE), Grok Responses API (with encrypted reasoning and OAuth), OpenRouter, Anthropic, Google Gemini, and custom local endpoints (Ollama, llama.cpp, vLLM).
-- **Fine-Grained Sandboxing and Safety**: Sandboxed shell execution with configurable read/write boundaries (`standard`, `strict`) and interactive unified diff previews.
-- **Shell Auto-Approval Modes**: Configure shell approval as `off`, `advisory`, `auto`, or `always`; `always` disables the shell sandbox.
-- **Multi-Agent Orchestration**: Spawn specialized foreground or background subagents (`explorer`, `worker`, `mentor`, `librarian`) across tiered model profiles (`smart`, `balanced`, `cheap`, `chore`).
-- **Provider-Neutral Context Compaction**: Intelligently compacts long conversation histories (via `/compact` or automatic thresholds) while preserving cold-prefix architectural facts and hot-tail tool ledger integrity.
-- **Time-Travel Rewind and Forking**: Non-destructive conversation rewinding with discard previews (`/rewind`, `/undo`, `/retry`), session resumption (`/resume`, `--resume`), and session branching (`--fork`).
-- **Native Remote Development**: Execute commands and manage codebases on remote servers seamlessly over SSH with SSH agent authentication.
-- **Five Focused Operating Modes**: Standard, Plan, Lite, Mentor, and Orchestrator modes adapted to every phase of engineering.
+## What you can do
 
----
+- **Work with a codebase.** Trace a failure, implement a feature, review a change, and run tests. Use `/plan` for read-only exploration.
+- **Choose your models.** Connect a supported hosted provider or a local model endpoint. Switch models with `/model` and manage connections with `/providers`.
+- **Delegate tasks.** Use [subagents](https://qduc.github.io/term2/modes/subagents/) for research, implementation, or review, including background work.
+- **Keep working across sessions.** Resume a conversation after a break, fork it to try another approach, or work with a remote project over SSH.
 
-## Comparison: Why term2?
+## Get started
 
-| Feature | term2 | Claude Code | Warp |
-| :--- | :---: | :---: | :---: |
-| **License** | **MIT Open Source** | Proprietary | Proprietary / Closed Core |
-| **Cost** | **Direct API / Pay-per-use** | Subscription / API | Freemium / Paid Tiers |
-| **Model Ecosystem** | **Any** (OpenAI, Codex, Grok, OpenRouter, Anthropic, Google, Local) | Anthropic only | Selected Cloud Models |
-| **OAuth Browser Login** | **Grok and Codex PKCE** | No | No |
-| **Local LLMs (Ollama/llama.cpp)** | **Native** | No | Partial (BYOLLM) |
-| **Subagent Swarms** | **Configurable 4-Tier Roles** | Limited | No |
-| **Remote SSH Workspaces** | **Native SSH Agent** | Remote Control | SSH Client |
-| **Dual-Model Mentor Mode** | **Built-in** | No | No |
-| **Context Compaction** | **Local and Provider-Native** | Automatic (Closed) | No |
+You need Node.js **20 or newer** (Node 24 recommended), an interactive terminal, and access to a model provider.
 
----
-
-## Installation
-
-### Prerequisites
-- Node.js version 20.0.0 or higher
-- An API key or OAuth account for your preferred AI provider
-
-### Install via Package Manager
 ```bash
-# Install globally via npm
 npm install --global @qduc/term2
-
-# Or using pnpm
-pnpm add --global @qduc/term2
+term2 --version
 ```
 
-### Quick Setup
+On Linux, sandboxed shell commands also need `bubblewrap` (`bwrap`), `socat`, and `ripgrep` (`rg`). Install them with your OS package manager. Shell execution fails closed when the enabled sandbox is unavailable. See [installation](https://qduc.github.io/term2/getting-started/installation/) for platform details.
 
-#### 1. OAuth Browser Login (Grok or Codex)
-term2 supports browser-based PKCE OAuth logins for OpenAI Codex/ChatGPT and xAI Grok:
-```bash
-# Log in to xAI Grok
-term2 --grok-login
-
-# Log in to OpenAI Codex / ChatGPT
-term2 --codex-login
-```
+Choose one authentication path:
 
-> **Multi-Account Management**: Manage multiple OAuth accounts and switch between active profiles at runtime via the `/providers` menu.
-
-#### 2. Environment Variables (API Keys)
-Alternatively, export the API key for your chosen provider:
 ```bash
-# OpenAI
-export OPENAI_API_KEY="sk-..."
-
-# OpenRouter (Claude, Gemini, DeepSeek, etc.)
-export OPENROUTER_API_KEY="sk-or-v1-..."
-
-# Anthropic
-export ANTHROPIC_API_KEY="sk-ant-..."
-
-# Google Gemini
-export GOOGLE_GENERATIVE_AI_API_KEY="..."
-
-# Web Search (Tavily or Exa)
-export TAVILY_API_KEY="tvly-..."
-export EXA_API_KEY="..."
-```
-
----
-
-## Quick Start
-
-Launch term2 in your project directory:
-
-```bash
-# Start standard interactive session
-term2
-
-# Start in lightweight mode (fast system/terminal operations, no codebase ingestion)
-term2 --lite
-
-# Start with a specific model and provider
-term2 -m claude-sonnet-4.6 -p openrouter
-
-# Resume the last conversation session
-term2 --resume
+# ChatGPT / Codex: complete the browser login
+term2 --codex-login
 
-# Run a non-interactive prompt from the command line
-term2 "Explain the architecture of source/agent.ts"
+# Or OpenAI API: set your own API key in this shell
+export OPENAI_API_KEY="your-api-key"
 ```
 
----
-
-## Operating Modes
-
-term2 features five distinct operating modes tailored to different stages of development:
-
-| Mode | Trigger / CLI | Purpose and Behavior | Tools and Context |
-| :--- | :--- | :--- | :--- |
-| **Standard** | `term2` *(default)* | High-velocity codebase development. Automatically approves workspace file patches while gating destructive commands. | Full codebase context, all tools, patch auto-approval. |
-| **Plan** | `/plan` or `Shift+Tab` | Architecture design, research, and code exploration. Enforces read-only safety guarantees (no file writes or mutating commands). | Full codebase context, read-only tools. |
-| **Lite** | `term2 --lite` or `/lite` | General terminal tasks, DevOps, and SSH administration without codebase ingestion overhead. | Session context only, read-only system tools. |
-| **Mentor** | `/mentor` | Collaborative problem-solving. Pairs the primary model with a higher-tier strategic model (`smartModel`) for architectural review. | Full codebase context, all tools and `ask_mentor`. |
-| **Orchestrator** | `/orchestrator` | High-level task coordination. Encourages prompt-guided delegation across specialized subagents while retaining end-to-end turn ownership. | Full codebase context, direct tools and subagent delegation. |
-
-> **Mode Switching:** Press `Shift+Tab` to cycle the application's operating modes. To run one direct shell command in any operating mode, prefix the input with `!`; the command and output are added to the conversation, then the input returns to normal mode. Use slash commands (`/plan`, `/lite`, `/mentor`, `/orchestrator`) to switch modes explicitly.
-
----
-
-## Subagents Architecture
-
-term2 includes a modular subagent execution system designed to isolate heavy computational, research, or modification workloads from the main conversation transcript. Each subagent operates under a defined role profile with explicit permissions, execution budgets, and capability tiers.
-
-### Explorer Subagent
-- **Purpose**: Rapid codebase reconnaissance, symbol search, file discovery, and dependency mapping.
-- **Permissions**: Read-only (`canRead: true`, `canWrite: false`, `canSearchWeb: true`, `canRunShell: true`).
-- **Safety Boundary**: Restricted to non-destructive read operations and safe read-only shell commands (`rg`, `fd`, `git status`).
-- **Model Tier**: Defaults to `agent.cheapModel` (e.g. `gpt-5.4-mini`) to minimize latency and token consumption during wide search operations.
-
-### Worker Subagent
-- **Purpose**: Autonomous implementation of discrete tasks, file modifications, refactoring slices, and running automated test suites.
-- **Permissions**: Full read/write (`canRead: true`, `canWrite: true`, `canSearchWeb: false`, `canRunShell: true`).
-- **Safety Boundary**: Sandboxed to the workspace root directory. Supports pinning execution to a dedicated Git worktree (`worktree` parameter) to isolate dirty state from the parent checkout.
-- **Model Tier**: Defaults to `agent.balancedModel` (e.g. `gpt-5.3-codex` or `claude-sonnet-4.6`).
-
-### Mentor Subagent
-- **Purpose**: High-level strategic consultation, algorithmic design analysis, and independent second-opinion code review.
-- **Permissions**: Advisory only (`canRead: false`, `canWrite: false`, `canSearchWeb: false`, `canRunShell: false`).
-- **Safety Boundary**: Completely detached from the workspace filesystem and tool execution. Receives architectural queries via the `ask_mentor` tool and returns structured recommendations without polluting turn history.
-- **Model Tier**: Defaults to `agent.smartModel` (e.g. `gpt-5.5`).
+For other providers, follow [provider setup](https://qduc.github.io/term2/providers/). OpenRouter and Grok are built in; add direct Anthropic, Gemini, and local endpoints through `/providers`. An API key alone does not register a custom provider. Hosted providers receive the conversation context and tool results used for your request; their billing and account limits apply.
 
-### Librarian Subagent
-- **Purpose**: Long-term persistent memory management, documentation discovery, and indexing domain knowledge across sessions.
-- **Permissions**: Memory management (`canRead: false`, `canWrite: false`, `canSearchWeb: false`, `canRunShell: false`). Interacts with memory exclusively via specialized `memory_*` tools and `activate_skill`.
-- **Safety Boundary**: Operates strictly within persistent memory storage without general filesystem, shell, or web access.
-- **Model Tier**: Defaults to `agent.cheapModel`.
-
-### Subagent Capability Tiers
-Subagent roles are mapped to four configurable model tiers in `settings.json`:
-- `agent.smartModel` / `agent.smartProvider` / `agent.smartReasoningEffort`: For deep reasoning and strategic advisory (used by Mentor).
-- `agent.balancedModel` / `agent.balancedProvider` / `agent.balancedReasoningEffort`: For implementation, editing, and execution (used by Worker).
-- `agent.cheapModel` / `agent.cheapProvider` / `agent.cheapReasoningEffort`: For high-throughput search and discovery (used by Explorer and Librarian).
-- `agent.choreModel` / `agent.choreProvider`: For narrow utility tasks such as AST patch self-healing and auto-approval evaluation.
-
-### Background Subagent Management and Lifecycle
-- **Concurrent Execution**: Subagents can run synchronously in the foreground or asynchronously in the background.
-- **Task Manager (`Ctrl+G`)**: Press `Ctrl+G` to open the Background Task Manager to inspect live running subagents, view stream output, pause runs, or terminate hung jobs.
-- **Resource Budgets**: Each subagent run is bounded by an execution turn budget (`maxTurns`) to prevent infinite recursive loops.
-
-### Subagent Check-In and Steering
-- **Periodic Check-In**: The `BackgroundCheckInScheduler` periodically wakes the launching agent with status summaries of background work without interrupting active user turns.
-- **Mid-Turn Steering**: Deliver real-time steering messages to running subagent tasks at clean request boundaries.
-- **Orchestrator Inquiries**: Subagents can query the parent orchestrator for clarifying decisions using the `ask_orchestrator` tool when encountering ambiguity.
-
----
-
-## Safety, Sandboxing, and Auto-Approval
-
-### Sandboxed Execution
-Shell commands execute inside an isolated execution boundary:
-- `standard` policy: Restricts file writes to the workspace and temporary directories while blocking reads of sensitive credential paths (`~/.ssh`, `~/.aws`, `~/.docker`, `~/.netrc`, `~/.kube`, etc.).
-- `strict` policy: Locks down file reading to the workspace and safe toolchains, blocking access to the user home directory and system roots (`/etc`, `/var`, `/root`).
-- Toggle sandbox enforcement anytime using `/sandbox`, or configure policies via `sandbox.readPolicy` in `/settings`.
-
-### Shell Auto-Approval Modes (`/auto-approve`)
-- `off` (Default): Shell commands are not auto-approved.
-- `advisory`: Commands require confirmation, but include real-time LLM-generated safety and consequence explanations.
-- `auto`: Safe, read-only, and idempotent workspace operations execute automatically; potentially destructive commands (e.g. `rm -rf`, `git reset --hard`, `git push --force`) are strictly blocked for user confirmation.
-- `always`: Unattended shell execution mode. Disables sandbox boundaries; use only in a trusted environment.
-
-The `--auto-approve` CLI flag is separate: it applies only to non-interactive prompts and allows tool execution without interactive confirmation.
-
----
-
-## Context Management and Compaction
-
-### Local and Server-Side Compaction
-- **Provider-Neutral Compactor**: When conversation tokens exceed configured ratios or thresholds, term2 performs request-boundary context compaction.
-- **Cold-Prefix Preservation**: Architectural decisions, environment constraints, and critical requirements are summarized and retained verbatim.
-- **Hot-Tail Invariants**: Recent conversation turns and active tool ledger call/result pairs remain completely intact, preventing broken references.
-- **Manual Trigger**: Run `/compact` at any point to compress transcript history immediately.
-- **OpenAI Native Compaction**: Leverages server-side context management when supported by the provider endpoint.
-
----
-
-## Conversation Resumption, Rewind, and Queuing
-
-### Time-Travel Rewind (`/rewind`, `/undo`, `/retry`)
-- **Inspection Picker**: Running `/rewind` opens an interactive menu detailing exactly which turns, replies, and modified files will be affected before committing.
-- **Disposition Modes**:
-  - `edit` (Default / `/undo`): Discards subsequent turns and places the selected user prompt back into the input editor for revision.
-  - `resend` (`/retry`): Discards subsequent turns and immediately resends the selected prompt to the model.
-- **Tool Retry (`/retry-tool`)**: Retries only the most recent failed or timed-out tool call without losing conversation context.
-
-### Session Persistence and Resumption
-- **Persistence Engine**: Sessions are auto-saved using append-only event streams.
-- **Resuming Sessions**:
-  ```bash
-  term2 --resume               # Resume the most recent session
-  term2 --resume <session-id>  # Resume a specific session by UUID
-  term2 --resume ls            # List recent saved sessions with metadata
-  ```
-- **Session Branching (`--fork`)**:
-  ```bash
-  term2 --resume <session-id> --fork
-  ```
-
-### Mid-Turn Injection and Prompt Queuing
-- **Input Queuing**: Press `Alt+Enter` or `Esc+Enter` to queue additional user instructions while a turn is actively generating.
-- **Queue Editing**: Edit or delete pending queued prompts before they are admitted at the next request boundary.
-
----
-
-## Remote Development over SSH
-
-term2 connects directly to remote hosts over SSH, executing commands and modifying remote files within a remote working directory:
+Open a project and select a model from the provider you authenticated:
 
 ```bash
-# Connect to remote server with working directory
-term2 --ssh user@hostname --remote-dir /path/to/project
-
-# Custom SSH port
-term2 --ssh user@hostname --remote-dir /path/to/project --ssh-port 2222
-
-# Lightweight remote administration without codebase indexing
-term2 --ssh user@hostname --lite
+cd /path/to/your/project
+term2 --model
 ```
 
-**Requirements and Behavior**:
-- Requires an SSH agent running locally (`ssh-agent`) with loaded keys (`ssh-add`), or configured identity keys in `~/.ssh/config`.
-- `--remote-dir` is required for full codebase sessions. In `--lite` mode, omitting `--remote-dir` defaults to the remote user's home directory.
-- Background command execution is restricted to local workspaces.
+Ask it to explain the project, fix a bug, or implement a feature. Standard mode can apply valid local workspace edits without a prompt; shell commands require approval under the default interactive settings. Review the resulting diff before committing. Enter `/plan` for read-only exploration, then `/plan` again to return to Standard mode.
 
----
+This README tracks `main`; npm releases may lag behind it. Check `term2 --version` and the [changelog](CHANGELOG.md) when following documentation for a newer feature.
 
-## Non-Interactive and Scripting Mode
+## Useful commands
 
-Execute one-off queries and automation tasks directly from shell scripts:
+| To…                         | Use…                                                          |
+| --------------------------- | ------------------------------------------------------------- |
+| Switch models               | `/model` or `Ctrl+O`                                          |
+| Manage provider connections | `/providers`                                                  |
+| Resume a conversation       | `term2 --resume`; `term2 --resume ls` lists saved sessions    |
+| Try another approach        | `term2 --resume <session-id> --fork`                          |
+| Work over SSH               | `term2 --ssh user@host --remote-dir /path/to/project`         |
+| Ask a one-off question      | `term2 "Explain the difference between a merge and a rebase"` |
 
-```bash
-# Standard non-interactive query (output prints to stdout)
-term2 "Summarize recent commits in this repository"
+A command-line prompt runs non-interactively, with tool execution disabled by default. `--auto-approve` enables tools without interactive confirmation; use it only when you intend unattended execution. See [scripting](https://qduc.github.io/term2/using-term2/non-interactive/) and [SSH setup](https://qduc.github.io/term2/remote/ssh/).
 
-# Automated task execution with tool approvals enabled
-term2 --auto-approve "Fix syntax errors in source/cli.tsx and run typecheck"
+## Approvals and sandbox
 
-# Capture output in shell pipeline
-TODO_LIST=$(term2 "List all TODO markers in source/")
-```
+Interactive sessions default to shell auto-approval `off` and an enabled shell sandbox. You can change these settings, so the active configuration matters. The sandbox constrains shell commands; it does not cover providers, external MCP servers, or every tool.
 
----
-
-## Reference Guide
-
-### Slash Commands
-
-| Command | Arguments | Description |
-| :--- | :--- | :--- |
-| `/model` | `[model-name]` | Open model selection menu or switch model directly (`Ctrl+O`). |
-| `/effort` | `[effort-level]` | Configure reasoning effort (`none`, `minimal`, `low`, `medium`, `high`, `xhigh`) (`Ctrl+T`). |
-| `/plan` | — | Toggle read-only Plan Mode (`Shift+Tab`). |
-| `/lite` | — | Toggle lightweight terminal mode. |
-| `/mentor` | — | Toggle collaborative Mentor Mode. |
-| `/orchestrator` | — | Toggle prompt-guided Orchestrator Mode. |
-| `/auto-approve` | `[off\|advisory\|auto\|always]` | Set shell command auto-approval level. |
-| `/sandbox` | — | Toggle shell sandbox isolation on or off. |
-| `/compact` | — | Manually trigger context compaction. |
-| `/providers` | — | Open interactive provider manager (list, add, edit, remove, switch accounts). |
-| `/mcp` | — | Show configured MCP servers and their connection state. |
-| `/skills` | `[skill-name]` | Activate a skill for the next request. |
-| `/rewind` | `[last\|<turn>] [edit\|resend]` | Rewind conversation history with interactive discard inspection. |
-| `/undo` | — | Alias for `/rewind edit` (places target turn back in input box). |
-| `/retry` | — | Alias for `/rewind resend` (immediately resends the last user turn). |
-| `/retry-tool` | — | Re-execute the last failed or timed-out tool call. |
-| `/copy` | `[N]` | Copy the latest assistant response (or $N$-th prior response) to clipboard. |
-| `/usage` | — | Display exact token counts, cost breakdown, and provider rate-limit metrics. |
-| `/resume` | `[conversation-id]` | Resume a saved conversation (browse with `/resume` or specify an ID). |
-| `/handoff` | — | Export current conversation context for handoff to another session or model. |
-| `/settings` | `[key] [value]` | View or modify runtime configuration settings. |
-| `/clear` | — | Clear current conversation history and start a fresh turn. |
-| `/quit` | — | Exit term2 session. |
-
-### Keyboard Shortcuts
-
-| Shortcut | Context | Action |
-| :--- | :--- | :--- |
-| `Ctrl + O` | Input Box | Open interactive model selection menu. |
-| `Ctrl + T` | Input Box | Open reasoning effort selection menu. |
-| `Ctrl + G` | Global | Open Background Task Manager (inspect/pause/stop tasks). |
-| `Shift + Tab` | Global | Cycle Operating Modes. |
-| `!` prefix | Input Box | Execute the following command directly in the shell; the red `!` prompt remains active until the command completes. |
-| `Alt + Enter` / `Esc + Enter` | Input Box | Queue input while a turn is executing. |
-| `Ctrl + R` | Model Menu | Refresh provider model catalog from API. |
-| `Ctrl + D` | Settings Menu | Reset highlighted setting to its default value. |
-| `Escape` | Composer | First press shows the Escape hint; a second press clears typed text. With an empty composer during a turn, double Escape interrupts the turn. |
-| `y` / `n` | Docker host-control approval | Approve once / reject. Other tool approvals use the numbered choices shown in the prompt. |
-| `Ctrl + C` | Global | Force immediate graceful exit. |
-
----
-
-## Configuration
-
-Settings are persisted across sessions in `settings.json`:
-- **macOS**: `~/Library/Logs/term2-nodejs/settings.json`
-- **Linux**: `~/.local/state/term2-nodejs/settings.json` *(or `$XDG_STATE_HOME/term2-nodejs/settings.json`)*
-- **Windows**: `%LOCALAPPDATA%\term2-nodejs\Log\settings.json`
-
-### Configuration Example
-
-```json
-{
-  "agent": {
-    "provider": "openai",
-    "model": "gpt-5.4",
-    "reasoningEffort": "medium",
-    "smartModel": "gpt-5.5",
-    "smartProvider": "openai",
-    "balancedModel": "claude-sonnet-4.6",
-    "balancedProvider": "openrouter",
-    "cheapModel": "gpt-5.4-mini",
-    "choreModel": "gpt-5.4-mini",
-    "temperature": 0.7,
-    "maxTurns": 100
-  },
-  "shell": {
-    "autoApproveMode": "auto",
-    "timeout": 120000,
-    "maxParallelToolCalls": 3
-  },
-  "sandbox": {
-    "readPolicy": "standard"
-  },
-  "webSearch": {
-    "provider": "tavily"
-  },
-  "providers": [
-    {
-      "name": "Local Ollama",
-      "type": "openai-compatible",
-      "baseUrl": "http://127.0.0.1:11434/v1"
-    },
-    {
-      "name": "Local llama.cpp",
-      "type": "llama.cpp",
-      "baseUrl": "http://127.0.0.1:8080/v1"
-    }
-  ]
-}
-```
-
----
+See [approvals](https://qduc.github.io/term2/safety/approvals/) and [sandbox boundaries](https://qduc.github.io/term2/safety/sandbox/) for the policies and their limits.
 
-## Development and Testing
+## Develop from source
 
-To set up the development environment from source:
+Use Node 24 and pnpm 11:
 
 ```bash
-# 1. Clone repository
 git clone https://github.com/qduc/term2.git
 cd term2
-
-# 2. Install dependencies (requires pnpm >= 11)
-pnpm install
-
-# 3. Start TypeScript compiler in watch mode
-pnpm dev
-
-# 4. Run test suites
-pnpm test                    # Run Vitest unit tests
-pnpm test:lane               # Run deterministic fast-lane tests
-pnpm test:provider-black-box # Run provider black-box wire tests
-pnpm typecheck               # Run strict type checking
-pnpm lint                    # Run ESLint and Prettier checks
-
-# 5. Build for distribution
+pnpm install --frozen-lockfile
 pnpm build
+pnpm start --help
 ```
 
----
+Read [AGENTS.md](AGENTS.md) before changing code. Use focused tests for the affected area; integration and provider tests have separate commands in `package.json`.
 
-## License
+The documentation site lives in [`website/`](website/):
 
-This project is licensed under the [MIT License](LICENSE).
+```bash
+pnpm --dir website install --frozen-lockfile
+pnpm --dir website build
+pnpm --dir website preview
+```
 
----
+The separate [`web-client/`](web-client/) app connects to the private `term2 serve` gateway. See [gateway setup](https://qduc.github.io/term2/remote/web-gateway/).
 
-## Acknowledgments
+## License
 
-term2 is built on open-source libraries:
-- [Ink](https://github.com/vadimdemedes/ink) — React for interactive command-line apps
-- [Vercel AI SDK](https://github.com/vercel/ai) — Multi-provider AI streaming primitives
-- [OpenAI Node SDK](https://github.com/openai/openai-node) — OpenAI API transport
-- [ssh2](https://github.com/mscdex/ssh2) — SSH2 client for remote server execution
-- [Zod](https://github.com/colinhacks/zod) — TypeScript-first schema validation
+[MIT](LICENSE).

package/dist/agent.d.ts

@@ -4,7 +4,7 @@ import type { BackgroundShellRegistry } from './services/shell/background-shell-
 import type { BackgroundShellOutputBundle } from './services/shell/background-shell-watches.js';
 import { type ForegroundRunSubagentParams } from './tools/agent/run-subagent.js';
 import { type ConfigureTaskCheckInParams, type ConfigureTaskCheckInResult } from './tools/agent/configure-task-check-in.js';
-import type { ToolRegistry } from './tools/types.js';
+import type { RequestSnapshot, ToolRegistry } from './tools/types.js';
 import type { NestedSubagentResult, SubagentResult, SubagentRunHandle, SubagentRunStatus } from './services/subagents/types.js';
 import type { ResolvedSubagentLaunch } from './lib/subagent-bridge.js';
 import type { CancelRunParams, GetSubagentResultParams, GetSubagentStatusParams, RunSubagentAsyncParams, SendMessageParams, SendMessageAcknowledgement, CancelRunAcknowledgement } from './tools/agent/run-subagent-async.js';
@@ -25,6 +25,7 @@ export interface AgentDefinition {
     name: string;
     instructions: string;
     resolveInstructionsForRequest?: () => string;
+    resolveRequestSnapshot?: () => RequestSnapshot;
     tools: ToolRegistry;
     model: string;
     /** Enable pinned global memory context only for profiles exposing root memory context. */

package/dist/agent.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../source/agent.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,4CAA4C,CAAC;AAIrF,OAAO,EAGL,KAAK,8BAA8B,EACpC,MAAM,yBAAyB,CAAC;AACjC,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,+CAA+C,CAAC;AAC7F,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,8CAA8C,CAAC;AAGhG,OAAO,EAAmC,KAAK,2BAA2B,EAAE,MAAM,+BAA+B,CAAC;AAclH,OAAO,EAEL,KAAK,0BAA0B,EAC/B,KAAK,0BAA0B,EAChC,MAAM,0CAA0C,CAAC;AAGlD,OAAO,KAAK,EAAqB,YAAY,EAAE,MAAM,kBAAkB,CAAC;AACxE,OAAO,KAAK,EACV,oBAAoB,EACpB,cAAc,EACd,iBAAiB,EACjB,iBAAiB,EAClB,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,KAAK,EACV,eAAe,EACf,uBAAuB,EACvB,uBAAuB,EACvB,sBAAsB,EACtB,iBAAiB,EACjB,0BAA0B,EAC1B,wBAAwB,EACzB,MAAM,qCAAqC,CAAC;AAI7C,OAAO,KAAK,EAAE,gBAAgB,EAAE,eAAe,EAAE,MAAM,kCAAkC,CAAC;AAC1F,OAAO,EAAE,gBAAgB,EAAE,MAAM,iCAAiC,CAAC;AAInE,OAAO,EAAE,aAAa,EAAE,MAAM,qCAAqC,CAAC;AAMpE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,2CAA2C,CAAC;AAI9E,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,4CAA4C,CAAC;AAIpF,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,uCAAuC,CAAC;AAEhF,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,4CAA4C,CAAC;AACjF,OAAO,KAAK,EAAE,sBAAsB,EAAE,6BAA6B,EAAE,MAAM,iCAAiC,CAAC;AAC7G,OAAO,EAAE,0BAA0B,EAAE,MAAM,sDAAsD,CAAC;AAElG,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,mCAAmC,CAAC;AACvE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,+CAA+C,CAAC;AAKjF,wBAAgB,UAAU,CACxB,eAAe,EAAE,gBAAgB,EACjC,gBAAgB,CAAC,EAAE,gBAAgB,EACnC,IAAI,UAAQ,GACX,MAAM,CAsCR;AAED,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,CA8BzD;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,YAAY,EAAE,MAAM,CAAC;IACrB,6BAA6B,CAAC,EAAE,MAAM,MAAM,CAAC;IAC7C,KAAK,EAAE,YAAY,CAAC;IACpB,KAAK,EAAE,MAAM,CAAC;IACd,0FAA0F;IAC1F,oBAAoB,EAAE,OAAO,CAAC;CAC/B;AAED,KAAK,kBAAkB,GAAG,IAAI,CAAC,cAAc,EAAE,WAAW,CAAC,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAAC;AA8C5F;;GAEG;AACH,eAAO,MAAM,kBAAkB,SACvB;IACJ,eAAe,EAAE,gBAAgB,CAAC;IAClC,cAAc,EAAE,eAAe,CAAC;IAChC,OAAO,CAAC,EAAE,MAAM,WAAW,GAAG,SAAS,CAAC;IACxC,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,sBAAsB,CAAC,EAAE,0BAA0B,CAAC;IACpD,SAAS,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IAClD,WAAW,CAAC,EAAE,CACZ,MAAM,EAAE,2BAA2B,EACnC,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,kBAAkB,CAAC,CAAC;IACjC,gBAAgB,CAAC,EAAE,CACjB,MAAM,EAAE,sBAAsB,EAC9B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,iBAAiB,CAAC,CAAC;IAChC,iBAAiB,CAAC,EAAE,CAClB,MAAM,EAAE,uBAAuB,EAC/B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,cAAc,CAAC,CAAC;IAC7B,iBAAiB,CAAC,EAAE,CAClB,MAAM,EAAE,uBAAuB,EAC/B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,iBAAiB,GAAG,iBAAiB,EAAE,CAAC;IAC7C,mBAAmB,CAAC,EAAE,CAAC,MAAM,EAAE,iBAAiB,KAAK,0BAA0B,CAAC;IAChF,iBAAiB,CAAC,EAAE,CAAC,MAAM,EAAE,eAAe,KAAK,wBAAwB,CAAC;IAC1E,iHAAiH;IACjH,mBAAmB,CAAC,EAAE,CACpB,MAAM,EAAE,sBAAsB,EAC9B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,oBAAoB,CAAC,CAAC;IACnC,wBAAwB,CAAC,EAAE,CAAC,MAAM,EAAE,sBAAsB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;IAC1F,gBAAgB,CAAC,EAAE,CAAC,MAAM,CAAC,EAAE,MAAM,KAAK,MAAM,GAAG,SAAS,CAAC;IAC3D,4FAA4F;IAC5F,WAAW,CAAC,EAAE;QAAE,UAAU,EAAE,CAAC,IAAI,EAAE,WAAW,KAAK,IAAI,CAAC;QAAC,gBAAgB,EAAE,MAAM,OAAO,CAAA;KAAE,CAAC;IAC3F,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,YAAY,CAAC,EAAE,IAAI,CAAC,YAAY,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAClD,qBAAqB,CAAC,EAAE,OAAO,CAAC;IAChC,aAAa,CAAC,EAAE,kBAAkB,CAAC;IACnC,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,gBAAgB,CAAC,EAAE,OAAO,CAAC;IAC3B,iFAAiF;IACjF,uBAAuB,CAAC,EAAE,uBAAuB,CAAC,8BAA8B,CAAC,CAAC;IAClF,wEAAwE;IACxE,qBAAqB,CAAC,EAAE,2BAA2B,CAAC;IACpD,kBAAkB,CAAC,EAAE,kBAAkB,CAAC;IACxC,qFAAqF;IACrF,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,2FAA2F;IAC3F,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,oFAAoF;IACpF,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,sBAAsB,CAAC,EAAE,CAAC,OAAO,EAAE,sBAAsB,KAAK,6BAA6B,CAAC;IAC5F,oBAAoB,CAAC,EAAE,CAAC,MAAM,EAAE,0BAA0B,KAAK,0BAA0B,CAAC;IAC1F,oBAAoB,CAAC,EAAE,CACrB,MAAM,EAAE;QAAE,IAAI,EAAE,OAAO,GAAG,UAAU,CAAC;QAAC,EAAE,EAAE,MAAM,CAAA;KAAE,EAClD,OAAO,EAAE;QAAE,OAAO,CAAC,EAAE,OAAO,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAA;KAAE,KAChD,IAAI,CAAC;IACV,gFAAgF;IAChF,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,2BAA2B,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,MAAM,KAAK,MAAM,CAAC;IAC7D,aAAa,CAAC,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC;IACtC,aAAa,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;IACjD,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;CACvC,UACO,MAAM,KACb,eAmeF,CAAC"}
+{"version":3,"file":"agent.d.ts","sourceRoot":"","sources":["../source/agent.ts"],"names":[],"mappings":"AAEA,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,4CAA4C,CAAC;AAIrF,OAAO,EAGL,KAAK,8BAA8B,EACpC,MAAM,yBAAyB,CAAC;AACjC,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,+CAA+C,CAAC;AAC7F,OAAO,KAAK,EAAE,2BAA2B,EAAE,MAAM,8CAA8C,CAAC;AAGhG,OAAO,EAAmC,KAAK,2BAA2B,EAAE,MAAM,+BAA+B,CAAC;AAclH,OAAO,EAEL,KAAK,0BAA0B,EAC/B,KAAK,0BAA0B,EAChC,MAAM,0CAA0C,CAAC;AAGlD,OAAO,KAAK,EAAqB,eAAe,EAAE,YAAY,EAAE,MAAM,kBAAkB,CAAC;AACzF,OAAO,KAAK,EACV,oBAAoB,EACpB,cAAc,EACd,iBAAiB,EACjB,iBAAiB,EAClB,MAAM,+BAA+B,CAAC;AACvC,OAAO,KAAK,EAAE,sBAAsB,EAAE,MAAM,0BAA0B,CAAC;AACvE,OAAO,KAAK,EACV,eAAe,EACf,uBAAuB,EACvB,uBAAuB,EACvB,sBAAsB,EACtB,iBAAiB,EACjB,0BAA0B,EAC1B,wBAAwB,EACzB,MAAM,qCAAqC,CAAC;AAI7C,OAAO,KAAK,EAAE,gBAAgB,EAAE,eAAe,EAAE,MAAM,kCAAkC,CAAC;AAC1F,OAAO,EAAE,gBAAgB,EAAE,MAAM,iCAAiC,CAAC;AAInE,OAAO,EAAE,aAAa,EAAE,MAAM,qCAAqC,CAAC;AAOpE,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,2CAA2C,CAAC;AAI9E,OAAO,KAAK,EAAE,iBAAiB,EAAE,MAAM,4CAA4C,CAAC;AAIpF,OAAO,KAAK,EAAE,kBAAkB,EAAE,MAAM,uCAAuC,CAAC;AAEhF,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,4CAA4C,CAAC;AACjF,OAAO,KAAK,EAAE,sBAAsB,EAAE,6BAA6B,EAAE,MAAM,iCAAiC,CAAC;AAC7G,OAAO,EAAE,0BAA0B,EAAE,MAAM,sDAAsD,CAAC;AAElG,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,mCAAmC,CAAC;AACvE,OAAO,KAAK,EAAE,WAAW,EAAE,MAAM,+CAA+C,CAAC;AAKjF,wBAAgB,UAAU,CACxB,eAAe,EAAE,gBAAgB,EACjC,gBAAgB,CAAC,EAAE,gBAAgB,EACnC,IAAI,UAAQ,GACX,MAAM,CAsCR;AAED,wBAAgB,qBAAqB,CAAC,GAAG,EAAE,MAAM,GAAG,MAAM,CA8BzD;AAED,MAAM,WAAW,eAAe;IAC9B,IAAI,EAAE,MAAM,CAAC;IACb,YAAY,EAAE,MAAM,CAAC;IACrB,6BAA6B,CAAC,EAAE,MAAM,MAAM,CAAC;IAC7C,sBAAsB,CAAC,EAAE,MAAM,eAAe,CAAC;IAC/C,KAAK,EAAE,YAAY,CAAC;IACpB,KAAK,EAAE,MAAM,CAAC;IACd,0FAA0F;IAC1F,oBAAoB,EAAE,OAAO,CAAC;CAC/B;AAED,KAAK,kBAAkB,GAAG,IAAI,CAAC,cAAc,EAAE,WAAW,CAAC,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAAC;AA8C5F;;GAEG;AACH,eAAO,MAAM,kBAAkB,SACvB;IACJ,eAAe,EAAE,gBAAgB,CAAC;IAClC,cAAc,EAAE,eAAe,CAAC;IAChC,OAAO,CAAC,EAAE,MAAM,WAAW,GAAG,SAAS,CAAC;IACxC,gBAAgB,CAAC,EAAE,gBAAgB,CAAC;IACpC,sBAAsB,CAAC,EAAE,0BAA0B,CAAC;IACpD,SAAS,CAAC,EAAE,CAAC,QAAQ,EAAE,MAAM,KAAK,OAAO,CAAC,MAAM,CAAC,CAAC;IAClD,WAAW,CAAC,EAAE,CACZ,MAAM,EAAE,2BAA2B,EACnC,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,kBAAkB,CAAC,CAAC;IACjC,gBAAgB,CAAC,EAAE,CACjB,MAAM,EAAE,sBAAsB,EAC9B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,iBAAiB,CAAC,CAAC;IAChC,iBAAiB,CAAC,EAAE,CAClB,MAAM,EAAE,uBAAuB,EAC/B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,cAAc,CAAC,CAAC;IAC7B,iBAAiB,CAAC,EAAE,CAClB,MAAM,EAAE,uBAAuB,EAC/B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,iBAAiB,GAAG,iBAAiB,EAAE,CAAC;IAC7C,mBAAmB,CAAC,EAAE,CAAC,MAAM,EAAE,iBAAiB,KAAK,0BAA0B,CAAC;IAChF,iBAAiB,CAAC,EAAE,CAAC,MAAM,EAAE,eAAe,KAAK,wBAAwB,CAAC;IAC1E,iHAAiH;IACjH,mBAAmB,CAAC,EAAE,CACpB,MAAM,EAAE,sBAAsB,EAC9B,OAAO,CAAC,EAAE,OAAO,EACjB,OAAO,CAAC,EAAE,OAAO,KACd,OAAO,CAAC,oBAAoB,CAAC,CAAC;IACnC,wBAAwB,CAAC,EAAE,CAAC,MAAM,EAAE,sBAAsB,KAAK,OAAO,CAAC,iBAAiB,CAAC,CAAC;IAC1F,gBAAgB,CAAC,EAAE,CAAC,MAAM,CAAC,EAAE,MAAM,KAAK,MAAM,GAAG,SAAS,CAAC;IAC3D,4FAA4F;IAC5F,WAAW,CAAC,EAAE;QAAE,UAAU,EAAE,CAAC,IAAI,EAAE,WAAW,KAAK,IAAI,CAAC;QAAC,gBAAgB,EAAE,MAAM,OAAO,CAAA;KAAE,CAAC;IAC3F,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,YAAY,CAAC,EAAE,IAAI,CAAC,YAAY,EAAE,OAAO,CAAC,GAAG,IAAI,CAAC;IAClD,qBAAqB,CAAC,EAAE,OAAO,CAAC;IAChC,aAAa,CAAC,EAAE,kBAAkB,CAAC;IACnC,QAAQ,CAAC,EAAE,OAAO,CAAC;IACnB,gBAAgB,CAAC,EAAE,OAAO,CAAC;IAC3B,iFAAiF;IACjF,uBAAuB,CAAC,EAAE,uBAAuB,CAAC,8BAA8B,CAAC,CAAC;IAClF,wEAAwE;IACxE,qBAAqB,CAAC,EAAE,2BAA2B,CAAC;IACpD,kBAAkB,CAAC,EAAE,kBAAkB,CAAC;IACxC,qFAAqF;IACrF,oBAAoB,CAAC,EAAE,OAAO,CAAC;IAC/B,2FAA2F;IAC3F,YAAY,CAAC,EAAE,OAAO,CAAC;IACvB,oFAAoF;IACpF,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,sBAAsB,CAAC,EAAE,CAAC,OAAO,EAAE,sBAAsB,KAAK,6BAA6B,CAAC;IAC5F,oBAAoB,CAAC,EAAE,CAAC,MAAM,EAAE,0BAA0B,KAAK,0BAA0B,CAAC;IAC1F,oBAAoB,CAAC,EAAE,CACrB,MAAM,EAAE;QAAE,IAAI,EAAE,OAAO,GAAG,UAAU,CAAC;QAAC,EAAE,EAAE,MAAM,CAAA;KAAE,EAClD,OAAO,EAAE;QAAE,OAAO,CAAC,EAAE,OAAO,CAAC;QAAC,UAAU,CAAC,EAAE,MAAM,CAAA;KAAE,KAChD,IAAI,CAAC;IACV,gFAAgF;IAChF,aAAa,CAAC,EAAE,aAAa,CAAC;IAC9B,2BAA2B,CAAC,EAAE,CAAC,IAAI,EAAE,MAAM,MAAM,KAAK,MAAM,CAAC;IAC7D,aAAa,CAAC,EAAE,aAAa,CAAC,MAAM,CAAC,CAAC;IACtC,aAAa,CAAC,EAAE,QAAQ,CAAC,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC,CAAC;IACjD,iBAAiB,CAAC,EAAE,iBAAiB,CAAC;CACvC,UACO,MAAM,KACb,eA4eF,CAAC"}

package/dist/agent.js

@@ -23,6 +23,7 @@ import { ProfileResolutionError, resolveActiveProfile } from './services/profile
 import { shouldPreferPatchEditingModel } from './lib/tool-selection-policy.js';
 import { createActivateSkillToolDefinition } from './tools/agent/activate-skill.js';
 import { createProposeGoalToolDefinition } from './tools/agent/propose-goal.js';
+import { createGoalCheckToolDefinition } from './tools/agent/goal-check.js';
 import { createRunAgentWorkflowToolDefinition } from './tools/run-agent-workflow.js';
 import { createWorktreeToolDefinitions } from './tools/system/worktree.js';
 import { createRunCodeToolDefinition } from './tools/system/run-code/index.js';
@@ -330,6 +331,10 @@ export const getAgentDefinition = (deps, model) => {
             hasPriorProposal: proposeGoal.hasPriorProposal,
         }));
     }
+    // The stop self-check is root-only (only the root receives `getGoal`) and is
+    // available in every mode: the stop guard requires it whenever a goal is active.
+    if (deps.getGoal)
+        tools.push(createGoalCheckToolDefinition({ getGoal: deps.getGoal }));
     if (hasCapability('user-interaction') && getAskUserAnswer && allowAskUser) {
         const askUserTool = createAskUserToolDefinition(getAskUserAnswer);
         if (askUserTool.name !== TOOL_NAME_ASK_USER) {
@@ -510,7 +515,12 @@ export const getAgentDefinition = (deps, model) => {
         memoryContextEnabled: hasCapability('memory') && memoryContextEnabled && memoryCapability.access !== 'none',
         instructions,
         ...(deps.getGoal
-            ? { resolveInstructionsForRequest: () => `${instructions}${renderDurableGoalContext(deps.getGoal?.())}` }
+            ? {
+                resolveInstructionsForRequest: () => `${instructions}${renderDurableGoalContext(deps.getGoal?.())}`,
+                // Read at the same request boundary as the rendered goal above, so
+                // goal_check records the goal the model judged, not a later one.
+                resolveRequestSnapshot: () => ({ goal: deps.getGoal?.() }),
+            }
             : {}),
         tools,
         model: resolvedModel,