microsoft/intelligent-terminal v0.2.23

Intelligent Terminal v0.2.2395

on GitHub

This servicing release focuses on sharpening the parts of Intelligent Terminal you use most. Autofix now works toward what you were actually trying to do instead of just focusing on the commands that failed, you can bring your own local models with API keys, and the Agent Pane gets a wave of polish for navigation, activity feedback, and switching agents or models in place. It also lands a batch of performance and reliability fixes to make things quicker and smoother.

Highlights

1. Bring your own local model securely. We've added support for API keys with local models. See the Settings>Agents to hook this up. API keys are stored in Windows Credential Manager, so authenticated cloud endpoints work right alongside keyless local providers like Ollama. #561

1. Autofix works toward your goal, not just the failed command. It can clarify what you meant, diagnose multi-step failures, ask for permissions, and route the proposed fix back to the pane where the failure happened. #625

1. Follow every agent action clearly. Reported commands, working directories, output previews, exit codes, file changes, and queued permission requests all show up in order alongside responses and plans. #601, #611, #612

1. Switch agents and models without rebuilding your pane. Changing the agent or model setting updates active tabs in place, so they keep their terminal, focus, and Agent Pane. #655

1. Reliable across long-running sessions. Native hooks, clearer setup diagnostics, correct session ownership, and deterministic process cleanup keep agent integrations stable over time. #571, #635, #636, #637, #646, #647, #649

Features and Improvements

Bring your own model and provider

1. Added secure API-key authentication for OpenAI-compatible providers. Authenticated cloud endpoints work alongside keyless local providers such as Ollama, with credentials stored in Windows Credential Manager instead of settings files. #561
2. Showed the active model in the Agent Pane title bar. Confirm which model is handling the current tab, including BYOK models, without opening the model picker. #634

Autofix

1. Made Autofix work toward your intended outcome instead of rewriting only the failed command. It can clarify your goal, diagnose multi-step failures, request permissions, and return a proposed fix to the pane where the failure occurred. #625
2. Made repeated Autofix attempts more efficient. The full planning instructions are reused across turns instead of being resent with every request. #622

Agent Pane experience

1. Improved the Agent Pane layout at shorter heights. Recommendations, pending actions, and the input box stay visible and usable in compact panes. #580
2. Made slash commands searchable by any part of their name. Case-insensitive matching, highlighted results, and prefix matches ranked first. #615
3. Made completed turns easier to navigate with the keyboard and mouse. Keyboard-selected history stays in view, and you can click prompt rows to expand or collapse them. #618, #624
4. Made agent tool activity easier to follow. Reported commands, working directories, output previews, exit codes, file changes, and queued permission requests appear while preserving the order of responses, plans, and tool calls. #601, #611, #612
5. Added structured questions to the Agent Pane. Agents can pause for a choice or freeform clarification and continue with the answer in the correct session. #606
6. Reduced duplicate Agent Pane notifications and hook overhead. Questions and permission requests appear once while turn-level activity stays visible. #631

Agents and connectivity

1. Applied agent and model setting changes without rebuilding active panes. Tabs across windows switch providers or models in place without losing their terminal, stealing focus, or unnecessarily replacing the Agent Pane. #655
2. Used detected Claude and Codex installations for ACP connections. Custom install locations work without downloading duplicate native packages, while missing CLIs fail early with the existing setup guidance. #644
3. Supported larger agent events in wtcli publish. Protocol payloads moved from command-line arguments to standard input, so long tool output and other large messages no longer exceed Windows limits. #652

Setup and diagnostics

1. Made agent session tracking hooks faster and more reliable. Replaced the PowerShell bridge with native wtcli integration, reducing per-event overhead while ensuring hook failures never block supported agent CLIs. #571
2. Made agent-hook setup easier to diagnose. Installed and bundled plugin versions are shown, the OpenCode status row is fixed, and each CLI whose hooks fail to install is named. #635, #636

Bug Fixes

Sessions and processes

1. Kept session management scoped to the correct agent and environment. Switching agents or settings no longer shows blank, stale, or cross-distribution session lists. #646
2. Closed ACP sessions when their tabs close. Stale conversations are no longer left active after the owning tab is destroyed, including across window moves and pending connections. #637
3. Cleaned up Agent helpers and their child processes when panes disconnect or crash. Closed panes no longer leave orphaned processes, stale sessions, or unwanted automatic restarts behind. #647, #649

Agent Pane

1. Fixed completed responses appearing truncated until the next interaction. Expanded turns settle into their final layout immediately. #595
2. Preserved multiline prompts when reopening completed turns. Shift+Enter line breaks stay readable in expanded history while collapsed summaries stay compact. #614
3. Refreshed the model catalog after BYOK providers are added or removed. The model dropdown recovers immediately and no longer shows an unusable expander before a provider exists. #632
4. Restored right-click copy and paste in the Agent Pane. Right-click copies an active selection or pastes into the current draft when nothing is selected. #642
5. Fixed agent-requested tab and pane actions being rejected when they included valid placement directions. Proposed actions can open in the intended location. #599

What's coming up?

So what’s next and exciting? Terminal session restore is something we're actively working on, alongside richer tab content. See the discussion on this here: #653. Stay tuned!

We'll keep iterating on bugs and feature requests rapidly while Intelligent Terminal is in this experimental stage. Please file issues to help us make this product better!