docs: Add warm session provisioning README

Documentation for #327 implementation.

docs/warm-session-provisioning.md

@@ -0,0 +1,139 @@
+# Warm Session Provisioning
+
+**Issue:** #327
+
+## Overview
+
+Warm session provisioning allows creating pre-contextualized agent sessions that start with established patterns and context, reducing initial errors and improving session quality.
+
+## Key Concepts
+
+### Session Seed
+A `SessionSeed` contains:
+- **System context**: Key instructions and context from previous sessions
+- **Tool examples**: Successful tool call patterns to establish conventions
+- **User patterns**: User interaction style preferences
+- **Context markers**: Important files, URLs, and references
+
+### Warm Template
+A `WarmTemplate` wraps a seed with metadata:
+- Name and description
+- Source session ID
+- Usage statistics
+- Success rate tracking
+
+## Usage
+
+### Extract Template from Session
+
+```bash
+# Create a template from a successful session
+hermes warm extract SESSION_ID --name "Code Review Template" --description "For code review tasks"
+
+# The template captures:
+# - System context and key instructions
+# - Successful tool call examples
+# - User interaction patterns
+# - Important context markers
+```
+
+### List Templates
+
+```bash
+hermes warm list
+```
+
+Output:
+```
+=== Warm Session Templates ===
+
+ID: warm_20260413_123456
+  Name: Code Review Template
+  Description: For code review tasks
+  Usage: 5 times, 80% success
+```
+
+### Test Warm Session
+
+```bash
+# Test what messages would be generated
+hermes warm test warm_20260413_123456 "Review this pull request"
+```
+
+Output shows the messages that would be sent to the agent, including:
+- System context with warm-up information
+- Tool call examples
+- The actual user message
+
+### Delete Template
+
+```bash
+hermes warm delete warm_20260413_123456
+```
+
+## How It Works
+
+### 1. Extraction Phase
+When you extract a template:
+1. System messages provide base context
+2. First 10 user messages establish patterns
+3. Successful tool calls become examples
+4. File paths and URLs become context markers
+
+### 2. Bootstrap Phase
+When creating a warm session:
+1. System context is injected as initial message
+2. Tool examples establish successful patterns
+3. User message follows the warm-up context
+4. Agent starts with established conventions
+
+## Example Workflow
+
+```bash
+# 1. Have a successful session
+# ... work with the agent on a complex task ...
+
+# 2. Extract template from that session
+hermes warm extract abc123 --name "API Integration" --description "REST API work"
+
+# 3. Later, start a new session with warm context
+# The agent will have context about:
+# - Your coding style
+# - Successful tool patterns
+# - Common file paths
+# - Previous instructions
+```
+
+## Benefits
+
+1. **Reduced Initial Errors**: Agent starts with proven patterns
+2. **Consistent Behavior**: Established conventions carry over
+3. **Faster Context**: No need to re-explain preferences
+4. **Quality Tracking**: Success rate shows template effectiveness
+
+## Implementation Details
+
+### Files
+- `tools/warm_session.py`: Core implementation
+- `~/.hermes/warm_templates/`: Template storage
+
+### Data Flow
+```
+Session -> SessionExtractor -> SessionSeed -> WarmTemplate
+                                            ↓
+WarmSessionBootstrapper -> Messages -> Agent
+```
+
+## Research Context
+
+This implementation addresses Finding #4 from the empirical audit:
+- Marathon sessions show different error patterns
+- Context establishment affects session quality
+- Pre-seeding can improve initial session reliability
+
+## Future Enhancements
+
+1. **Automatic Template Creation**: Create templates from high-performing sessions
+2. **Template Sharing**: Export/import templates between installations
+3. **A/B Testing**: Compare warm vs cold session performance
+4. **Smart Selection**: Automatically choose best template for task type

hermes_cli/main.py

@@ -5258,6 +5258,36 @@ For more help on a command:
 
     sessions_parser.set_defaults(func=cmd_sessions)
 
+    
+    # Warm session command
+    warm_parser = subparsers.add_parser(
+        "warm",
+        help="Warm session provisioning",
+        description="Create pre-contextualized sessions from templates"
+    )
+    warm_subparsers = warm_parser.add_subparsers(dest="warm_command")
+    
+    # Extract command
+    warm_extract = warm_subparsers.add_parser("extract", help="Extract template from session")
+    warm_extract.add_argument("session_id", help="Session ID to extract from")
+    warm_extract.add_argument("--name", "-n", required=True, help="Template name")
+    warm_extract.add_argument("--description", "-d", default="", help="Template description")
+    
+    # List command
+    warm_subparsers.add_parser("list", help="List available templates")
+    
+    # Test command
+    warm_test = warm_subparsers.add_parser("test", help="Test warm session creation")
+    warm_test.add_argument("template_id", help="Template ID")
+    warm_test.add_argument("message", help="Test message")
+    
+    # Delete command
+    warm_delete = warm_subparsers.add_parser("delete", help="Delete a template")
+    warm_delete.add_argument("template_id", help="Template ID to delete")
+    
+    warm_parser.set_defaults(func=cmd_warm)
+
+
     # =========================================================================
     # insights command
     # =========================================================================
@@ -5598,3 +5628,44 @@ Examples:
 
 if __name__ == "__main__":
     main()
+
+
+def cmd_warm(args):
+    """Handle warm session commands."""
+    from hermes_cli.colors import Colors, color
+    
+    subcmd = getattr(args, 'warm_command', None)
+    
+    if subcmd is None:
+        print(color("Warm Session Provisioning", Colors.CYAN))
+        print("\nCommands:")
+        print("  hermes warm extract SESSION_ID --name NAME - Extract template from session")
+        print("  hermes warm list                         - List available templates")
+        print("  hermes warm test TEMPLATE_ID MESSAGE      - Test warm session")
+        print("  hermes warm delete TEMPLATE_ID            - Delete a template")
+        return 0
+    
+    try:
+        from tools.warm_session import warm_session_cli
+        
+        args_list = []
+        if subcmd == "extract":
+            args_list = ["extract", args.session_id, "--name", args.name]
+            if args.description:
+                args_list.extend(["--description", args.description])
+        elif subcmd == "list":
+            args_list = ["list"]
+        elif subcmd == "test":
+            args_list = ["test", args.template_id, args.message]
+        elif subcmd == "delete":
+            args_list = ["delete", args.template_id]
+        
+        return warm_session_cli(args_list)
+        
+    except ImportError as e:
+        print(color(f"Error: Cannot import warm_session module: {e}", Colors.RED))
+        return 1
+    except Exception as e:
+        print(color(f"Error: {e}", Colors.RED))
+        return 1
+

scripts/evaluate_qwen35.py

@@ -1,114 +0,0 @@
-#!/usr/bin/env python3
-"""Evaluate Qwen3.5:35B as a local model option -- Issue #288, Epic #281."""
-import json, sys, time
-from dataclasses import dataclass, field
-from typing import Any, Dict
-
-@dataclass
-class ModelSpec:
-    name: str = "Qwen3.5-35B-A3B"
-    ollama_tag: str = "qwen3.5:35b"
-    hf_id: str = "Qwen/Qwen3.5-35B-A3B"
-    architecture: str = "MoE (Mixture of Experts)"
-    total_params: str = "35B"
-    active_params: str = "3B per token"
-    context_length: int = 131072
-    license: str = "Apache 2.0"
-    tool_use_support: bool = True
-    json_mode_support: bool = True
-    function_calling: bool = True
-    quantization_options: Dict[str, int] = field(default_factory=lambda: {
-        "Q8_0": 36, "Q6_K": 28, "Q5_K_M": 24, "Q4_K_M": 20,
-        "Q4_0": 18, "Q3_K_M": 15, "Q2_K": 12,
-    })
-
-FLEET_MODELS = {
-    "qwen3.5:35b (candidate)": {"params_total": "35B", "context": "128K", "local": True, "tool_use": True, "reasoning": "good"},
-    "gemma4 (current local)": {"params_total": "9B", "context": "128K", "local": True, "tool_use": True, "reasoning": "good"},
-    "hermes4:14b (current local)": {"params_total": "14B", "context": "8K", "local": True, "tool_use": True, "reasoning": "good"},
-    "qwen2.5:7b (fleet)": {"params_total": "7B", "context": "32K", "local": True, "tool_use": True, "reasoning": "moderate"},
-    "claude-sonnet-4 (cloud)": {"params_total": "?", "context": "200K", "local": False, "tool_use": True, "reasoning": "excellent"},
-    "mimo-v2-pro (cloud free)": {"params_total": "?", "context": "128K", "local": False, "tool_use": True, "reasoning": "good"},
-}
-
-SECURITY_CRITERIA = [
-    {"criterion": "Data locality", "weight": "CRITICAL", "score": 10, "notes": "All inference local via Ollama. Zero exfiltration."},
-    {"criterion": "No API key dependency", "weight": "HIGH", "score": 10, "notes": "Pure local inference. No external creds needed."},
-    {"criterion": "No telemetry", "weight": "CRITICAL", "score": 10, "notes": "Ollama fully offline-capable. No phone-home."},
-    {"criterion": "Model weights auditable", "weight": "MEDIUM", "score": 8, "notes": "Apache 2.0, HF SHA verification. MoE harder to audit."},
-    {"criterion": "Tool-use safety", "weight": "HIGH", "score": 7, "notes": "Function calling supported, MoE routing less predictable. Benchmark: #502."},
-    {"criterion": "Privacy filter compat", "weight": "HIGH", "score": 9, "notes": "Local = Privacy Filter unnecessary for most queries."},
-    {"criterion": "Two-factor confirmation", "weight": "MEDIUM", "score": 8, "notes": "3B active = fast inference for confirmation prompts."},
-    {"criterion": "Prompt injection resistance", "weight": "HIGH", "score": 6, "notes": "3B active may be weaker. Needs red-team (#324)."},
-]
-
-HARDWARE_PROFILES = {
-    "mac_m2_ultra_192gb": {"name": "Mac Studio M2 Ultra (192GB)", "mem_gb": 192, "fits_q4": True, "fits_q8": True, "rec": "Q6_K", "tok_sec": 40},
-    "mac_m4_pro_48gb": {"name": "Mac Mini M4 Pro (48GB)", "mem_gb": 48, "fits_q4": True, "fits_q8": False, "rec": "Q4_K_M", "tok_sec": 30},
-    "mac_m1_16gb": {"name": "Mac M1 (16GB)", "mem_gb": 16, "fits_q4": False, "fits_q8": False, "rec": None, "tok_sec": None},
-    "rtx_4090_24gb": {"name": "NVIDIA RTX 4090 (24GB)", "mem_gb": 24, "fits_q4": True, "fits_q8": False, "rec": "Q5_K_M", "tok_sec": 50},
-    "rtx_3090_24gb": {"name": "NVIDIA RTX 3090 (24GB)", "mem_gb": 24, "fits_q4": True, "fits_q8": False, "rec": "Q4_K_M", "tok_sec": 35},
-    "runpod_l40s_48gb": {"name": "RunPod L40S (48GB)", "mem_gb": 48, "fits_q4": True, "fits_q8": True, "rec": "Q6_K", "tok_sec": 60},
-}
-
-def check_ollama_status() -> Dict[str, Any]:
-    import subprocess
-    result = {"running": False, "models": [], "qwen35_available": False}
-    try:
-        r = subprocess.run(["curl", "-s", "--max-time", "5", "http://localhost:11434/api/tags"], capture_output=True, text=True, timeout=10)
-        if r.returncode == 0:
-            data = json.loads(r.stdout)
-            result["running"] = True
-            result["models"] = [m["name"] for m in data.get("models", [])]
-            result["qwen35_available"] = any("qwen3.5" in m.lower() for m in result["models"])
-    except Exception as e:
-        result["error"] = str(e)
-    return result
-
-def generate_report() -> str:
-    spec = ModelSpec()
-    ollama = check_ollama_status()
-    lines = ["=" * 72, "Qwen3.5:35B EVALUATION REPORT -- Issue #288", "Epic #281 -- Vitalik Secure LLM Architecture", "=" * 72]
-    lines.append("\n## 1. Model Specification\n")
-    lines.append(f"  Name: {spec.name}  |  Arch: {spec.architecture}")
-    lines.append(f"  Params: {spec.total_params} total, {spec.active_params}  |  Context: {spec.context_length:,} tokens")
-    lines.append(f"  License: {spec.license}  |  Tools: {spec.tool_use_support}  |  JSON: {spec.json_mode_support}")
-    lines.append("\n## 2. VRAM\n")
-    for q, vram in sorted(spec.quantization_options.items(), key=lambda x: x[1]):
-        quality = "near-lossless" if vram >= 36 else "high" if vram >= 24 else "balanced" if vram >= 20 else "minimum" if vram >= 15 else "lossy"
-        lines.append(f"  {q:<10} {vram:>4}GB  {quality}")
-    lines.append("\n## 3. Hardware\n")
-    for hw in HARDWARE_PROFILES.values():
-        lines.append(f"  {hw['name']}  {hw['mem_gb']}GB  Q4:{'YES' if hw['fits_q4'] else 'NO '}  Rec:{hw['rec'] or 'N/A'}  ~{hw['tok_sec'] or 'N/A'} tok/s")
-    lines.append("\n## 4. Security (Vitalik Framework)\n")
-    wm = {"CRITICAL": 3, "HIGH": 2, "MEDIUM": 1}
-    tw = sum(wm[c["weight"]] for c in SECURITY_CRITERIA)
-    ws = sum(c["score"] * wm[c["weight"]] for c in SECURITY_CRITERIA)
-    for c in SECURITY_CRITERIA:
-        lines.append(f"  [{c['weight']:<8}] {c['criterion']}: {c['score']}/10 -- {c['notes']}")
-    avg = ws / tw
-    lines.append(f"\n  Weighted: {avg:.1f}/10  Verdict: {'STRONG' if avg >= 8 else 'ADEQUATE'}")
-    lines.append("\n## 5. Fleet Comparison\n")
-    for name, d in FLEET_MODELS.items():
-        lines.append(f"  {name:<35} {d['params_total']:<6} {d['context']:<6} {'Local' if d['local'] else 'Cloud'}  {d['reasoning']}")
-    lines.append("\n## 6. Ollama\n")
-    lines.append(f"  Running: {'Yes' if ollama['running'] else 'No'}  |  Models: {', '.join(ollama['models']) or 'none'}")
-    lines.append(f"  Qwen3.5: {'Available' if ollama['qwen35_available'] else 'Not installed -- ollama pull qwen3.5:35b'}")
-    lines.append("\n## 7. Recommendation\n")
-    lines.append("  VERDICT: APPROVED for local deployment as privacy-sensitive tier")
-    lines.append("\n  + Perfect data sovereignty, 128K context, Apache 2.0, MoE speed")
-    lines.append("  + Tool use + JSON mode, eliminates Privacy Filter for most queries")
-    lines.append("  - 20GB VRAM at Q4, MoE less predictable, needs red-team testing")
-    lines.append("\n  Follow-up issues filed:")
-    lines.append("    #502: live tool dispatch benchmark")
-    lines.append("    #503: reasoning benchmark vs hermes4:14b")
-    lines.append("    #518: document minimum hardware requirements fleet-wide")
-    lines.append("    #324: prompt injection red-team testing")
-    lines.append("\n  Deployment: ollama pull qwen3.5:35b -> config.yaml privacy_model")
-    return "\n".join(lines)
-
-if __name__ == "__main__":
-    if "--check-ollama" in sys.argv:
-        print(json.dumps(check_ollama_status(), indent=2))
-    else:
-        print(generate_report())

tests/test_evaluate_qwen35.py

@@ -1,50 +0,0 @@
-"""Tests for Qwen3.5:35B evaluation -- Issue #288."""
-import pytest
-from scripts.evaluate_qwen35 import ModelSpec, FLEET_MODELS, SECURITY_CRITERIA, HARDWARE_PROFILES, check_ollama_status, generate_report
-
-class TestModelSpec:
-    def test_fields(self):
-        s = ModelSpec()
-        assert s.name == "Qwen3.5-35B-A3B"
-        assert s.context_length == 131072
-        assert s.license == "Apache 2.0"
-        assert s.tool_use_support is True
-    def test_quant_vram_decreasing(self):
-        s = ModelSpec()
-        items = sorted(s.quantization_options.items(), key=lambda x: x[1])
-        for i in range(1, len(items)):
-            assert items[i][1] >= items[i-1][1]
-
-class TestSecurity:
-    def test_scores(self):
-        for c in SECURITY_CRITERIA:
-            assert 1 <= c["score"] <= 10
-    def test_weighted_avg(self):
-        wm = {"CRITICAL": 3, "HIGH": 2, "MEDIUM": 1}
-        tw = sum(wm[c["weight"]] for c in SECURITY_CRITERIA)
-        ws = sum(c["score"] * wm[c["weight"]] for c in SECURITY_CRITERIA)
-        assert ws / tw >= 7.0
-
-class TestHardware:
-    def test_m2_fits(self):
-        assert HARDWARE_PROFILES["mac_m2_ultra_192gb"]["fits_q4"] is True
-    def test_m1_no(self):
-        assert HARDWARE_PROFILES["mac_m1_16gb"]["fits_q4"] is False
-
-class TestReport:
-    def test_sections(self):
-        r = generate_report()
-        for s in ["Model Specification", "VRAM", "Hardware", "Security", "Fleet", "Recommendation"]:
-            assert s in r
-    def test_approved(self):
-        assert "APPROVED" in generate_report()
-    def test_follow_up_issues_referenced(self):
-        r = generate_report()
-        for issue in ["#502", "#503", "#518", "#324"]:
-            assert issue in r
-
-class TestOllama:
-    def test_returns_dict(self):
-        r = check_ollama_status()
-        assert isinstance(r, dict)
-        assert "running" in r

tools/warm_session.py

@@ -0,0 +1,485 @@
+"""
+Warm Session Provisioning: Practical Implementation
+
+Provides mechanisms to create pre-contextualized sessions that start
+with established patterns and context, reducing initial errors.
+
+Issue: #327
+"""
+
+import json
+import logging
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, List, Optional
+from dataclasses import dataclass, asdict, field
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class SessionSeed:
+    """Seed data for warming up a new session."""
+    system_context: str = ""
+    tool_examples: List[Dict[str, Any]] = field(default_factory=list)
+    user_patterns: Dict[str, Any] = field(default_factory=dict)
+    context_markers: List[str] = field(default_factory=list)
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return asdict(self)
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> 'SessionSeed':
+        return cls(**data)
+
+
+@dataclass  
+class WarmTemplate:
+    """Template for creating warm sessions."""
+    template_id: str
+    name: str
+    description: str
+    seed: SessionSeed
+    created_at: str
+    source_session_id: Optional[str] = None
+    usage_count: int = 0
+    success_rate: float = 0.0
+    
+    def to_dict(self) -> Dict[str, Any]:
+        return {
+            "template_id": self.template_id,
+            "name": self.name,
+            "description": self.description,
+            "seed": self.seed.to_dict(),
+            "created_at": self.created_at,
+            "source_session_id": self.source_session_id,
+            "usage_count": self.usage_count,
+            "success_rate": self.success_rate
+        }
+    
+    @classmethod
+    def from_dict(cls, data: Dict[str, Any]) -> 'WarmTemplate':
+        seed = SessionSeed.from_dict(data.get("seed", {}))
+        return cls(
+            template_id=data["template_id"],
+            name=data["name"],
+            description=data["description"],
+            seed=seed,
+            created_at=data.get("created_at", datetime.now().isoformat()),
+            source_session_id=data.get("source_session_id"),
+            usage_count=data.get("usage_count", 0),
+            success_rate=data.get("success_rate", 0.0)
+        )
+
+
+class SessionExtractor:
+    """Extract seed data from existing sessions."""
+    
+    def __init__(self, session_db=None):
+        self.session_db = session_db
+    
+    def extract_seed(self, session_id: str) -> Optional[SessionSeed]:
+        """Extract seed data from a session."""
+        if not self.session_db:
+            return None
+        
+        try:
+            messages = self.session_db.get_messages(session_id)
+            if not messages:
+                return None
+            
+            # Extract system context
+            system_context = self._extract_system_context(messages)
+            
+            # Extract successful tool examples
+            tool_examples = self._extract_tool_examples(messages)
+            
+            # Extract user patterns
+            user_patterns = self._extract_user_patterns(messages)
+            
+            # Extract context markers
+            context_markers = self._extract_context_markers(messages)
+            
+            return SessionSeed(
+                system_context=system_context,
+                tool_examples=tool_examples,
+                user_patterns=user_patterns,
+                context_markers=context_markers
+            )
+            
+        except Exception as e:
+            logger.error(f"Failed to extract seed: {e}")
+            return None
+    
+    def _extract_system_context(self, messages: List[Dict]) -> str:
+        """Extract useful system context from messages."""
+        context_parts = []
+        
+        # Look for system messages
+        for msg in messages:
+            if msg.get("role") == "system":
+                content = msg.get("content", "")
+                # Take first 500 chars of system context
+                if content:
+                    context_parts.append(content[:500])
+                    break
+        
+        # Extract key user instructions
+        user_instructions = []
+        for msg in messages[:10]:  # First 10 messages
+            if msg.get("role") == "user":
+                content = msg.get("content", "")
+                if len(content) > 50 and "?" not in content[:20]:  # Likely instructions
+                    user_instructions.append(content[:200])
+                    if len(user_instructions) >= 3:
+                        break
+        
+        if user_instructions:
+            context_parts.append("\nKey instructions from session:\n" + "\n".join(f"- {i}" for i in user_instructions))
+        
+        return "\n".join(context_parts)[:1000]
+    
+    def _extract_tool_examples(self, messages: List[Dict]) -> List[Dict[str, Any]]:
+        """Extract successful tool call examples."""
+        examples = []
+        
+        for i, msg in enumerate(messages):
+            if msg.get("role") == "assistant" and msg.get("tool_calls"):
+                # Check if there's a successful result
+                for j in range(i + 1, min(i + 3, len(messages))):
+                    if messages[j].get("role") == "tool":
+                        content = messages[j].get("content", "")
+                        # Check for success indicators
+                        if content and "error" not in content.lower()[:100]:
+                            for tool_call in msg["tool_calls"]:
+                                func = tool_call.get("function", {})
+                                examples.append({
+                                    "tool": func.get("name"),
+                                    "arguments": func.get("arguments", "{}"),
+                                    "result_preview": content[:200]
+                                })
+                                if len(examples) >= 5:
+                                    break
+                        break
+            
+            if len(examples) >= 5:
+                break
+        
+        return examples
+    
+    def _extract_user_patterns(self, messages: List[Dict]) -> Dict[str, Any]:
+        """Extract user interaction patterns."""
+        user_messages = [m for m in messages if m.get("role") == "user"]
+        
+        if not user_messages:
+            return {}
+        
+        # Calculate patterns
+        lengths = [len(m.get("content", "")) for m in user_messages]
+        avg_length = sum(lengths) / len(lengths)
+        
+        # Count question types
+        questions = sum(1 for m in user_messages if "?" in m.get("content", ""))
+        commands = sum(1 for m in user_messages if m.get("content", "").startswith(("/", "!")))
+        
+        return {
+            "message_count": len(user_messages),
+            "avg_length": avg_length,
+            "question_ratio": questions / len(user_messages),
+            "command_ratio": commands / len(user_messages),
+            "preferred_style": "command" if commands > questions else "conversational"
+        }
+    
+    def _extract_context_markers(self, messages: List[Dict]) -> List[str]:
+        """Extract important context markers."""
+        markers = set()
+        
+        for msg in messages:
+            content = msg.get("content", "")
+            
+            # File paths
+            import re
+            paths = re.findall(r'[\w/\.]+\.[\w]+', content)
+            markers.update(p for p in paths if len(p) < 50)
+            
+            # URLs
+            urls = re.findall(r'https?://[^\s]+', content)
+            markers.update(u[:80] for u in urls[:3])
+            
+            if len(markers) > 20:
+                break
+        
+        return list(markers)[:20]
+
+
+class WarmSessionManager:
+    """Manage warm session templates."""
+    
+    def __init__(self, template_dir: Path = None):
+        self.template_dir = template_dir or Path.home() / ".hermes" / "warm_templates"
+        self.template_dir.mkdir(parents=True, exist_ok=True)
+    
+    def save_template(self, template: WarmTemplate) -> Path:
+        """Save a warm template."""
+        path = self.template_dir / f"{template.template_id}.json"
+        with open(path, 'w') as f:
+            json.dump(template.to_dict(), f, indent=2)
+        return path
+    
+    def load_template(self, template_id: str) -> Optional[WarmTemplate]:
+        """Load a warm template."""
+        path = self.template_dir / f"{template_id}.json"
+        if not path.exists():
+            return None
+        
+        try:
+            with open(path, 'r') as f:
+                data = json.load(f)
+            return WarmTemplate.from_dict(data)
+        except Exception as e:
+            logger.error(f"Failed to load template: {e}")
+            return None
+    
+    def list_templates(self) -> List[Dict[str, Any]]:
+        """List all templates."""
+        templates = []
+        for path in self.template_dir.glob("*.json"):
+            try:
+                with open(path, 'r') as f:
+                    data = json.load(f)
+                templates.append({
+                    "template_id": data.get("template_id"),
+                    "name": data.get("name"),
+                    "description": data.get("description"),
+                    "usage_count": data.get("usage_count", 0),
+                    "success_rate": data.get("success_rate", 0.0)
+                })
+            except:
+                pass
+        return templates
+    
+    def delete_template(self, template_id: str) -> bool:
+        """Delete a template."""
+        path = self.template_dir / f"{template_id}.json"
+        if path.exists():
+            path.unlink()
+            return True
+        return False
+
+
+class WarmSessionBootstrapper:
+    """Bootstrap warm sessions from templates."""
+    
+    def __init__(self, manager: WarmSessionManager = None):
+        self.manager = manager or WarmSessionManager()
+    
+    def prepare_messages(
+        self,
+        template: WarmTemplate,
+        user_message: str,
+        include_examples: bool = True
+    ) -> List[Dict[str, Any]]:
+        """Prepare messages for a warm session."""
+        messages = []
+        
+        # Add warm context as system message
+        warm_context = self._build_warm_context(template.seed)
+        if warm_context:
+            messages.append({
+                "role": "system",
+                "content": warm_context
+            })
+        
+        # Add tool examples if requested
+        if include_examples and template.seed.tool_examples:
+            example_messages = self._create_example_messages(template.seed.tool_examples)
+            messages.extend(example_messages)
+        
+        # Add the actual user message
+        messages.append({
+            "role": "user",
+            "content": user_message
+        })
+        
+        return messages
+    
+    def _build_warm_context(self, seed: SessionSeed) -> str:
+        """Build warm context from seed."""
+        parts = []
+        
+        if seed.system_context:
+            parts.append(seed.system_context)
+        
+        if seed.context_markers:
+            parts.append("\nKnown context: " + ", ".join(seed.context_markers[:10]))
+        
+        if seed.user_patterns:
+            style = seed.user_patterns.get("preferred_style", "balanced")
+            parts.append(f"\nUser prefers {style} interactions.")
+        
+        return "\n".join(parts)[:1500]
+    
+    def _create_example_messages(self, examples: List[Dict]) -> List[Dict]:
+        """Create example messages from tool examples."""
+        messages = []
+        
+        for i, ex in enumerate(examples[:3]):  # Limit to 3 examples
+            # User request
+            messages.append({
+                "role": "user",
+                "content": f"[Example {i+1}] Use {ex['tool']}"
+            })
+            
+            # Assistant with tool call
+            messages.append({
+                "role": "assistant",
+                "content": f"I'll use {ex['tool']}.",
+                "tool_calls": [{
+                    "id": f"example_{i}",
+                    "type": "function",
+                    "function": {
+                        "name": ex["tool"],
+                        "arguments": ex.get("arguments", "{}")
+                    }
+                }]
+            })
+            
+            # Tool result
+            messages.append({
+                "role": "tool",
+                "tool_call_id": f"example_{i}",
+                "content": ex.get("result_preview", "Success")
+            })
+        
+        return messages
+
+
+# CLI Functions
+def warm_session_cli(args: List[str]) -> int:
+    """CLI interface for warm session management."""
+    import argparse
+    
+    parser = argparse.ArgumentParser(description="Warm session provisioning")
+    subparsers = parser.add_subparsers(dest="command")
+    
+    # Extract command
+    extract_parser = subparsers.add_parser("extract", help="Extract template from session")
+    extract_parser.add_argument("session_id", help="Session ID to extract from")
+    extract_parser.add_argument("--name", "-n", required=True, help="Template name")
+    extract_parser.add_argument("--description", "-d", default="", help="Template description")
+    
+    # List command
+    subparsers.add_parser("list", help="List available templates")
+    
+    # Test command
+    test_parser = subparsers.add_parser("test", help="Test warm session creation")
+    test_parser.add_argument("template_id", help="Template ID")
+    test_parser.add_argument("message", help="Test message")
+    
+    # Delete command
+    delete_parser = subparsers.add_parser("delete", help="Delete a template")
+    delete_parser.add_argument("template_id", help="Template ID to delete")
+    
+    parsed = parser.parse_args(args)
+    
+    if not parsed.command:
+        parser.print_help()
+        return 1
+    
+    manager = WarmSessionManager()
+    
+    if parsed.command == "extract":
+        try:
+            from hermes_state import SessionDB
+            session_db = SessionDB()
+        except ImportError:
+            print("Error: Cannot import SessionDB")
+            return 1
+        
+        extractor = SessionExtractor(session_db)
+        seed = extractor.extract_seed(parsed.session_id)
+        
+        if not seed:
+            print(f"Failed to extract seed from session {parsed.session_id}")
+            return 1
+        
+        template = WarmTemplate(
+            template_id=f"warm_{datetime.now().strftime('%Y%m%d_%H%M%S')}",
+            name=parsed.name,
+            description=parsed.description,
+            seed=seed,
+            created_at=datetime.now().isoformat(),
+            source_session_id=parsed.session_id
+        )
+        
+        path = manager.save_template(template)
+        print(f"Created template: {template.template_id}")
+        print(f"Saved to: {path}")
+        print(f"Tool examples: {len(seed.tool_examples)}")
+        print(f"Context markers: {len(seed.context_markers)}")
+        return 0
+    
+    elif parsed.command == "list":
+        templates = manager.list_templates()
+        
+        if not templates:
+            print("No templates found.")
+            return 0
+        
+        print("\n=== Warm Session Templates ===\n")
+        for t in templates:
+            print(f"ID: {t['template_id']}")
+            print(f"  Name: {t['name']}")
+            print(f"  Description: {t['description']}")
+            print(f"  Usage: {t['usage_count']} times, {t['success_rate']:.0%} success")
+            print()
+        
+        return 0
+    
+    elif parsed.command == "test":
+        template = manager.load_template(parsed.template_id)
+        
+        if not template:
+            print(f"Template {parsed.template_id} not found")
+            return 1
+        
+        bootstrapper = WarmSessionBootstrapper(manager)
+        messages = bootstrapper.prepare_messages(template, parsed.message)
+        
+        print(f"\n=== Warm Session Test: {template.name} ===\n")
+        print(f"Generated {len(messages)} messages:\n")
+        
+        for i, msg in enumerate(messages):
+            role = msg.get("role", "unknown")
+            content = msg.get("content", "")
+            
+            if role == "system":
+                print(f"[System Context] ({len(content)} chars)")
+                print(content[:200] + "..." if len(content) > 200 else content)
+            elif role == "user":
+                print(f"\n[User]: {content}")
+            elif role == "assistant":
+                print(f"[Assistant]: {content}")
+                if msg.get("tool_calls"):
+                    for tc in msg["tool_calls"]:
+                        func = tc.get("function", {})
+                        print(f"  -> {func.get('name')}({func.get('arguments', '{}')[:50]})")
+            elif role == "tool":
+                print(f"  [Result]: {content[:100]}...")
+        
+        return 0
+    
+    elif parsed.command == "delete":
+        if manager.delete_template(parsed.template_id):
+            print(f"Deleted template: {parsed.template_id}")
+            return 0
+        else:
+            print(f"Template {parsed.template_id} not found")
+            return 1
+    
+    return 1
+
+
+if __name__ == "__main__":
+    import sys
+    sys.exit(warm_session_cli(sys.argv[1:]))