v0.7.0

What's new

Model switching mid-session

New /model command lets you switch models without restarting lmcode:

• /model list — show all downloaded models with size and load status
• /model load <id> — load a different model and reset the conversation
• /model unload — unload the current model from memory

Auto-start LM Studio at startup

When you run lmcode with LM Studio closed, it now automatically tries to bring it up:

1. Tries lms server start (if LM Studio GUI is already open but server is off)
2. Falls back to lms daemon up (fully headless, no GUI needed)

Animated dots keep you informed while it starts — no more frozen terminal.

Interactive model picker

When no model is loaded, an arrow-key menu lets you pick from your downloaded models directly from lmcode. Animated dots while the model loads.

ASCII art logo in startup menus

The lmcode banner now appears above every startup menu.

Fixes

• Fast startup: socket pre-check (0.5 s) before SDK probe eliminates the multi-second freeze when the server is not running
• lms load now uses correct --gpu auto arg format (was --gpu=auto)
• Model identifier resolved from modelKey field in lms ls --json

v0.6.1

What's changed

Refactored

• agent/core.py split into focused submodules — the 1 259-line monolith is now four files:
  • _noise.py — SDK WebSocket noise suppression (installed at import time)
  • _display.py — all print/render helpers, diff rendering, SLASH_COMMANDS
  • _prompt.py — prompt-toolkit PromptSession factory, Tab key bindings, autocomplete
  • core.py — slimmed to ~820 lines, owns Agent, run_chat, system prompt

Fixed

• write_file mixed newline unescape — the previous guard "\n" not in content silently skipped unescaping when a model (e.g. Qwen 7B) emitted a mix of real and literal \n sequences, producing syntactically invalid Python files. Guard removed; unescape now always runs when a literal \n is present.

Tests

• tests/test_agent/test_display.py — full coverage for _ctx_usage_line, _build_stats_line, _format_tool_signature, _render_diff_sidebyside, _print_history, SLASH_COMMANDS.
• tests/test_agent/test_noise.py — coverage for _FilterSDKNoise stderr wrapper and _FilteredLastResort logging interceptor.

Full changelog: https://github.com/VforVitorio/lmcode/blob/main/CHANGELOG.md

v0.6.0

What's new in v0.6.0

Fixed

• LM Studio disconnect — closing LM Studio mid-session now shows a clean message instead of a raw traceback:

  LM Studio disconnected
  → restart LM Studio and run lmcode again
  

• SDK WebSocket JSON noise — the {"event": "Websocket failed, terminating session."} lines that appeared on disconnect are now fully suppressed. Root cause: Python's logging.lastResort bypasses root-logger filters; the fix wraps lastResort.handle() directly.

Closes #70.

v0.5.0

What's new in v0.5.0

Added

• Ctrl+C interrupt mid-generation — pressing Ctrl+C returns to the input prompt instead of exiting lmcode. Shows ^C + italic interrupted with a Rule separator. Chat history is rolled back cleanly so the next turn has no dangling context.
• Syntax-highlighted diff blocks for write_file — when a file is overwritten the agent displays a side-by-side diff with +/- counts and Catppuccin colours. New files get a dedicated "new file" panel.
• Playground folder — playground/ sandbox for end-to-end manual testing of all UI features.
• /history [n] slash command to review past turns as Rich panels.
• Ctrl+R reverse history search in the input prompt.
• Tab mode cycling — Tab cycles through ask → auto → strict → ask permission modes.
• Ghost-text autocomplete for slash commands (accepted with Tab or →).

Changed

• run_shell output is now wrapped in an IN/OUT panel with a separator Rule and uppercase labels.
• read_file results render in a one-dark themed panel with violet rounded border and line numbers.
• System prompt rewritten with tool-first framing, hard constraints, and environment block — tuned for reliable tool-calling on 7B models.
• Tool docstrings rewritten with explicit use-case guidance for better tool-selection on small models.

Fixed

• Verbose panels always shown — the SDK calls tools with positional args, leaving kwargs={}. Fixed with inspect.signature positional-arg merge in _wrap_tool_verbose.
• write_file escape sequences — models that emit literal \n / \t / \" (no real newlines) now have those sequences unescaped before writing.
• SDK "already closed channel" noise — the warning emitted after Ctrl+C is suppressed via a root-logger logging.Filter and a sys.stderr wrapper.
• Completion menu colours, CI lint/typecheck failures.

Tests

79 tests passing. New coverage for write_file escape sequences, _wrap_tool_verbose positional args, keyboard interrupt propagation, and chat rollback after Ctrl+C.

v0.4.0 — animated spinner, ghost-text autocomplete, file preview & /history

What's new in v0.4.0

Spinner & status labels

The thinking spinner now shows animated dots and changes label based on what the agent is doing:

• thinking… — waiting for the model's first token
• working… — a tool call is in progress (shows tool name + path)
• finishing… — model is generating the final response

Ghost-text slash autocomplete

Type / and the first matching command appears as dim ghost text. Press Tab to accept it inline — no popup menu, no noise.

• /h → ghost text elp → Tab → /help

Tab mode cycling

Tab without a / prefix cycles the permission mode: ask → auto → strict.

Persistent history (Ctrl+R)

Prompts are saved to ~/.lmcode/history. Use Ctrl+R or the Up-arrow to recall previous inputs across sessions.

read_file syntax panel

When the agent reads a file, the result is shown as a monokai syntax-highlighted panel (up to 20 lines with line numbers) instead of a plain one-liner.

/history [N]

New slash command — prints the last N conversation turns as bordered panels (default 5).

/history      # last 5 turns
/history 10   # last 10 turns

Banner fix

The compact banner fallback now triggers at 70 columns (was 90), so the full ASCII logo shows in more terminal widths.

Issues closed

• #44 ghost-text slash autocomplete
• #45 animated spinner state labels
• #51 spinner style improvements

v0.3.0 — /compact, context indicator & cycling tips

What's new in v0.3.0

/compact — summarise history to free context space

Run /compact when the conversation is getting long. lmcode asks the model to summarise everything discussed so far, resets the chat, and injects the summary as context. A bouncingBar spinner shows while compacting, then a panel displays the summary preview and how many messages were collapsed.

Context window indicator (#41)

• /status and /tokens now show a live arc indicator: ◔ 38% (12.4k / 32k tok)
• Arcs cycle through ○ ◔ ◑ ◕ ● as usage grows
• One-time warning printed after a response when the context exceeds 80%: ⚠ context at 83% — run /compact to summarise the conversation
• Percentage is calculated from the last turn's prompt tokens (the actual history size sent to the model), not a cumulative sum

Cycling tips (#38)

Tips below the spinner now rotate every 8 seconds through a shuffled list instead of staying static for the whole turn.

/tokens (#29)

New slash command showing session-wide token totals:

session tokens
  prompt (↑)        14.2k
  generated (↓)     1.8k
  total             16.0k
  context           ◔ 38%  (14.2k / 32k tok)

/hide-model (#36)

Toggles the model name in the live prompt:

• On: ● lmcode (qwen2.5-1.5b-instruct) [ask] ›
• Off: ● lmcode [ask] ›

Spinner keepalive

The spinner label now refreshes every 100 ms via an asyncio.create_task keepalive, keeping the animation smooth during the model prefill phase.

Requirements

• Python 3.12+
• LM Studio with local server enabled and a model loaded

Usage

pip install lmcode
lmcode

Closes: #29 #31 #36 #38 #41

v0.2.0 — UI polish, slash commands & error handling

What's new in v0.2.0

UI & interaction

• Spinner: dots style with violet (brand) colour instead of the old arc
• History entries: submitted prompt line rewrites in-place to a dim › message style — no more cluttered scrollback
• Exchange separator: plain Rule line after each response cleanly separates turns
• Rotating tips: a random tip appears below the spinner while the model is thinking
• Responsive banner: full ASCII art panel on wide terminals, compact 2-line fallback on narrow ones
• Mode badge removed from the startup banner (mode is always visible in the live prompt)

Slash commands

Command	What it does
/tools	List all registered tools with their signatures
/status	Show model, mode, verbose, tips, stats, and turn count
/tips	Toggle rotating tips during thinking
/stats	Toggle the token stats line after each response
/verbose	Toggle tool call and result display

Error handling

• No server: exits cleanly with a clear message when LM Studio's local server is not reachable
• No model loaded: exits with instructions before showing the banner
• Model ejected mid-session: catches LMStudioModelNotFoundError and exits gracefully instead of throwing a traceback
• XML in responses: response text is now printed via Text + highlight=False — Rich no longer mis-highlights XML tags that small models sometimes emit

System prompt

• Model is explicitly instructed to only call tools when the request requires file/shell access
• Model is instructed never to output raw XML, HTML, or JSON schemas

Settings

• UISettings added to lmcode.toml: spinner, show_tips, show_stats
• Token-aware file read limit: max_file_bytes is now computed from the model's actual context length via model.get_context_length()

Developer

• mypy clean: added Iterator[Path] return types to generator helpers in filesystem.py and search.py

Requirements

• Python 3.12+
• LM Studio with the local server enabled (default: localhost:1234) and a model loaded

Usage

pip install lmcode
# or: uv add lmcode
lmcode

Closes: #2 #13 #16 #20 #25 #26 #28 #30 #32 #33 #35

v0.1.0 — Basic chat with LM Studio

What's in this release

• CLI with ASCII art banner and color palette
• Interactive chat loop connecting to LM Studio via \model.act()
• Auto-detects loaded model (no need to specify --model)
• read_file\ tool with binary detection and truncation
• Per-repo context via \LMCODE.md\ files
• Plugin system scaffolding (pluggy)
• Session event models (JSONL format, recorder coming in v0.2.0)

Requirements

• Python 3.12+
• LM Studio running locally with a model loaded and the local server enabled (default: \localhost:1234)

Usage

pip install lmcode  
# or: uv add lmcode
lmcode chat