CMDAI CODE v1.0.1

CMDAI CODE: Timeline Navigation, Session Tabs & Full Persistence 🚀

This release introduces major TUI workflow enhancements designed to make multi-turn agent interactions faster, smoother, and completely resilient to state loss.

---

🌟 Highlights

⏱️ Interactive Timeline Navigation Rail

• Dedicated Navigation Rail: A clean, vertical timeline dock on the left side of the chat featuring interactive dots for every turn in the conversation.
• Instant Jump: Click any dot to smoothly scroll directly to that specific user prompt and agent response.
• Smart Hover Previews (Tooltips): Hovering over a dot reveals an informative preview tooltip showing the prompt summary, turn number, and exact timestamp.
• Zero Dot Overlap: Tooltips now anchor smartly to the right of the rail (x=3 into the chat area), so the hovered dot and all dots above and below remain 100% visible.

📑 Session Tabs Bar

• Browser-Style Tabs: A persistent tab bar at the top of the screen allows seamless switching between multiple active sessions with a single click.
• Tab Management: Easily open new sessions (+) or close existing ones (×) without losing conversation history.
• Smart Tab Replacement: Submitting a prompt from an empty initial session reuses the current tab instead of cluttering your workspace with duplicate empty sessions.

🛡️ Full Plan & Approval Persistence (Approval Bar & TODOs)

• No More Disappearing State: Fixed an issue where the file edit approval bar (● EDIT: 2 files [Approve] [Remove]) and the TODO preview bar would vanish after 3 seconds upon session load.
• Background Task Protection: Background task completions (e.g., long-running PowerShell commands) no longer dismiss the pending approval bar or prematurely trigger automatic generations.
• Instant /diff Inspection: Clicking click to inspect on the approval bar opens the diff editor immediately, even right after reloading an archived session.

📅 Accurate Session Dates (No More False "Today")

• Introduced strict last_written_ts tracking.
• A session is moved to "Today" only when you actually send a new message or the agent generates new code. Merely opening, viewing, or switching sessions preserves their original timestamps.

👁️ Dynamic Multimodal Detection (Vision & Screenshots)

• Dynamic detection replaces static model lists, automatically identifying vision capabilities across modern models
• The <tool:vision> tool is enabled and guarded dynamically so text-only models won't attempt unhandled vision calls.

---

🛠️ Detailed Changelog

🚀 Features & Architecture

• src/cmdai/tui/widgets.py: Added TimelineRail, TimelineDot, and SessionTabsBar components with click and hover interactions.
• src/cmdai/tui/styles.py: Added styling for the timeline rail, session tabs, and -timeline-tooltip offset classes.
• src/cmdai/core/sessions.py: Extended session persistence with todo_list, pending_approval, and non-bumping timestamp guards.

🐛 Bug Fixes & Improvements

• src/cmdai/tui/app.py:
  • Positioned timeline tooltips to the right of the rail to prevent obstructing adjacent dots.
  • Guarded _tick_meta_spinner against clearing the approval bar while user action is pending.
  • Rebuilt _replay_transcript to reconstruct active TODO queues and code-mode turn edits on reload.
  • Added fallback file resolution in _open_diff_window for pending approval targets.
• src/cmdai/core/capabilities.py: Added dynamic multimodal detection heuristics for API and local LLM backends.

CMDAI CODE v1.0.0

Autonomous coding assistant. Combines local GGUF models (llama.cpp) with multi-provider cloud AI and the MCP protocol.

---

Highlights

1. ⚡ Advanced Local GGUF (llama.cpp) Handling

• Full tool-calling loop parity with API models: eliminated the premature “Done” with empty responses on retries or failed tool steps. Local models now receive consistent feedback and continue until the task is fully resolved.
• Automatic Synthesis Pass: after tool execution the model always produces a clean, complete Markdown final answer for the user.
• Smart path suggestions: when read is called on a non-existent file it automatically suggests real files present in the working directory.
• Dynamic context limits in /modelinfo: accurate detection of context windows (e.g. 32k for Qwen2.5-Coder family) and safe RAM budgeting (70 % headroom rule).

2. 🎨 Unified TUI & New Animations

• Command tool: all terminal commands now surface under a single clean name (Command) instead of mixed shell / terminal / bash labels.
• Summarizing fixed: new two-phase context-compression animation (⌬ ↔ ✻) with a live timer and full expand/collapse support (same as the Thinking block).
• New /ctx alias for quick context info (alongside /context).
• Working-directory sync: 100 % consistency between workdir and current_cwd — ls, read, write and edit always operate on the correct directory.

3. 🌐 New AI Providers & Background Subagents

• Custom OpenAI-compatible endpoints in /models (Ctrl+A): easily add any compatible provider (e.g. Groq, Together AI, vLLM, DeepInfra).
• Background subagents (/bg / Ctrl+B): delegate research or verification tasks to subagents that run in the background with live progress view.

4. 📦 One-Command Installation

• All installation methods work
• More in README.md

And more under the hood improvements, polish and stability fixes.