bone-agent 1.3.2 → 1.3.3

package/src/llm/providers.py

@@ -173,17 +173,27 @@ class AnthropicHandler:
         # OpenAI format: {"choices": [{"message": {"content": "..."}}], "usage": {...}}
 
         # Convert Anthropic usage format (input_tokens/output_tokens) to OpenAI format (prompt_tokens/completion_tokens)
+        # Anthropic's input_tokens does NOT include cache tokens; total input =
+        #   input_tokens + cache_read_input_tokens + cache_creation_input_tokens
         anthropic_usage = response_json.get("usage", {})
+        cache_read = anthropic_usage.get('cache_read_input_tokens', 0)
+        cache_creation = anthropic_usage.get('cache_creation_input_tokens', 0)
+        prompt_tokens = anthropic_usage.get('input_tokens', 0) + cache_read + cache_creation
+        completion_tokens = anthropic_usage.get('output_tokens', 0)
         openai_format_usage = {
-            'prompt_tokens': anthropic_usage.get('input_tokens', 0),
-            'completion_tokens': anthropic_usage.get('output_tokens', 0),
-            'total_tokens': anthropic_usage.get('input_tokens', 0) + anthropic_usage.get('output_tokens', 0),
+            'prompt_tokens': prompt_tokens,
+            'completion_tokens': completion_tokens,
+            'total_tokens': prompt_tokens + completion_tokens,
         }
         # Preserve Anthropic cache token fields for the token tracker
         if 'cache_read_input_tokens' in anthropic_usage:
             openai_format_usage['cache_read_input_tokens'] = anthropic_usage['cache_read_input_tokens']
         if 'cache_creation_input_tokens' in anthropic_usage:
             openai_format_usage['cache_creation_input_tokens'] = anthropic_usage['cache_creation_input_tokens']
+        # Preserve non-cache input count so cost estimation can bill only the
+        # non-cache portion without relying on fragile prompt_tokens subtraction.
+        if 'input_tokens' in anthropic_usage:
+            openai_format_usage['input_tokens'] = anthropic_usage['input_tokens']
 
         result = {
             "choices": [],
@@ -279,17 +289,26 @@ class AnthropicHandler:
 
         # Yield usage data as final item if captured
         # Convert Anthropic format (input_tokens/output_tokens) to OpenAI format (prompt_tokens/completion_tokens)
+        # Anthropic's input_tokens does NOT include cache tokens; total input =
+        #   input_tokens + cache_read_input_tokens + cache_creation_input_tokens
         if usage_data:
+            cache_read = usage_data.get('cache_read_input_tokens', 0)
+            cache_creation = usage_data.get('cache_creation_input_tokens', 0)
+            prompt_tokens = usage_data.get('input_tokens', 0) + cache_read + cache_creation
+            completion_tokens = usage_data.get('output_tokens', 0)
             openai_format_usage = {
-                'prompt_tokens': usage_data.get('input_tokens', 0),
-                'completion_tokens': usage_data.get('output_tokens', 0),
-                'total_tokens': usage_data.get('input_tokens', 0) + usage_data.get('output_tokens', 0),
+                'prompt_tokens': prompt_tokens,
+                'completion_tokens': completion_tokens,
+                'total_tokens': prompt_tokens + completion_tokens,
             }
             # Preserve Anthropic cache token fields for the token tracker
             if 'cache_read_input_tokens' in usage_data:
                 openai_format_usage['cache_read_input_tokens'] = usage_data['cache_read_input_tokens']
             if 'cache_creation_input_tokens' in usage_data:
                 openai_format_usage['cache_creation_input_tokens'] = usage_data['cache_creation_input_tokens']
+            # Preserve non-cache input count for accurate cost estimation
+            if 'input_tokens' in usage_data:
+                openai_format_usage['input_tokens'] = usage_data['input_tokens']
             yield {'__usage__': openai_format_usage}
 
     @staticmethod

package/src/llm/token_tracker.py

@@ -55,6 +55,10 @@ class TokenTracker:
         self.total_cache_creation_tokens = 0   # Cumulative input tokens written to cache
         self.conv_cache_read_tokens = 0        # Per-conversation cache read tokens
         self.conv_cache_creation_tokens = 0    # Per-conversation cache creation tokens
+
+        # Active prompt variant (loaded from prompts/ directory)
+        self.current_variant = "main"
+
     def add_usage(self, usage_data, model_name: str = ""):
         """Add token usage from an API response.
 
@@ -122,7 +126,19 @@ class TokenTracker:
             # Fallback: look up cost rates from config
             cost_in, cost_out = get_model_cost(model_name)
             if cost_in > 0 or cost_out > 0:
-                computed = self._calculate_cost(prompt_tokens, completion_tokens, cost_in, cost_out)
+                # Compute the billable (non-cache) input token count for cost
+                # estimation.  Providers normalize `prompt_tokens` differently:
+                #   - Anthropic handler: sums input + cache_read + cache_creation
+                #   - OpenAI: prompt_tokens natively includes cached_tokens
+                #   - Future providers: may exclude cache tokens from prompt_tokens
+                # Use the explicit `input_tokens` field (Anthropic native,
+                # non-cache portion) when available; otherwise subtract cache
+                # tokens from prompt_tokens (assumes prompt_tokens includes
+                # cache counts).
+                base_prompt = usage_data.get('input_tokens')
+                if base_prompt is None:
+                    base_prompt = max(0, prompt_tokens - cache_read - cache_creation)
+                computed = self._calculate_cost(base_prompt, completion_tokens, cost_in, cost_out)
                 self.add_estimated_cost(computed['total_cost'])
 
     def add_actual_cost(self, cost_usd: float):

package/src/tools/edit.py

@@ -223,9 +223,11 @@ def _prepare_edit(arguments, repo_root, gitignore_spec=None, vault_root=None) ->
     if not path or not isinstance(path, str) or not path.strip():
         raise FileEditError("Missing or invalid 'path' parameter")
 
-    # Memory files (.bone/ and user_memory.md) are auto-approved writes that the
-    # system itself adds to .gitignore, so gitignore filtering would block them.
-    is_memory = Path(path).parent.name == ".bone" or Path(path).name == "user_memory.md"
+    # Memory files (.bone/ under repo root and user_memory.md) are auto-approved
+    # writes that the system itself adds to .gitignore, so gitignore filtering
+    # would block them. Must anchor to repo_root to avoid matching any .bone/ dir.
+    _resolved = (repo_root / path).resolve()
+    is_memory = str(_resolved).startswith(str((repo_root / ".bone").resolve()) + os.sep) or Path(path).name == "user_memory.md"
 
     # Resolve and validate path using PathResolver
     try:
@@ -236,9 +238,9 @@ def _prepare_edit(arguments, repo_root, gitignore_spec=None, vault_root=None) ->
         raise FileEditError(str(e), details=e.details)
 
     if not file_path.exists():
-        # Auto-create memory files (.bone/) with default header on first write.
-        # These are already auto-approved, so directory+file creation is safe.
-        if file_path.parent.name == ".bone" or file_path.name == "user_memory.md":
+        # Auto-create memory files (.bone/ under repo root) with default header
+        # on first write. Already auto-approved, so creation is safe.
+        if is_memory:
             file_path.parent.mkdir(parents=True, exist_ok=True)
             header = "# Project Memory\n\n" if file_path.name == "agents.md" else "# User Memory\n\n"
             file_path.write_text(header, encoding="utf-8")

package/src/tools/helpers/path_resolver.py

@@ -107,24 +107,30 @@ class PathResolver:
             # Resolve to absolute path (handles .. and symlinks)
             path = path.resolve()
 
-            # Step 2b: Security boundary — path must be within repo_root or vault_path
+            # Step 2b: Security boundary — path must be within repo_root, vault_path,
+            # or the agent's own data directory (~/.bone/).
             if enforce_boundary:
                 try:
                     path.relative_to(self.repo_root)
                 except ValueError:
-                    if self.vault_path is not None:
-                        try:
-                            path.relative_to(self.vault_path)
-                        except ValueError:
+                    # Check ~/.bone/ — agent data dir is always accessible
+                    bone_root = Path.home() / ".bone"
+                    try:
+                        path.relative_to(bone_root)
+                    except ValueError:
+                        if self.vault_path is not None:
+                            try:
+                                path.relative_to(self.vault_path)
+                            except ValueError:
+                                elapsed = time.time() - start_time
+                                _track_validation_error("outside_allowed_roots")
+                                _path_resolution_times.append(elapsed)
+                                return None, f"Path is outside allowed directories: {path_str}"
+                        else:
                             elapsed = time.time() - start_time
-                            _track_validation_error("outside_allowed_roots")
+                            _track_validation_error("outside_repo")
                             _path_resolution_times.append(elapsed)
-                            return None, f"Path is outside allowed directories: {path_str}"
-                    else:
-                        elapsed = time.time() - start_time
-                        _track_validation_error("outside_repo")
-                        _path_resolution_times.append(elapsed)
-                        return None, f"Path is outside repository: {path_str}"
+                            return None, f"Path is outside repository: {path_str}"
 
             # Step 3: Check existence if required
             if must_exist:

package/src/tools/rg_search.py

@@ -2,21 +2,84 @@
 
 import logging
 import re
-import shlex
+import stat
 import subprocess
 from pathlib import Path
 from typing import Optional
 
 from .helpers.base import tool
 from .helpers.formatters import format_tool_result
-from .shell import _prepare_execution_environment, run_shell_command
+from .shell import _execute_direct_command, _prepare_execution_environment
 from .helpers.converters import coerce_bool, coerce_int
+from utils.settings import tool_settings
 
 logger = logging.getLogger(__name__)
 
 # Default match limit for vault searches (separate from repo limit)
 _VAULT_MAX_MATCHES = 20
 
+# Regex for detecting file-path lines in rg output (shared by _annotate_file_sizes and _search_vault)
+_path_line_re = re.compile(r"^[^\s:|].*[/.]")
+
+
+def _format_file_size(size_bytes: int) -> str:
+    """Format file size in human-readable form."""
+    if size_bytes < 1024:
+        return f"{size_bytes} B"
+    elif size_bytes < 1024 * 1024:
+        return f"{size_bytes / 1024:.1f} KB"
+    else:
+        return f"{size_bytes / (1024 * 1024):.1f} MB"
+
+
+def _annotate_file_sizes(formatted_output: str, base_path: Path, output_mode: str = "files_with_matches") -> str:
+    """Append human-readable file sizes to each file path line in rg output.
+
+    Works on files_with_matches and count output modes where each content
+    line starts with a file path. Skips metadata, truncation, and section
+    header lines. Skipped entirely for content mode (no benefit there).
+    """
+    if output_mode == "content":
+        return formatted_output
+
+    lines = formatted_output.split("\n")
+    annotated = []
+    for line in lines:
+        stripped = line.strip()
+        if (not stripped
+                or stripped.startswith("exit_code=")
+                or stripped.startswith("matches=")
+                or stripped.startswith("files=")
+                or stripped.startswith("... (")
+                or stripped.startswith("[repo]")
+                or stripped.startswith("[vault]")):
+            annotated.append(line)
+            continue
+        # Only annotate pure file-path lines (files_with_matches: "file",
+        # count: "file:N"). Skip content-mode match lines ("file:line:match")
+        # which always have 2+ colons or a colon-digit-dash pattern.
+        parts = line.split(":")
+        is_file_line = (
+            _path_line_re.match(line)
+            and len(parts) <= 2
+            and (len(parts) == 1 or parts[1].strip().isdigit())
+        )
+        if is_file_line:
+            file_part = parts[0].strip()
+            full_path = base_path / file_part
+            try:
+                st = full_path.stat()
+                if stat.S_ISREG(st.st_mode):
+                    size = _format_file_size(st.st_size)
+                    annotated.append(f"{line}  {size:>8}")
+                else:
+                    annotated.append(line)
+            except (OSError, ValueError):
+                annotated.append(line)
+        else:
+            annotated.append(line)
+    return "\n".join(annotated)
+
 
 @tool(
     name="rg",
@@ -103,72 +166,74 @@ def rg(
     if not isinstance(pattern, str) or not pattern.strip():
         return "exit_code=1\nrg requires a non-empty 'pattern' argument."
 
-    # Build rg command from arguments
-    cmd_parts = ["rg"]
+    # Build rg args as a list (bypass string roundtrip to avoid shlex issues with regex metacharacters)
+    args = []
 
     # Add --line-number for content mode
     if output_mode == "content":
-        cmd_parts.append("--line-number")
+        args.append("--line-number")
 
     # Add multiline flag
     multiline = coerce_bool(kwargs.get("multiline"), default=False)
     if multiline:
-        cmd_parts.append("-U")
-        cmd_parts.append("--multiline-dotall")
+        args.append("-U")
+        args.append("--multiline-dotall")
 
     # Add case insensitive flag
     case_insensitive = coerce_bool(kwargs.get("case_insensitive"), default=False)
     if case_insensitive:
-        cmd_parts.append("--ignore-case")
+        args.append("--ignore-case")
 
     # Add context lines flag
     context_lines = coerce_int(kwargs.get("context_lines"))[0] if kwargs.get("context_lines") else None
     if context_lines:
-        cmd_parts.append(f"--context={context_lines}")
+        args.append(f"--context={context_lines}")
 
     # Add glob pattern
     if glob:
-        cmd_parts.append(f"--glob={glob}")
+        args.append(f"--glob={glob}")
 
     # Add file type filter
     file_type = kwargs.get("type")
     if file_type:
-        cmd_parts.append(f"--type={file_type}")
+        args.append(f"--type={file_type}")
 
-    # Add files-with-matches flag for count mode
+    # Add output mode flags
     if output_mode == "files_with_matches":
-        cmd_parts.append("--files-with-matches")
+        args.append("--files-with-matches")
     elif output_mode == "count":
-        cmd_parts.append("--count")
+        args.append("--count")
 
-    # Add pattern - quote if it contains spaces
-    if " " in pattern:
-        cmd_parts.append(shlex.quote(pattern))
-    else:
-        cmd_parts.append(pattern)
+    # Pattern and search path — no quoting needed, subprocess list form bypasses shell
+    args.append(pattern)
 
-    # Add path (default to current directory)
     search_path = path or "."
-    cmd_parts.append(search_path)
-
-    # Build command string
-    command = " ".join(cmd_parts)
+    args.append(search_path)
 
     # Get max_matches from kwargs (default: 100, set to 0 for no limit)
     raw = coerce_int(kwargs.get("max_matches"))[0] if kwargs.get("max_matches") is not None else None
     max_matches = raw if raw is not None and raw >= 0 else 100
 
-    # Execute repo search
+    # Execute repo search directly (no string→shlex roundtrip)
     try:
-        repo_result = run_shell_command(
-            command, repo_root, rg_exe_path, console, debug_mode, gitignore_spec,
-            max_matches=max_matches
+        env = _prepare_execution_environment(repo_root, rg_exe_path)
+
+        result = _execute_direct_command(
+            [str(rg_exe_path)] + args,
+            repo_root, env, debug_mode, console,
+        )
+
+        command_display = "rg " + " ".join(args)
+        repo_result = format_tool_result(
+            result, command=command_display, is_rg=True,
+            debug_mode=debug_mode, max_matches=max_matches,
         )
     except Exception as e:
         return f"exit_code=1\nrg command failed: {str(e)}"
 
     # If no vault configured, return repo results directly
     if not vault_root:
+        repo_result = _annotate_file_sizes(repo_result, repo_root, output_mode)
         return repo_result
 
     # Run vault search and merge results
@@ -184,9 +249,12 @@ def rg(
     )
 
     if not vault_output:
+        repo_result = _annotate_file_sizes(repo_result, repo_root, output_mode)
         return repo_result
 
     # Merge results: repo section + vault section with absolute paths
+    repo_result = _annotate_file_sizes(repo_result, repo_root, output_mode)
+    vault_output = _annotate_file_sizes(vault_output, Path(vault_root), output_mode)
     return _merge_results(repo_result, vault_output, output_mode)
 
 
@@ -257,7 +325,7 @@ def _search_vault(vault_root, rg_exe_path, output_mode, debug_mode, console,
             text=True,
             encoding="utf-8",
             errors="replace",
-            timeout=30,
+            timeout=tool_settings.command_timeout_sec,
             cwd=str(vault_path),
             env=env,
         )
@@ -288,7 +356,6 @@ def _search_vault(vault_root, rg_exe_path, output_mode, debug_mode, console,
         # rg output: "relative/path:linenum:match" or "relative/path-linenum-context"
         # or "relative/path:count" (count mode).  Must contain / or . before any
         # colon to avoid matching content-only lines or binary headers.
-        _path_line_re = re.compile(r"^[^\s:|].*[/.]")
         vault_prefix = str(vault_path)
 
         lines = formatted.split("\n")

package/src/ui/commands.py

@@ -1,6 +1,8 @@
 """Command routing and help display."""
 
+import os
 import re
+import subprocess
 from dataclasses import dataclass
 from typing import Optional
 from llm import config
@@ -155,6 +157,9 @@ def _cron_remove(console, sub_args, cron_config, notify_scheduler):
     if not job_id:
         console.print("[red]Usage: /cron remove <id>[/red]")
         return CommandResult(status="handled")
+    if job_id == "dream":
+        console.print("[red]The 'dream' job is managed by DREAM_SETTINGS.enabled in config.yaml and cannot be removed.[/red]")
+        return CommandResult(status="handled")
     if cron_config.remove_job(job_id):
         notify_scheduler()
         console.print(f"[green]Removed cron job '{job_id}'[/green]")
@@ -171,6 +176,9 @@ def _cron_toggle(console, sub_args, cron_config, notify_scheduler, enable):
     if not job_id:
         console.print(f"[red]Usage: /cron {verb} <id>[/red]")
         return CommandResult(status="handled")
+    if not enable and job_id == "dream":
+        console.print("[red]The 'dream' job is managed by DREAM_SETTINGS.enabled in config.yaml and cannot be disabled via /cron.[/red]")
+        return CommandResult(status="handled")
     if job_id in cron_config.jobs:
         cron_config.update_job(job_id, enabled=enable)
         notify_scheduler()
@@ -381,6 +389,12 @@ def _handle_config(chat_manager, console, debug_mode_container, args, cron_sched
                 {"value": "danger", "text": "DANGER"},
             ],
         ),
+        SettingOption(
+            key="memory_enabled", text="Memory",
+            value=config.MEMORY_SETTINGS.get("enabled", True),
+            input_type="boolean",
+            on_text="ON", off_text="OFF",
+        ),
     ]
 
     # Build status bar settings
@@ -480,6 +494,10 @@ def _handle_config(chat_manager, console, debug_mode_container, args, cron_sched
                 console.print("[bold red on default]  Dangerous git commands are still blocked.[/bold red on default]")
                 console.print("[bold yellow on default]  Use at your own risk![/bold yellow on default]")
                 console.print()
+        elif key == "memory_enabled":
+            config.update_memory_settings({"enabled": value})
+            state = "enabled" if value else "disabled"
+            change_lines.append(f"  Memory: {state}")
         elif key == "compact_trigger_tokens":
             context_settings.compact_trigger_tokens = int(value)
             change_lines.append(f"  Compaction Threshold: {value:,} tokens")
@@ -531,6 +549,17 @@ def _handle_config(chat_manager, console, debug_mode_container, args, cron_sched
         except Exception as e:
             console.print(f"[red]Failed to save status bar settings: {e}[/red]")
 
+    # Persist memory setting to config
+    if "memory_enabled" in changes:
+        try:
+            cfg_data = config_manager.load(force_reload=True)
+            if "MEMORY_SETTINGS" not in cfg_data:
+                cfg_data["MEMORY_SETTINGS"] = {}
+            cfg_data["MEMORY_SETTINGS"]["enabled"] = changes["memory_enabled"]
+            config_manager.save(cfg_data)
+        except Exception as e:
+            console.print(f"[red]Failed to save memory settings: {e}[/red]")
+
     # Display summary
     console.print(f"[green]Settings updated:[/green]")
     for line in change_lines:
@@ -561,12 +590,13 @@ def _handle_clear(chat_manager, console, debug_mode_container, args, cron_schedu
     conv_cache_read = chat_manager.token_tracker.conv_cache_read_tokens
     conv_cache_creation = chat_manager.token_tracker.conv_cache_creation_tokens
     if conv_cache_read > 0 or conv_cache_creation > 0:
+        total_cached = conv_cache_read + conv_cache_creation
         cache_hit_pct = (
-            conv_cache_read / conv_in * 100
-        ) if conv_in > 0 else 0
+            conv_cache_read / total_cached * 100
+        ) if total_cached > 0 else 0
         console.print(f"  Cache read:   {conv_cache_read:,} tokens")
         console.print(f"  Cache write:  {conv_cache_creation:,} tokens")
-        console.print(f"  ({cache_hit_pct:.0f}% of input served from cache)")
+        console.print(f"  ({cache_hit_pct:.0f}% cache hit rate)")
 
     # Display cost — combined actual + estimated, with config-based fallback
     tracker_conv = chat_manager.token_tracker
@@ -1065,10 +1095,11 @@ def _handle_usage(chat_manager, console, debug_mode_container, args, cron_schedu
     # Display cache token breakdown (if any cache tokens were recorded)
     has_cache = tracker.total_cache_read_tokens > 0 or tracker.total_cache_creation_tokens > 0
     if has_cache:
+        total_cached = tracker.total_cache_read_tokens + tracker.total_cache_creation_tokens
         cache_hit_pct = (
             tracker.total_cache_read_tokens
-            / tracker.total_prompt_tokens * 100
-        ) if tracker.total_prompt_tokens > 0 else 0
+            / total_cached * 100
+        ) if total_cached > 0 else 0
         console.print()
         console.print(f"[#5F9EA0]Input Cache ({cache_hit_pct:.0f}% hit rate):[/#5F9EA0]")
         console.print(f"  Cache read:   {tracker.total_cache_read_tokens:,} tokens")
@@ -2433,6 +2464,57 @@ def _handle_cd(chat_manager, console, debug_mode_container, args, cron_scheduler
     return CommandResult(status="handled")
 
 
+def _handle_prompt(chat_manager, console, debug_mode_container, args, cron_scheduler=None):
+    """Handle /prompt command — show/swap prompt variants."""
+    from utils.settings import prompt_settings
+    from llm.prompts import _variant_available, _list_variants
+
+    cfg_manager = config_manager
+
+    if not args or args.strip() == "list":
+        variants = _list_variants()
+        current = prompt_settings.variant
+        console.print()
+        console.print(f"[bold #5F9EA0]Prompt Variants[/bold #5F9EA0]  (current: [bold]{current}[/bold])")
+        console.print()
+        for v in variants:
+            marker = "[bold green]active[/bold green]" if v == current else ""
+            console.print(f"  [bold]{v}[/bold]  {marker}")
+        console.print()
+        console.print("[dim]Switch with: [bold #5F9EA0]/prompt main[/bold #5F9EA0] or [bold #5F9EA0]/prompt micro[/bold #5F9EA0][/dim]")
+        return CommandResult(status="handled")
+
+    # Single arg: variant name to switch to
+    target = args.strip().lower()
+
+    if not _variant_available(target):
+        variants = _list_variants()
+        console.print(f"[red]Unknown variant: '{target}'[/red]")
+        console.print(f"[dim]Available: {', '.join(variants)}[/dim]")
+        return CommandResult(status="handled")
+
+    # Update settings
+    prompt_settings.variant = target
+
+    # Persist to config
+    try:
+        cfg_data = cfg_manager.load(force_reload=True)
+        if "PROMPT_SETTINGS" not in cfg_data:
+            cfg_data["PROMPT_SETTINGS"] = {}
+        cfg_data["PROMPT_SETTINGS"]["variant"] = target
+        cfg_manager.save(cfg_data)
+    except Exception as e:
+        console.print(f"[red]Failed to save variant to config: {e}[/red]")
+        console.print("[yellow]Variant applied for this session only — it will revert on restart.[/yellow]")
+
+    # Rebuild system prompt in-place (no restart)
+    chat_manager.update_system_prompt(variant=target)
+    console.print(f"[green]Switched to '{target}' variant[/green]")
+    console.print("[dim]System prompt rebuilt in-place.[/dim]")
+
+    return CommandResult(status="handled")
+
+
 def _handle_obsidian_init(console, obsidian_settings):
     """Handle /obsidian init — scaffold project folder structure in vault."""
     if not obsidian_settings.is_active():
@@ -2658,9 +2740,34 @@ _COMMAND_HANDLERS = {
     "/cd": _handle_cd,
     "/setup": _handle_setup,
     "/cron": _handle_cron,
+    "/prompt": _handle_prompt,
 }
 
 
+def _handle_shell_command(console, command):
+    """Execute a shell command prefixed with : and display output."""
+    from utils.settings import tool_settings
+    try:
+        result = subprocess.run(
+            ["/bin/sh", "-c", command], capture_output=True, text=True,
+            encoding="utf-8", errors="replace", timeout=tool_settings.command_timeout_sec,
+        )
+        output = ((result.stdout or "") + (result.stderr or "")).strip() or "(no output)"
+        lines = output.splitlines()
+        if len(lines) > 200:
+            output = "\n".join(lines[:100]) + f"\n\n... ({len(lines) - 200} lines omitted) ...\n\n" + "\n".join(lines[-100:])
+        console.print()
+        if result.returncode != 0:
+            console.print(f"[red]exit code: {result.returncode}[/red]")
+        console.print(output)
+        console.print()
+    except subprocess.TimeoutExpired:
+        console.print(f"[red]Command timed out after {tool_settings.command_timeout_sec}s[/red]")
+    except Exception as e:
+        console.print(f"[red]Error: {e}[/red]")
+    return CommandResult(status="handled")
+
+
 def process_command(chat_manager, user_input, console, debug_mode_container, cron_scheduler=None):
     """Process command and optionally return replacement content.
 
@@ -2681,6 +2788,14 @@ def process_command(chat_manager, user_input, console, debug_mode_container, cro
     cmd = parts[0].lower()
     args = parts[1] if len(parts) > 1 else None
 
+    # Shell command prefix (:command)
+    if user_input.startswith(":"):
+        shell_cmd = user_input[1:].strip()
+        if shell_cmd:
+            result = _handle_shell_command(console, shell_cmd)
+            return (result.status, result.replacement_input)
+        return ("handled", None)
+
     # Look up handler in registry
     handler = _COMMAND_HANDLERS.get(cmd)
     if handler:

package/src/ui/displays.py

@@ -70,6 +70,7 @@ def show_help_table(console):
     table.add_row("[bold #5F9EA0]/tools[/bold #5F9EA0] [list|enable|disable|enable-group|disable-group]", "Toggle tools or groups (e.g. file_ops, task_mgmt)")
     table.add_row("[bold #5F9EA0]/setup[/bold #5F9EA0]", "Re-run the first-run setup wizard")
     table.add_row("[bold #5F9EA0]/cron[/bold #5F9EA0] [list|add|remove|enable|disable|run]", "Manage scheduled cron jobs")
+    table.add_row("[bold #5F9EA0]:[/bold #5F9EA0]<command>", "Run a shell command (e.g. :git status)")
 
 
     console.print(Panel(table, title="[bold #5F9EA0]Commands[/bold #5F9EA0]", border_style="grey23", padding=(0, 2)))

package/src/ui/main.py

@@ -507,6 +507,7 @@ def main():
                 thinking_indicator.start()
                 INPUT_BLOCKED['blocked'] = True
                 try:
+                    console.print("─" * console.width, style="rgb(30,30,30)")
                     console.print()  # Extra newline after user input to separate from LLM response
                     # Add user message
                     if TOOLS_ENABLED:

package/src/utils/settings.py

@@ -28,10 +28,14 @@ def left_align_headings(text: str) -> str:
 @dataclass
 class ServerSettings:
     """Local llama-server configuration."""
-    ngl_layers: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("ngl_layers", 30))
+    ngl_layers: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("ngl_layers", 99))
     ctx_size: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("ctx_size", 8192))
     n_predict: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("n_predict", 8192))
     rope_scale: float = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("rope_scale", 1.0))
+    threads: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("threads", 4))
+    batch_size: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("batch_size", 2048))
+    ubatch_size: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("ubatch_size", 512))
+    flash_attn: bool = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("flash_attn", True))
     health_check_timeout_sec: int = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("health_check_timeout_sec", 120))
     health_check_interval_sec: float = field(default_factory=lambda: _CONFIG.get("SERVER_SETTINGS", {}).get("health_check_interval_sec", 1.0))
 
@@ -104,6 +108,18 @@ class ContextSettings:
             self.hard_limit_tokens = int(self.max_context_window * 0.9)
 
 
+@dataclass
+class PromptSettings:
+    """Prompt variant selection."""
+    variant: str = field(default_factory=lambda: _CONFIG.get("PROMPT_SETTINGS", {}).get("variant", "micro"))
+
+
+@dataclass
+class DreamSettings:
+    """Dream memory consolidation settings."""
+    enabled: bool = field(default_factory=lambda: _CONFIG.get("DREAM_SETTINGS", {}).get("enabled", True))
+
+
 @dataclass
 class ObsidianSettings:
     """Obsidian vault integration settings.
@@ -165,8 +181,9 @@ tool_settings = ToolSettings()
 file_settings = FileSettings()
 context_settings = ContextSettings()
 sub_agent_settings = SubAgentSettings()
+dream_settings = DreamSettings()
 obsidian_settings = ObsidianSettings()
-
+prompt_settings = PromptSettings()
 # Tool execution constants
 MAX_TOOL_CALLS = tool_settings.max_tool_calls
 MAX_COMMAND_OUTPUT_LINES = tool_settings.max_command_output_lines