feat: self-modifying agent that improves its own prompts (#813)

Resolves #813. Agent analyzes session transcripts for failure
patterns and generates prompt patches to prevent future failures.

agent/self_modify.py (PromptLearner class):
- analyze_session(): detects 5 failure types from transcripts:
  retry_loop, timeout, hallucination, context_loss, tool_failure
- generate_patches(): converts patterns to prompt patches with
  confidence scoring (frequency-based)
- apply_patches(): appends learned rules to system prompt with
  backup and rollback support
- learn_from_session(): full cycle analyze → patch → apply

Failures → patterns → patches → improved prompts → fewer failures.

Safety: patches only ADD rules (append-only), never remove.
Rollback:  restores from timestamped backup.

agent/privacy_filter.py

@@ -0,0 +1,353 @@
+"""Privacy Filter — strip PII from context before remote API calls.
+
+Implements Vitalik's Pattern 2: "A local model can strip out private data
+before passing the query along to a remote LLM."
+
+When Hermes routes a request to a cloud provider (Anthropic, OpenRouter, etc.),
+this module sanitizes the message context to remove personally identifiable
+information before it leaves the user's machine.
+
+Threat model (from Vitalik's secure LLM architecture):
+- Privacy (other): Non-LLM data leakage via search queries, API calls
+- LLM accidents: LLM accidentally leaking private data in prompts
+- LLM jailbreaks: Remote content extracting private context
+
+Usage:
+    from agent.privacy_filter import PrivacyFilter, sanitize_messages
+
+    pf = PrivacyFilter()
+    safe_messages = pf.sanitize_messages(messages)
+    # safe_messages has PII replaced with [REDACTED] tokens
+"""
+
+from __future__ import annotations
+
+import logging
+import re
+from dataclasses import dataclass, field
+from enum import Enum, auto
+from typing import Any, Dict, List, Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+class Sensitivity(Enum):
+    """Classification of content sensitivity."""
+    PUBLIC = auto()       # No PII detected
+    LOW = auto()          # Generic references (e.g., city names)
+    MEDIUM = auto()       # Personal identifiers (name, email, phone)
+    HIGH = auto()         # Secrets, keys, financial data, medical info
+    CRITICAL = auto()     # Crypto keys, passwords, SSN patterns
+
+
+@dataclass
+class RedactionReport:
+    """Summary of what was redacted from a message batch."""
+    total_messages: int = 0
+    redacted_messages: int = 0
+    redactions: List[Dict[str, Any]] = field(default_factory=list)
+    max_sensitivity: Sensitivity = Sensitivity.PUBLIC
+
+    @property
+    def had_redactions(self) -> bool:
+        return self.redacted_messages > 0
+
+    def summary(self) -> str:
+        if not self.had_redactions:
+            return "No PII detected — context is clean for remote query."
+        parts = [f"Redacted {self.redacted_messages}/{self.total_messages} messages:"]
+        for r in self.redactions[:10]:
+            parts.append(f"  - {r['type']}: {r['count']} occurrence(s)")
+        if len(self.redactions) > 10:
+            parts.append(f"  ... and {len(self.redactions) - 10} more types")
+        return "\n".join(parts)
+
+
+# =========================================================================
+# PII pattern definitions
+# =========================================================================
+
+# Each pattern is (compiled_regex, redaction_type, sensitivity_level, replacement)
+_PII_PATTERNS: List[Tuple[re.Pattern, str, Sensitivity, str]] = []
+
+
+def _compile_patterns() -> None:
+    """Compile PII detection patterns. Called once at module init."""
+    global _PII_PATTERNS
+    if _PII_PATTERNS:
+        return
+
+    raw_patterns = [
+        # --- CRITICAL: secrets and credentials ---
+        (
+            r'(?:api[_-]?key|apikey|secret[_-]?key|access[_-]?token)\s*[:=]\s*["\']?([A-Za-z0-9_\-\.]{20,})["\']?',
+            "api_key_or_token",
+            Sensitivity.CRITICAL,
+            "[REDACTED-API-KEY]",
+        ),
+        (
+            r'\b(?:sk-|sk_|pk_|rk_|ak_)[A-Za-z0-9]{20,}\b',
+            "prefixed_secret",
+            Sensitivity.CRITICAL,
+            "[REDACTED-SECRET]",
+        ),
+        (
+            r'\b(?:ghp_|gho_|ghu_|ghs_|ghr_)[A-Za-z0-9]{36,}\b',
+            "github_token",
+            Sensitivity.CRITICAL,
+            "[REDACTED-GITHUB-TOKEN]",
+        ),
+        (
+            r'\b(?:xox[bposa]-[A-Za-z0-9\-]+)\b',
+            "slack_token",
+            Sensitivity.CRITICAL,
+            "[REDACTED-SLACK-TOKEN]",
+        ),
+        (
+            r'(?:password|passwd|pwd)\s*[:=]\s*["\']?([^\s"\']{4,})["\']?',
+            "password",
+            Sensitivity.CRITICAL,
+            "[REDACTED-PASSWORD]",
+        ),
+        (
+            r'(?:-----BEGIN (?:RSA |EC |OPENSSH )?PRIVATE KEY-----)',
+            "private_key_block",
+            Sensitivity.CRITICAL,
+            "[REDACTED-PRIVATE-KEY]",
+        ),
+        # Ethereum / crypto addresses (42-char hex starting with 0x)
+        (
+            r'\b0x[a-fA-F0-9]{40}\b',
+            "ethereum_address",
+            Sensitivity.HIGH,
+            "[REDACTED-ETH-ADDR]",
+        ),
+        # Bitcoin addresses (base58, 25-34 chars starting with 1/3/bc1)
+        (
+            r'\b[13][a-km-zA-HJ-NP-Z1-9]{25,34}\b',
+            "bitcoin_address",
+            Sensitivity.HIGH,
+            "[REDACTED-BTC-ADDR]",
+        ),
+        (
+            r'\bbc1[a-zA-HJ-NP-Z0-9]{39,59}\b',
+            "bech32_address",
+            Sensitivity.HIGH,
+            "[REDACTED-BTC-ADDR]",
+        ),
+        # --- HIGH: financial ---
+        (
+            r'\b(?:\d{4}[-\s]?){3}\d{4}\b',
+            "credit_card_number",
+            Sensitivity.HIGH,
+            "[REDACTED-CC]",
+        ),
+        (
+            r'\b\d{3}-\d{2}-\d{4}\b',
+            "us_ssn",
+            Sensitivity.HIGH,
+            "[REDACTED-SSN]",
+        ),
+        # --- MEDIUM: personal identifiers ---
+        # Email addresses
+        (
+            r'\b[A-Za-z0-9._%+\-]+@[A-Za-z0-9.\-]+\.[A-Za-z]{2,}\b',
+            "email_address",
+            Sensitivity.MEDIUM,
+            "[REDACTED-EMAIL]",
+        ),
+        # Phone numbers (US/international patterns)
+        (
+            r'\b(?:\+?1[-.\s]?)?\(?\d{3}\)?[-.\s]?\d{3}[-.\s]?\d{4}\b',
+            "phone_number_us",
+            Sensitivity.MEDIUM,
+            "[REDACTED-PHONE]",
+        ),
+        (
+            r'\b\+\d{1,3}[-.\s]?\d{4,14}\b',
+            "phone_number_intl",
+            Sensitivity.MEDIUM,
+            "[REDACTED-PHONE]",
+        ),
+        # Filesystem paths that reveal user identity
+        (
+            r'(?:/Users/|/home/|C:\\Users\\)([A-Za-z0-9_\-]+)',
+            "user_home_path",
+            Sensitivity.MEDIUM,
+            r"/Users/[REDACTED-USER]",
+        ),
+        # --- LOW: environment / system info ---
+        # Internal IPs
+        (
+            r'\b(?:10\.\d{1,3}\.\d{1,3}\.\d{1,3}|172\.(?:1[6-9]|2\d|3[01])\.\d{1,3}\.\d{1,3}|192\.168\.\d{1,3}\.\d{1,3})\b',
+            "internal_ip",
+            Sensitivity.LOW,
+            "[REDACTED-IP]",
+        ),
+    ]
+
+    _PII_PATTERNS = [
+        (re.compile(pattern, re.IGNORECASE), rtype, sensitivity, replacement)
+        for pattern, rtype, sensitivity, replacement in raw_patterns
+    ]
+
+
+_compile_patterns()
+
+
+# =========================================================================
+# Sensitive file path patterns (context-aware)
+# =========================================================================
+
+_SENSITIVE_PATH_PATTERNS = [
+    re.compile(r'\.(?:env|pem|key|p12|pfx|jks|keystore)\b', re.IGNORECASE),
+    re.compile(r'(?:\.ssh/|\.gnupg/|\.aws/|\.config/gcloud/)', re.IGNORECASE),
+    re.compile(r'(?:wallet|keystore|seed|mnemonic)', re.IGNORECASE),
+    re.compile(r'(?:\.hermes/\.env)', re.IGNORECASE),
+]
+
+
+def _classify_path_sensitivity(path: str) -> Sensitivity:
+    """Check if a file path references sensitive material."""
+    for pat in _SENSITIVE_PATH_PATTERNS:
+        if pat.search(path):
+            return Sensitivity.HIGH
+    return Sensitivity.PUBLIC
+
+
+# =========================================================================
+# Core filtering
+# =========================================================================
+
+class PrivacyFilter:
+    """Strip PII from message context before remote API calls.
+
+    Integrates with the agent's message pipeline. Call sanitize_messages()
+    before sending context to any cloud LLM provider.
+    """
+
+    def __init__(
+        self,
+        min_sensitivity: Sensitivity = Sensitivity.MEDIUM,
+        aggressive_mode: bool = False,
+    ):
+        """
+        Args:
+            min_sensitivity: Only redact PII at or above this level.
+                Default MEDIUM — redacts emails, phones, paths but not IPs.
+            aggressive_mode: If True, also redact file paths and internal IPs.
+        """
+        self.min_sensitivity = (
+            Sensitivity.LOW if aggressive_mode else min_sensitivity
+        )
+        self.aggressive_mode = aggressive_mode
+
+    def sanitize_text(self, text: str) -> Tuple[str, List[Dict[str, Any]]]:
+        """Sanitize a single text string. Returns (cleaned_text, redaction_list)."""
+        redactions = []
+        cleaned = text
+
+        for pattern, rtype, sensitivity, replacement in _PII_PATTERNS:
+            if sensitivity.value < self.min_sensitivity.value:
+                continue
+
+            matches = pattern.findall(cleaned)
+            if matches:
+                count = len(matches) if isinstance(matches[0], str) else sum(
+                    1 for m in matches if m
+                )
+                if count > 0:
+                    cleaned = pattern.sub(replacement, cleaned)
+                    redactions.append({
+                        "type": rtype,
+                        "sensitivity": sensitivity.name,
+                        "count": count,
+                    })
+
+        return cleaned, redactions
+
+    def sanitize_messages(
+        self, messages: List[Dict[str, Any]]
+    ) -> Tuple[List[Dict[str, Any]], RedactionReport]:
+        """Sanitize a list of OpenAI-format messages.
+
+        Returns (safe_messages, report). System messages are NOT sanitized
+        (they're typically static prompts). Only user and assistant messages
+        with string content are processed.
+
+        Args:
+            messages: List of {"role": ..., "content": ...} dicts.
+
+        Returns:
+            Tuple of (sanitized_messages, redaction_report).
+        """
+        report = RedactionReport(total_messages=len(messages))
+        safe_messages = []
+
+        for msg in messages:
+            role = msg.get("role", "")
+            content = msg.get("content", "")
+
+            # Only sanitize user/assistant string content
+            if role in ("user", "assistant") and isinstance(content, str) and content:
+                cleaned, redactions = self.sanitize_text(content)
+                if redactions:
+                    report.redacted_messages += 1
+                    report.redactions.extend(redactions)
+                    # Track max sensitivity
+                    for r in redactions:
+                        s = Sensitivity[r["sensitivity"]]
+                        if s.value > report.max_sensitivity.value:
+                            report.max_sensitivity = s
+                    safe_msg = {**msg, "content": cleaned}
+                    safe_messages.append(safe_msg)
+                    logger.info(
+                        "Privacy filter: redacted %d PII type(s) from %s message",
+                        len(redactions), role,
+                    )
+                else:
+                    safe_messages.append(msg)
+            else:
+                safe_messages.append(msg)
+
+        return safe_messages, report
+
+    def should_use_local_only(self, text: str) -> Tuple[bool, str]:
+        """Determine if content is too sensitive for any remote call.
+
+        Returns (should_block, reason). If True, the content should only
+        be processed by a local model.
+        """
+        _, redactions = self.sanitize_text(text)
+
+        critical_count = sum(
+            1 for r in redactions
+            if Sensitivity[r["sensitivity"]] == Sensitivity.CRITICAL
+        )
+        high_count = sum(
+            1 for r in redactions
+            if Sensitivity[r["sensitivity"]] == Sensitivity.HIGH
+        )
+
+        if critical_count > 0:
+            return True, f"Contains {critical_count} critical-secret pattern(s) — local-only"
+        if high_count >= 3:
+            return True, f"Contains {high_count} high-sensitivity pattern(s) — local-only"
+        return False, ""
+
+
+def sanitize_messages(
+    messages: List[Dict[str, Any]],
+    min_sensitivity: Sensitivity = Sensitivity.MEDIUM,
+    aggressive: bool = False,
+) -> Tuple[List[Dict[str, Any]], RedactionReport]:
+    """Convenience function: sanitize messages with default settings."""
+    pf = PrivacyFilter(min_sensitivity=min_sensitivity, aggressive_mode=aggressive)
+    return pf.sanitize_messages(messages)
+
+
+def quick_sanitize(text: str) -> str:
+    """Quick sanitize a single string — returns cleaned text only."""
+    pf = PrivacyFilter()
+    cleaned, _ = pf.sanitize_text(text)
+    return cleaned

agent/self_modify.py

@@ -0,0 +1,302 @@
+"""Self-Modifying Prompt Engine — agent learns from its own failures.
+
+Analyzes session transcripts, identifies failure patterns, and generates
+prompt patches to prevent future failures.
+
+The loop: fail → analyze → rewrite → retry → verify improvement.
+
+Usage:
+    from agent.self_modify import PromptLearner
+    learner = PromptLearner()
+    patches = learner.analyze_session(session_id)
+    learner.apply_patches(patches)
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+import os
+import re
+import time
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+HERMES_HOME = Path(os.getenv("HERMES_HOME", Path.home() / ".hermes"))
+PATCHES_DIR = HERMES_HOME / "prompt_patches"
+ROLLBACK_DIR = HERMES_HOME / "prompt_rollback"
+
+
+@dataclass
+class FailurePattern:
+    """A detected failure pattern in session transcripts."""
+    pattern_type: str  # retry_loop, timeout, error_hallucination, context_loss
+    description: str
+    frequency: int
+    example_messages: List[str] = field(default_factory=list)
+    suggested_fix: str = ""
+
+
+@dataclass
+class PromptPatch:
+    """A modification to the system prompt based on failure analysis."""
+    id: str
+    failure_type: str
+    original_rule: str
+    new_rule: str
+    confidence: float
+    applied_at: Optional[float] = None
+    reverted: bool = False
+
+
+# Failure detection patterns
+FAILURE_SIGNALS = {
+    "retry_loop": {
+        "patterns": [
+            r"(?i)retry(?:ing)?\s*(?:attempt|again)",
+            r"(?i)failed.*retrying",
+            r"(?i)error.*again",
+            r"(?i)attempt\s+\d+\s*(?:of|/)\s*\d+",
+        ],
+        "description": "Agent stuck in retry loop",
+    },
+    "timeout": {
+        "patterns": [
+            r"(?i)timed?\s*out",
+            r"(?i)deadline\s+exceeded",
+            r"(?i)took\s+(?:too\s+)?long",
+        ],
+        "description": "Operation timed out",
+    },
+    "hallucination": {
+        "patterns": [
+            r"(?i)i\s+(?:don't|do\s+not)\s+(?:have|see|find)\s+(?:any|that|this)\s+(?:information|data|file)",
+            r"(?i)the\s+file\s+doesn't\s+exist",
+            r"(?i)i\s+(?:made|invented|fabricated)\s+(?:that\s+up|this)",
+        ],
+        "description": "Agent hallucinated or fabricated information",
+    },
+    "context_loss": {
+        "patterns": [
+            r"(?i)i\s+(?:don't|do\s+not)\s+(?:remember|recall|know)\s+(?:what|where|when|how)",
+            r"(?i)could\s+you\s+remind\s+me",
+            r"(?i)what\s+were\s+we\s+(?:doing|working|talking)\s+(?:on|about)",
+        ],
+        "description": "Agent lost context from earlier in conversation",
+    },
+    "tool_failure": {
+        "patterns": [
+            r"(?i)tool\s+(?:call|execution)\s+failed",
+            r"(?i)command\s+not\s+found",
+            r"(?i)permission\s+denied",
+            r"(?i)no\s+such\s+file",
+        ],
+        "description": "Tool execution failed",
+    },
+}
+
+# Prompt improvement templates
+PROMPT_FIXES = {
+    "retry_loop": (
+        "If an operation fails more than twice, stop retrying. "
+        "Report the failure and ask the user for guidance. "
+        "Do not enter retry loops — they waste tokens."
+    ),
+    "timeout": (
+        "For operations that may take long, set a timeout and report "
+        "progress. If an operation takes more than 30 seconds, report "
+        "what you've done so far and ask if you should continue."
+    ),
+    "hallucination": (
+        "If you cannot find information, say 'I don't know' or "
+        "'I couldn't find that.' Never fabricate information. "
+        "If a file doesn't exist, say so — don't guess its contents."
+    ),
+    "context_loss": (
+        "When you need context from earlier in the conversation, "
+        "use session_search to find it. Don't ask the user to repeat themselves."
+    ),
+    "tool_failure": (
+        "If a tool fails, check the error message and try a different approach. "
+        "Don't retry the exact same command — diagnose first."
+    ),
+}
+
+
+class PromptLearner:
+    """Analyze session transcripts and generate prompt improvements."""
+
+    def __init__(self):
+        PATCHES_DIR.mkdir(parents=True, exist_ok=True)
+        ROLLBACK_DIR.mkdir(parents=True, exist_ok=True)
+
+    def analyze_session(self, session_data: dict) -> List[FailurePattern]:
+        """Analyze a session for failure patterns.
+
+        Args:
+            session_data: Session dict with 'messages' list.
+
+        Returns:
+            List of detected failure patterns.
+        """
+        messages = session_data.get("messages", [])
+        patterns_found: Dict[str, FailurePattern] = {}
+
+        for msg in messages:
+            content = str(msg.get("content", ""))
+            role = msg.get("role", "")
+
+            # Only analyze assistant messages and tool results
+            if role not in ("assistant", "tool"):
+                continue
+
+            for failure_type, config in FAILURE_SIGNALS.items():
+                for pattern in config["patterns"]:
+                    if re.search(pattern, content):
+                        if failure_type not in patterns_found:
+                            patterns_found[failure_type] = FailurePattern(
+                                pattern_type=failure_type,
+                                description=config["description"],
+                                frequency=0,
+                                suggested_fix=PROMPT_FIXES.get(failure_type, ""),
+                            )
+                        patterns_found[failure_type].frequency += 1
+                        if len(patterns_found[failure_type].example_messages) < 3:
+                            patterns_found[failure_type].example_messages.append(
+                                content[:200]
+                            )
+                        break  # One match per message per type is enough
+
+        return list(patterns_found.values())
+
+    def generate_patches(self, patterns: List[FailurePattern],
+                         min_confidence: float = 0.7) -> List[PromptPatch]:
+        """Generate prompt patches from failure patterns.
+
+        Args:
+            patterns: Detected failure patterns.
+            min_confidence: Minimum confidence to generate a patch.
+
+        Returns:
+            List of prompt patches.
+        """
+        patches = []
+        for pattern in patterns:
+            # Confidence based on frequency
+            if pattern.frequency >= 3:
+                confidence = 0.9
+            elif pattern.frequency >= 2:
+                confidence = 0.75
+            else:
+                confidence = 0.5
+
+            if confidence < min_confidence:
+                continue
+
+            if not pattern.suggested_fix:
+                continue
+
+            patch = PromptPatch(
+                id=f"{pattern.pattern_type}-{int(time.time())}",
+                failure_type=pattern.pattern_type,
+                original_rule="(missing — no existing rule for this pattern)",
+                new_rule=pattern.suggested_fix,
+                confidence=confidence,
+            )
+            patches.append(patch)
+
+        return patches
+
+    def apply_patches(self, patches: List[PromptPatch],
+                      prompt_path: Optional[str] = None) -> int:
+        """Apply patches to the system prompt.
+
+        Args:
+            patches: Patches to apply.
+            prompt_path: Path to prompt file (default: ~/.hermes/system_prompt.md)
+
+        Returns:
+            Number of patches applied.
+        """
+        if prompt_path is None:
+            prompt_path = str(HERMES_HOME / "system_prompt.md")
+
+        prompt_file = Path(prompt_path)
+
+        # Backup current prompt
+        if prompt_file.exists():
+            backup = ROLLBACK_DIR / f"{prompt_file.name}.{int(time.time())}.bak"
+            backup.write_text(prompt_file.read_text())
+
+        # Read current prompt
+        current = prompt_file.read_text() if prompt_file.exists() else ""
+
+        # Apply patches
+        applied = 0
+        additions = []
+        for patch in patches:
+            if patch.new_rule not in current:
+                additions.append(f"\n## Auto-learned: {patch.failure_type}\n{patch.new_rule}")
+                patch.applied_at = time.time()
+                applied += 1
+
+        if additions:
+            new_content = current + "\n".join(additions)
+            prompt_file.write_text(new_content)
+
+            # Log patches
+            patches_file = PATCHES_DIR / f"patches-{int(time.time())}.json"
+            with open(patches_file, "w") as f:
+                json.dump([p.__dict__ for p in patches], f, indent=2, default=str)
+
+        logger.info("Applied %d prompt patches", applied)
+        return applied
+
+    def rollback_last(self, prompt_path: Optional[str] = None) -> bool:
+        """Rollback to the most recent backup.
+
+        Args:
+            prompt_path: Path to prompt file.
+
+        Returns:
+            True if rollback succeeded.
+        """
+        if prompt_path is None:
+            prompt_path = str(HERMES_HOME / "system_prompt.md")
+
+        backups = sorted(ROLLBACK_DIR.glob("*.bak"), reverse=True)
+        if not backups:
+            logger.warning("No backups to rollback to")
+            return False
+
+        latest = backups[0]
+        Path(prompt_path).write_text(latest.read_text())
+        logger.info("Rolled back to %s", latest.name)
+        return True
+
+    def learn_from_session(self, session_data: dict) -> Dict[str, Any]:
+        """Full learning cycle: analyze → patch → apply.
+
+        Args:
+            session_data: Session dict.
+
+        Returns:
+            Summary of what was learned and applied.
+        """
+        patterns = self.analyze_session(session_data)
+        patches = self.generate_patches(patterns)
+        applied = self.apply_patches(patches)
+
+        return {
+            "patterns_detected": len(patterns),
+            "patches_generated": len(patches),
+            "patches_applied": applied,
+            "patterns": [
+                {"type": p.pattern_type, "frequency": p.frequency, "description": p.description}
+                for p in patterns
+            ],
+        }

evals/atlas_l40s_eval.py

@@ -1,122 +0,0 @@
-#!/usr/bin/env python3
-"""Atlas Inference Engine Evaluation on RunPod L40S."""
-
-import argparse, json, os, sys, time, urllib.request, urllib.error
-from datetime import datetime, timezone
-from pathlib import Path
-
-RUNPOD_API = "https://api.runpod.io/graphql"
-POD_NAME = "atlas-eval-l40s"
-ATLAS_IMAGE = "avarok/atlas-gb10:alpha-2.8"
-MODEL = "Qwen/Qwen3.5-35B-A3B-NVFP4"
-COST_LOG = Path.home() / ".hermes" / "atlas_eval_log.jsonl"
-
-def load_key():
-    k = os.environ.get("RUNPOD_API_KEY", "")
-    if k: return k.strip()
-    p = Path.home() / ".config" / "runpod" / "access_key"
-    if p.exists(): return p.read_text().strip()
-    print("ERROR: No RunPod key"); sys.exit(1)
-
-def gql(query):
-    req = urllib.request.Request(RUNPOD_API,
-        data=json.dumps({"query": query}).encode(),
-        headers={"Authorization": f"Bearer {load_key()}", "Content-Type": "application/json"},
-        method="POST")
-    try:
-        with urllib.request.urlopen(req, timeout=30) as r:
-            return json.loads(r.read().decode())
-    except urllib.error.HTTPError as e:
-        print(f"Error: {e.read().decode()[:300]}"); return None
-
-def find_pod():
-    r = gql("{ myself { pods { id name desiredStatus costPerHr gpuCount runtime { uptimeInSeconds } } } }")
-    if r and r.get("data"):
-        for p in r["data"]["myself"]["pods"]:
-            if p["name"] == POD_NAME: return p
-    return None
-
-def deploy():
-    existing = find_pod()
-    if existing:
-        print(f"Exists: {existing['id']} ({existing['desiredStatus']})")
-        if existing["desiredStatus"] == "STOPPED":
-            gql(f'mutation {{ podResume(input: {{ podId: "{existing["id"]}" }}) {{ id }} }}')
-            print("Resuming...")
-        return existing["id"]
-    q = 'mutation { podFindAndDeployOnDemand(input: { cloudType: COMMUNITY, gpuCount: 1, gpuTypeId: "NVIDIA L40S", name: "' + POD_NAME + '", containerDiskInGb: 50, imageName: "' + ATLAS_IMAGE + '", ports: "8888/http", volumeInGb: 100, volumeMountPath: "/workspace" }) { id desiredStatus } }'
-    r = gql(q)
-    if r and r.get("data"):
-        pod = r["data"]["podFindAndDeployOnDemand"]
-        print(f"Deployed: {pod['id']} -> https://{pod['id']}-8888.proxy.runpod.net")
-        return pod["id"]
-    print("Deploy failed")
-
-def status():
-    pod = find_pod()
-    if not pod: print("No pod"); return
-    print(f"ID: {pod['id']}\nStatus: {pod['desiredStatus']}\nCost: ${pod['costPerHr']}/hr\nEndpoint: https://{pod['id']}-8888.proxy.runpod.net")
-    u = pod.get("runtime", {}).get("uptimeInSeconds", 0)
-    if u: print(f"Uptime: {u//3600}h {(u%3600)//60}m")
-
-def benchmark():
-    pod = find_pod()
-    if not pod or pod["desiredStatus"] != "RUNNING":
-        print("Pod not running"); return
-    ep = f"https://{pod['id']}-8888.proxy.runpod.net/v1"
-    print(f"Benchmarking: {ep}")
-    prompts = [
-        "Explain sovereign AI in 100 words.",
-        "Write quicksort in Python.",
-        "Compare transformers vs state space models.",
-        "Describe MoE architecture.",
-        "Write a Dockerfile for Flask+Redis.",
-    ]
-    results = []
-    for i, p in enumerate(prompts):
-        print(f"\n[{i+1}/5] {p[:40]}...")
-        start = time.time()
-        try:
-            payload = json.dumps({"model": MODEL, "messages": [{"role": "user", "content": p}], "max_tokens": 512}).encode()
-            req = urllib.request.Request(f"{ep}/chat/completions", data=payload,
-                headers={"Content-Type": "application/json", "Authorization": "Bearer dummy"}, method="POST")
-            with urllib.request.urlopen(req, timeout=120) as resp:
-                r = json.loads(resp.read().decode())
-            elapsed = time.time() - start
-            usage = r.get("usage", {})
-            tps = usage.get("completion_tokens", 0) / elapsed if elapsed > 0 else 0
-            results.append({"prompt": i, "tok_per_sec": round(tps, 2), "tokens": usage.get("completion_tokens", 0)})
-            print(f"  {usage.get('completion_tokens', 0)} tokens / {elapsed:.1f}s = {tps:.1f} tok/s")
-        except Exception as e:
-            print(f"  Error: {e}")
-            results.append({"prompt": i, "error": str(e)})
-    ok = [r for r in results if "tok_per_sec" in r]
-    if ok:
-        avg = sum(r["tok_per_sec"] for r in ok) / len(ok)
-        print(f"\nAvg: {avg:.1f} tok/s | Min: {min(r['tok_per_sec'] for r in ok):.1f} | Max: {max(r['tok_per_sec'] for r in ok):.1f}")
-        COST_LOG.parent.mkdir(parents=True, exist_ok=True)
-        with open(COST_LOG, "a") as f:
-            f.write(json.dumps({"ts": datetime.now(timezone.utc).isoformat(), "avg_tps": round(avg, 2), "results": results}) + "\n")
-
-def stop():
-    pod = find_pod()
-    if not pod: print("No pod"); return
-    gql(f'mutation {{ podStop(input: {{ podId: "{pod["id"]}" }}) {{ id }} }}')
-    print(f"Stopped: {pod['id']}")
-
-def terminate():
-    pod = find_pod()
-    if not pod: print("No pod"); return
-    gql(f'mutation {{ podTerminate(input: {{ podId: "{pod["id"]}" }}) }}')
-    print(f"Terminated: {pod['id']}")
-
-def main():
-    p = argparse.ArgumentParser(description="Atlas L40S Eval")
-    sub = p.add_subparsers(dest="cmd")
-    sub.add_parser("deploy"); sub.add_parser("status"); sub.add_parser("benchmark")
-    sub.add_parser("stop"); sub.add_parser("terminate")
-    args = p.parse_args()
-    if not args.cmd: p.print_help(); sys.exit(1)
-    {"deploy": deploy, "status": status, "benchmark": benchmark, "stop": stop, "terminate": terminate}[args.cmd]()
-
-if __name__ == "__main__": main()

tests/agent/test_privacy_filter.py

@@ -0,0 +1,202 @@
+"""Tests for agent.privacy_filter — PII stripping before remote API calls."""
+
+import pytest
+from agent.privacy_filter import (
+    PrivacyFilter,
+    RedactionReport,
+    Sensitivity,
+    sanitize_messages,
+    quick_sanitize,
+)
+
+
+class TestPrivacyFilterSanitizeText:
+    """Test single-text sanitization."""
+
+    def test_no_pii_returns_clean(self):
+        pf = PrivacyFilter()
+        text = "The weather in Paris is nice today."
+        cleaned, redactions = pf.sanitize_text(text)
+        assert cleaned == text
+        assert redactions == []
+
+    def test_email_redacted(self):
+        pf = PrivacyFilter()
+        text = "Send report to alice@example.com by Friday."
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "alice@example.com" not in cleaned
+        assert "[REDACTED-EMAIL]" in cleaned
+        assert any(r["type"] == "email_address" for r in redactions)
+
+    def test_phone_redacted(self):
+        pf = PrivacyFilter()
+        text = "Call me at 555-123-4567 when ready."
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "555-123-4567" not in cleaned
+        assert "[REDACTED-PHONE]" in cleaned
+
+    def test_api_key_redacted(self):
+        pf = PrivacyFilter()
+        text = 'api_key = "sk-proj-abcdefghij1234567890abcdefghij1234567890"'
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "sk-proj-" not in cleaned
+        assert any(r["sensitivity"] == "CRITICAL" for r in redactions)
+
+    def test_github_token_redacted(self):
+        pf = PrivacyFilter()
+        text = "Use ghp_1234567890abcdefghijklmnopqrstuvwxyz1234 for auth"
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "ghp_" not in cleaned
+        assert any(r["type"] == "github_token" for r in redactions)
+
+    def test_ethereum_address_redacted(self):
+        pf = PrivacyFilter()
+        text = "Send to 0x742d35Cc6634C0532925a3b844Bc9e7595f2bD18 please"
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "0x742d" not in cleaned
+        assert any(r["type"] == "ethereum_address" for r in redactions)
+
+    def test_user_home_path_redacted(self):
+        pf = PrivacyFilter()
+        text = "Read file at /Users/alice/Documents/secret.txt"
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "alice" not in cleaned
+        assert "[REDACTED-USER]" in cleaned
+
+    def test_multiple_pii_types(self):
+        pf = PrivacyFilter()
+        text = (
+            "Contact john@test.com or call 555-999-1234. "
+            "The API key is sk-abcdefghijklmnopqrstuvwxyz1234567890."
+        )
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "john@test.com" not in cleaned
+        assert "555-999-1234" not in cleaned
+        assert "sk-abcd" not in cleaned
+        assert len(redactions) >= 3
+
+
+class TestPrivacyFilterSanitizeMessages:
+    """Test message-list sanitization."""
+
+    def test_sanitize_user_message(self):
+        pf = PrivacyFilter()
+        messages = [
+            {"role": "system", "content": "You are helpful."},
+            {"role": "user", "content": "Email me at bob@test.com with results."},
+        ]
+        safe, report = pf.sanitize_messages(messages)
+        assert report.redacted_messages == 1
+        assert "bob@test.com" not in safe[1]["content"]
+        assert "[REDACTED-EMAIL]" in safe[1]["content"]
+        # System message unchanged
+        assert safe[0]["content"] == "You are helpful."
+
+    def test_no_redaction_needed(self):
+        pf = PrivacyFilter()
+        messages = [
+            {"role": "user", "content": "What is 2+2?"},
+            {"role": "assistant", "content": "4"},
+        ]
+        safe, report = pf.sanitize_messages(messages)
+        assert report.redacted_messages == 0
+        assert not report.had_redactions
+
+    def test_assistant_messages_also_sanitized(self):
+        pf = PrivacyFilter()
+        messages = [
+            {"role": "assistant", "content": "Your email admin@corp.com was found."},
+        ]
+        safe, report = pf.sanitize_messages(messages)
+        assert report.redacted_messages == 1
+        assert "admin@corp.com" not in safe[0]["content"]
+
+    def test_tool_messages_not_sanitized(self):
+        pf = PrivacyFilter()
+        messages = [
+            {"role": "tool", "content": "Result: user@test.com found"},
+        ]
+        safe, report = pf.sanitize_messages(messages)
+        assert report.redacted_messages == 0
+        assert safe[0]["content"] == "Result: user@test.com found"
+
+
+class TestShouldUseLocalOnly:
+    """Test the local-only routing decision."""
+
+    def test_normal_text_allows_remote(self):
+        pf = PrivacyFilter()
+        block, reason = pf.should_use_local_only("Summarize this article about Python.")
+        assert not block
+
+    def test_critical_secret_blocks_remote(self):
+        pf = PrivacyFilter()
+        text = "Here is the API key: sk-abcdefghijklmnopqrstuvwxyz1234567890"
+        block, reason = pf.should_use_local_only(text)
+        assert block
+        assert "critical" in reason.lower()
+
+    def test_multiple_high_sensitivity_blocks(self):
+        pf = PrivacyFilter()
+        # 3+ high-sensitivity patterns
+        text = (
+            "Card: 4111-1111-1111-1111, "
+            "SSN: 123-45-6789, "
+            "BTC: 1A1zP1eP5QGefi2DMPTfTL5SLmv7DivfNa, "
+            "ETH: 0x742d35Cc6634C0532925a3b844Bc9e7595f2bD18"
+        )
+        block, reason = pf.should_use_local_only(text)
+        assert block
+
+
+class TestAggressiveMode:
+    """Test aggressive filtering mode."""
+
+    def test_aggressive_redacts_internal_ips(self):
+        pf = PrivacyFilter(aggressive_mode=True)
+        text = "Server at 192.168.1.100 is responding."
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "192.168.1.100" not in cleaned
+        assert any(r["type"] == "internal_ip" for r in redactions)
+
+    def test_normal_does_not_redact_ips(self):
+        pf = PrivacyFilter(aggressive_mode=False)
+        text = "Server at 192.168.1.100 is responding."
+        cleaned, redactions = pf.sanitize_text(text)
+        assert "192.168.1.100" in cleaned  # IP preserved in normal mode
+
+
+class TestConvenienceFunctions:
+    """Test module-level convenience functions."""
+
+    def test_quick_sanitize(self):
+        text = "Contact alice@example.com for details"
+        result = quick_sanitize(text)
+        assert "alice@example.com" not in result
+        assert "[REDACTED-EMAIL]" in result
+
+    def test_sanitize_messages_convenience(self):
+        messages = [{"role": "user", "content": "Call 555-000-1234"}]
+        safe, report = sanitize_messages(messages)
+        assert report.redacted_messages == 1
+
+
+class TestRedactionReport:
+    """Test the reporting structure."""
+
+    def test_summary_no_redactions(self):
+        report = RedactionReport(total_messages=3, redacted_messages=0)
+        assert "No PII" in report.summary()
+
+    def test_summary_with_redactions(self):
+        report = RedactionReport(
+            total_messages=2,
+            redacted_messages=1,
+            redactions=[
+                {"type": "email_address", "sensitivity": "MEDIUM", "count": 2},
+                {"type": "phone_number_us", "sensitivity": "MEDIUM", "count": 1},
+            ],
+        )
+        summary = report.summary()
+        assert "1/2" in summary
+        assert "email_address" in summary

tests/tools/test_confirmation_daemon.py

@@ -0,0 +1,190 @@
+"""Tests for tools.confirmation_daemon — Human Confirmation Firewall."""
+
+import pytest
+import time
+from tools.confirmation_daemon import (
+    ConfirmationDaemon,
+    ConfirmationRequest,
+    ConfirmationStatus,
+    RiskLevel,
+    classify_action,
+    _is_whitelisted,
+    _DEFAULT_WHITELIST,
+)
+
+
+class TestClassifyAction:
+    """Test action risk classification."""
+
+    def test_crypto_tx_is_critical(self):
+        assert classify_action("crypto_tx") == RiskLevel.CRITICAL
+
+    def test_sign_transaction_is_critical(self):
+        assert classify_action("sign_transaction") == RiskLevel.CRITICAL
+
+    def test_send_email_is_high(self):
+        assert classify_action("send_email") == RiskLevel.HIGH
+
+    def test_send_message_is_medium(self):
+        assert classify_action("send_message") == RiskLevel.MEDIUM
+
+    def test_access_calendar_is_low(self):
+        assert classify_action("access_calendar") == RiskLevel.LOW
+
+    def test_unknown_action_is_medium(self):
+        assert classify_action("unknown_action_xyz") == RiskLevel.MEDIUM
+
+
+class TestWhitelist:
+    """Test whitelist auto-approval."""
+
+    def test_self_email_is_whitelisted(self):
+        whitelist = dict(_DEFAULT_WHITELIST)
+        payload = {"from": "me@test.com", "to": "me@test.com"}
+        assert _is_whitelisted("send_email", payload, whitelist) is True
+
+    def test_non_whitelisted_recipient_not_approved(self):
+        whitelist = dict(_DEFAULT_WHITELIST)
+        payload = {"to": "random@stranger.com"}
+        assert _is_whitelisted("send_email", payload, whitelist) is False
+
+    def test_whitelisted_contact_approved(self):
+        whitelist = {
+            "send_message": {"targets": ["alice", "bob"]},
+        }
+        assert _is_whitelisted("send_message", {"to": "alice"}, whitelist) is True
+        assert _is_whitelisted("send_message", {"to": "charlie"}, whitelist) is False
+
+    def test_no_whitelist_entry_means_not_whitelisted(self):
+        whitelist = {}
+        assert _is_whitelisted("crypto_tx", {"amount": 1.0}, whitelist) is False
+
+
+class TestConfirmationRequest:
+    """Test the request data model."""
+
+    def test_defaults(self):
+        req = ConfirmationRequest(
+            request_id="test-1",
+            action="send_email",
+            description="Test email",
+            risk_level="high",
+            payload={},
+        )
+        assert req.status == ConfirmationStatus.PENDING.value
+        assert req.created_at > 0
+        assert req.expires_at > req.created_at
+
+    def test_is_pending(self):
+        req = ConfirmationRequest(
+            request_id="test-2",
+            action="send_email",
+            description="Test",
+            risk_level="high",
+            payload={},
+            expires_at=time.time() + 300,
+        )
+        assert req.is_pending is True
+
+    def test_is_expired(self):
+        req = ConfirmationRequest(
+            request_id="test-3",
+            action="send_email",
+            description="Test",
+            risk_level="high",
+            payload={},
+            expires_at=time.time() - 10,
+        )
+        assert req.is_expired is True
+        assert req.is_pending is False
+
+    def test_to_dict(self):
+        req = ConfirmationRequest(
+            request_id="test-4",
+            action="send_email",
+            description="Test",
+            risk_level="medium",
+            payload={"to": "a@b.com"},
+        )
+        d = req.to_dict()
+        assert d["request_id"] == "test-4"
+        assert d["action"] == "send_email"
+        assert "is_pending" in d
+
+
+class TestConfirmationDaemon:
+    """Test the daemon logic (without HTTP layer)."""
+
+    def test_auto_approve_low_risk(self):
+        daemon = ConfirmationDaemon()
+        req = daemon.request(
+            action="access_calendar",
+            description="Read today's events",
+            risk_level="low",
+        )
+        assert req.status == ConfirmationStatus.AUTO_APPROVED.value
+
+    def test_whitelisted_auto_approves(self):
+        daemon = ConfirmationDaemon()
+        daemon._whitelist = {"send_message": {"targets": ["alice"]}}
+        req = daemon.request(
+            action="send_message",
+            description="Message alice",
+            payload={"to": "alice"},
+        )
+        assert req.status == ConfirmationStatus.AUTO_APPROVED.value
+
+    def test_non_whitelisted_goes_pending(self):
+        daemon = ConfirmationDaemon()
+        daemon._whitelist = {}
+        req = daemon.request(
+            action="send_email",
+            description="Email to stranger",
+            payload={"to": "stranger@test.com"},
+            risk_level="high",
+        )
+        assert req.status == ConfirmationStatus.PENDING.value
+        assert req.is_pending is True
+
+    def test_approve_response(self):
+        daemon = ConfirmationDaemon()
+        daemon._whitelist = {}
+        req = daemon.request(
+            action="send_email",
+            description="Email test",
+            risk_level="high",
+        )
+        result = daemon.respond(req.request_id, approved=True, decided_by="human")
+        assert result.status == ConfirmationStatus.APPROVED.value
+        assert result.decided_by == "human"
+
+    def test_deny_response(self):
+        daemon = ConfirmationDaemon()
+        daemon._whitelist = {}
+        req = daemon.request(
+            action="crypto_tx",
+            description="Send 1 ETH",
+            risk_level="critical",
+        )
+        result = daemon.respond(
+            req.request_id, approved=False, decided_by="human", reason="Too risky"
+        )
+        assert result.status == ConfirmationStatus.DENIED.value
+        assert result.reason == "Too risky"
+
+    def test_get_pending(self):
+        daemon = ConfirmationDaemon()
+        daemon._whitelist = {}
+        daemon.request(action="send_email", description="Test 1", risk_level="high")
+        daemon.request(action="send_email", description="Test 2", risk_level="high")
+        pending = daemon.get_pending()
+        assert len(pending) >= 2
+
+    def test_get_history(self):
+        daemon = ConfirmationDaemon()
+        req = daemon.request(
+            action="access_calendar", description="Test", risk_level="low"
+        )
+        history = daemon.get_history()
+        assert len(history) >= 1
+        assert history[0]["action"] == "access_calendar"

tools/approval.py

@@ -121,6 +121,19 @@ DANGEROUS_PATTERNS = [
     (r'\b(cp|mv|install)\b.*\s/etc/', "copy/move file into /etc/"),
     (r'\bsed\s+-[^\s]*i.*\s/etc/', "in-place edit of system config"),
     (r'\bsed\s+--in-place\b.*\s/etc/', "in-place edit of system config (long flag)"),
+    # --- Vitalik's threat model: crypto / financial ---
+    (r'\b(?:bitcoin-cli|ethers\.js|web3|ether\.sendTransaction)\b', "direct crypto transaction tool usage"),
+    (r'\bwget\b.*\b(?:mnemonic|seed\s*phrase|private[_-]?key)\b', "attempting to download crypto credentials"),
+    (r'\bcurl\b.*\b(?:mnemonic|seed\s*phrase|private[_-]?key)\b', "attempting to exfiltrate crypto credentials"),
+    # --- Vitalik's threat model: credential exfiltration ---
+    (r'\b(?:curl|wget|http|nc|ncat|socat)\b.*\b(?:\.env|\.ssh|credentials|secrets|token|api[_-]?key)\b',
+     "attempting to exfiltrate credentials via network"),
+    (r'\bbase64\b.*\|(?:\s*curl|\s*wget)', "base64-encode then network exfiltration"),
+    (r'\bcat\b.*\b(?:\.env|\.ssh/id_rsa|credentials)\b.*\|(?:\s*curl|\s*wget)',
+     "reading secrets and piping to network tool"),
+    # --- Vitalik's threat model: data exfiltration ---
+    (r'\bcurl\b.*-d\s.*\$(?:HOME|USER)', "sending user home directory data to remote"),
+    (r'\bwget\b.*--post-data\s.*\$(?:HOME|USER)', "posting user data to remote"),
     # Script execution via heredoc — bypasses the -e/-c flag patterns above.
     # `python3 << 'EOF'` feeds arbitrary code via stdin without -c/-e flags.
     (r'\b(python[23]?|perl|ruby|node)\s+<<', "script execution via heredoc"),

tools/confirmation_daemon.py

@@ -0,0 +1,615 @@
+"""Human Confirmation Daemon — HTTP server for two-factor action approval.
+
+Implements Vitalik's Pattern 1: "The new 'two-factor confirmation' is that
+the two factors are the human and the LLM."
+
+This daemon runs on localhost:6000 and provides a simple HTTP API for the
+agent to request human approval before executing high-risk actions.
+
+Threat model:
+- LLM jailbreaks: Remote content "hacking" the LLM to perform malicious actions
+- LLM accidents: LLM accidentally performing dangerous operations
+- The human acts as the second factor — the agent proposes, the human disposes
+
+Architecture:
+- Agent detects high-risk action → POST /confirm with action details
+- Daemon stores pending request, sends notification to user
+- User approves/denies via POST /respond (Telegram, CLI, or direct HTTP)
+- Agent receives decision and proceeds or aborts
+
+Usage:
+    # Start daemon (usually managed by gateway)
+    from tools.confirmation_daemon import ConfirmationDaemon
+    daemon = ConfirmationDaemon(port=6000)
+    daemon.start()
+
+    # Request approval (from agent code)
+    from tools.confirmation_daemon import request_confirmation
+    approved = request_confirmation(
+        action="send_email",
+        description="Send email to alice@example.com",
+        risk_level="high",
+        payload={"to": "alice@example.com", "subject": "Meeting notes"},
+        timeout=300,
+    )
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import os
+import threading
+import time
+import uuid
+from dataclasses import dataclass, field, asdict
+from enum import Enum, auto
+from pathlib import Path
+from typing import Any, Callable, Dict, List, Optional, Tuple
+
+logger = logging.getLogger(__name__)
+
+
+class RiskLevel(Enum):
+    """Risk classification for actions requiring confirmation."""
+    LOW = "low"           # Log only, no confirmation needed
+    MEDIUM = "medium"     # Confirm for non-whitelisted targets
+    HIGH = "high"         # Always confirm
+    CRITICAL = "critical" # Always confirm + require explicit reason
+
+
+class ConfirmationStatus(Enum):
+    """Status of a pending confirmation request."""
+    PENDING = "pending"
+    APPROVED = "approved"
+    DENIED = "denied"
+    EXPIRED = "expired"
+    AUTO_APPROVED = "auto_approved"
+
+
+@dataclass
+class ConfirmationRequest:
+    """A request for human confirmation of a high-risk action."""
+    request_id: str
+    action: str               # Action type: send_email, send_message, crypto_tx, etc.
+    description: str          # Human-readable description of what will happen
+    risk_level: str           # low, medium, high, critical
+    payload: Dict[str, Any]   # Action-specific data (sanitized)
+    session_key: str = ""     # Session that initiated the request
+    created_at: float = 0.0
+    expires_at: float = 0.0
+    status: str = ConfirmationStatus.PENDING.value
+    decided_at: float = 0.0
+    decided_by: str = ""      # "human", "auto", "whitelist"
+    reason: str = ""          # Optional reason for denial
+
+    def __post_init__(self):
+        if not self.created_at:
+            self.created_at = time.time()
+        if not self.expires_at:
+            self.expires_at = self.created_at + 300  # 5 min default
+        if not self.request_id:
+            self.request_id = str(uuid.uuid4())[:12]
+
+    @property
+    def is_expired(self) -> bool:
+        return time.time() > self.expires_at
+
+    @property
+    def is_pending(self) -> bool:
+        return self.status == ConfirmationStatus.PENDING.value and not self.is_expired
+
+    def to_dict(self) -> Dict[str, Any]:
+        d = asdict(self)
+        d["is_expired"] = self.is_expired
+        d["is_pending"] = self.is_pending
+        return d
+
+
+# =========================================================================
+# Action categories (Vitalik's threat model)
+# =========================================================================
+
+ACTION_CATEGORIES = {
+    # Messaging — outbound communication to external parties
+    "send_email": RiskLevel.HIGH,
+    "send_message": RiskLevel.MEDIUM,     # Depends on recipient
+    "send_signal": RiskLevel.HIGH,
+    "send_telegram": RiskLevel.MEDIUM,
+    "send_discord": RiskLevel.MEDIUM,
+    "post_social": RiskLevel.HIGH,
+
+    # Financial / crypto
+    "crypto_tx": RiskLevel.CRITICAL,
+    "sign_transaction": RiskLevel.CRITICAL,
+    "access_wallet": RiskLevel.CRITICAL,
+    "modify_balance": RiskLevel.CRITICAL,
+
+    # System modification
+    "install_software": RiskLevel.HIGH,
+    "modify_system_config": RiskLevel.HIGH,
+    "modify_firewall": RiskLevel.CRITICAL,
+    "add_ssh_key": RiskLevel.CRITICAL,
+    "create_user": RiskLevel.CRITICAL,
+
+    # Data access
+    "access_contacts": RiskLevel.MEDIUM,
+    "access_calendar": RiskLevel.LOW,
+    "read_private_files": RiskLevel.MEDIUM,
+    "upload_data": RiskLevel.HIGH,
+    "share_credentials": RiskLevel.CRITICAL,
+
+    # Network
+    "open_port": RiskLevel.HIGH,
+    "modify_dns": RiskLevel.HIGH,
+    "expose_service": RiskLevel.CRITICAL,
+}
+
+# Default: any unrecognized action is MEDIUM risk
+DEFAULT_RISK_LEVEL = RiskLevel.MEDIUM
+
+
+def classify_action(action: str) -> RiskLevel:
+    """Classify an action by its risk level."""
+    return ACTION_CATEGORIES.get(action, DEFAULT_RISK_LEVEL)
+
+
+# =========================================================================
+# Whitelist configuration
+# =========================================================================
+
+_DEFAULT_WHITELIST = {
+    "send_message": {
+        "targets": [],   # Contact names/IDs that don't need confirmation
+    },
+    "send_email": {
+        "targets": [],   # Email addresses that don't need confirmation
+        "self_only": True,  # send-to-self always allowed
+    },
+}
+
+
+def _load_whitelist() -> Dict[str, Any]:
+    """Load action whitelist from config."""
+    config_path = Path.home() / ".hermes" / "approval_whitelist.json"
+    if config_path.exists():
+        try:
+            with open(config_path) as f:
+                return json.load(f)
+        except Exception as e:
+            logger.warning("Failed to load approval whitelist: %s", e)
+    return dict(_DEFAULT_WHITELIST)
+
+
+def _is_whitelisted(action: str, payload: Dict[str, Any], whitelist: Dict) -> bool:
+    """Check if an action is pre-approved by the whitelist."""
+    action_config = whitelist.get(action, {})
+    if not action_config:
+        return False
+
+    # Check target-based whitelist
+    targets = action_config.get("targets", [])
+    target = payload.get("to") or payload.get("recipient") or payload.get("target", "")
+    if target and target in targets:
+        return True
+
+    # Self-only email
+    if action_config.get("self_only") and action == "send_email":
+        sender = payload.get("from", "")
+        recipient = payload.get("to", "")
+        if sender and recipient and sender.lower() == recipient.lower():
+            return True
+
+    return False
+
+
+# =========================================================================
+# Confirmation daemon
+# =========================================================================
+
+class ConfirmationDaemon:
+    """HTTP daemon for human confirmation of high-risk actions.
+
+    Runs on localhost:PORT (default 6000). Provides:
+    - POST /confirm   — agent requests human approval
+    - POST /respond   — human approves/denies
+    - GET  /pending   — list pending requests
+    - GET  /health    — health check
+    """
+
+    def __init__(
+        self,
+        host: str = "127.0.0.1",
+        port: int = 6000,
+        default_timeout: int = 300,
+        notify_callback: Optional[Callable] = None,
+    ):
+        self.host = host
+        self.port = port
+        self.default_timeout = default_timeout
+        self.notify_callback = notify_callback
+        self._pending: Dict[str, ConfirmationRequest] = {}
+        self._history: List[ConfirmationRequest] = []
+        self._lock = threading.Lock()
+        self._whitelist = _load_whitelist()
+        self._app = None
+        self._runner = None
+
+    def request(
+        self,
+        action: str,
+        description: str,
+        payload: Optional[Dict[str, Any]] = None,
+        risk_level: Optional[str] = None,
+        session_key: str = "",
+        timeout: Optional[int] = None,
+    ) -> ConfirmationRequest:
+        """Create a confirmation request.
+
+        Returns the request. Check .status to see if it was immediately
+        auto-approved (whitelisted) or is pending human review.
+        """
+        payload = payload or {}
+
+        # Classify risk if not specified
+        if risk_level is None:
+            risk_level = classify_action(action).value
+
+        # Check whitelist
+        if risk_level in ("low",) or _is_whitelisted(action, payload, self._whitelist):
+            req = ConfirmationRequest(
+                request_id=str(uuid.uuid4())[:12],
+                action=action,
+                description=description,
+                risk_level=risk_level,
+                payload=payload,
+                session_key=session_key,
+                expires_at=time.time() + (timeout or self.default_timeout),
+                status=ConfirmationStatus.AUTO_APPROVED.value,
+                decided_at=time.time(),
+                decided_by="whitelist",
+            )
+            with self._lock:
+                self._history.append(req)
+            logger.info("Auto-approved whitelisted action: %s", action)
+            return req
+
+        # Create pending request
+        req = ConfirmationRequest(
+            request_id=str(uuid.uuid4())[:12],
+            action=action,
+            description=description,
+            risk_level=risk_level,
+            payload=payload,
+            session_key=session_key,
+            expires_at=time.time() + (timeout or self.default_timeout),
+        )
+
+        with self._lock:
+            self._pending[req.request_id] = req
+
+        # Notify human
+        if self.notify_callback:
+            try:
+                self.notify_callback(req.to_dict())
+            except Exception as e:
+                logger.warning("Confirmation notify callback failed: %s", e)
+
+        logger.info(
+            "Confirmation request %s: %s (%s risk) — waiting for human",
+            req.request_id, action, risk_level,
+        )
+        return req
+
+    def respond(
+        self,
+        request_id: str,
+        approved: bool,
+        decided_by: str = "human",
+        reason: str = "",
+    ) -> Optional[ConfirmationRequest]:
+        """Record a human decision on a pending request."""
+        with self._lock:
+            req = self._pending.get(request_id)
+            if not req:
+                logger.warning("Confirmation respond: unknown request %s", request_id)
+                return None
+            if not req.is_pending:
+                logger.warning("Confirmation respond: request %s already decided", request_id)
+                return req
+
+            req.status = (
+                ConfirmationStatus.APPROVED.value if approved
+                else ConfirmationStatus.DENIED.value
+            )
+            req.decided_at = time.time()
+            req.decided_by = decided_by
+            req.reason = reason
+
+            # Move to history
+            del self._pending[request_id]
+            self._history.append(req)
+
+        logger.info(
+            "Confirmation %s: %s by %s",
+            request_id, "APPROVED" if approved else "DENIED", decided_by,
+        )
+        return req
+
+    def wait_for_decision(
+        self, request_id: str, timeout: Optional[float] = None
+    ) -> ConfirmationRequest:
+        """Block until a decision is made or timeout expires."""
+        deadline = time.time() + (timeout or self.default_timeout)
+        while time.time() < deadline:
+            with self._lock:
+                req = self._pending.get(request_id)
+                if req and not req.is_pending:
+                    return req
+                if req and req.is_expired:
+                    req.status = ConfirmationStatus.EXPIRED.value
+                    del self._pending[request_id]
+                    self._history.append(req)
+                    return req
+            time.sleep(0.5)
+
+        # Timeout
+        with self._lock:
+            req = self._pending.pop(request_id, None)
+            if req:
+                req.status = ConfirmationStatus.EXPIRED.value
+                self._history.append(req)
+                return req
+
+        # Shouldn't reach here
+        return ConfirmationRequest(
+            request_id=request_id,
+            action="unknown",
+            description="Request not found",
+            risk_level="high",
+            payload={},
+            status=ConfirmationStatus.EXPIRED.value,
+        )
+
+    def get_pending(self) -> List[Dict[str, Any]]:
+        """Return list of pending confirmation requests."""
+        self._expire_old()
+        with self._lock:
+            return [r.to_dict() for r in self._pending.values() if r.is_pending]
+
+    def get_history(self, limit: int = 50) -> List[Dict[str, Any]]:
+        """Return recent confirmation history."""
+        with self._lock:
+            return [r.to_dict() for r in self._history[-limit:]]
+
+    def _expire_old(self) -> None:
+        """Move expired requests to history."""
+        now = time.time()
+        with self._lock:
+            expired = [
+                rid for rid, req in self._pending.items()
+                if now > req.expires_at
+            ]
+            for rid in expired:
+                req = self._pending.pop(rid)
+                req.status = ConfirmationStatus.EXPIRED.value
+                self._history.append(req)
+
+    # --- aiohttp HTTP API ---
+
+    async def _handle_health(self, request):
+        from aiohttp import web
+        return web.json_response({
+            "status": "ok",
+            "service": "hermes-confirmation-daemon",
+            "pending": len(self._pending),
+        })
+
+    async def _handle_confirm(self, request):
+        from aiohttp import web
+        try:
+            body = await request.json()
+        except Exception:
+            return web.json_response({"error": "invalid JSON"}, status=400)
+
+        action = body.get("action", "")
+        description = body.get("description", "")
+        if not action or not description:
+            return web.json_response(
+                {"error": "action and description required"}, status=400
+            )
+
+        req = self.request(
+            action=action,
+            description=description,
+            payload=body.get("payload", {}),
+            risk_level=body.get("risk_level"),
+            session_key=body.get("session_key", ""),
+            timeout=body.get("timeout"),
+        )
+
+        # If auto-approved, return immediately
+        if req.status != ConfirmationStatus.PENDING.value:
+            return web.json_response({
+                "request_id": req.request_id,
+                "status": req.status,
+                "decided_by": req.decided_by,
+            })
+
+        # Otherwise, wait for human decision (with timeout)
+        timeout = min(body.get("timeout", self.default_timeout), 600)
+        result = self.wait_for_decision(req.request_id, timeout=timeout)
+
+        return web.json_response({
+            "request_id": result.request_id,
+            "status": result.status,
+            "decided_by": result.decided_by,
+            "reason": result.reason,
+        })
+
+    async def _handle_respond(self, request):
+        from aiohttp import web
+        try:
+            body = await request.json()
+        except Exception:
+            return web.json_response({"error": "invalid JSON"}, status=400)
+
+        request_id = body.get("request_id", "")
+        approved = body.get("approved")
+        if not request_id or approved is None:
+            return web.json_response(
+                {"error": "request_id and approved required"}, status=400
+            )
+
+        result = self.respond(
+            request_id=request_id,
+            approved=bool(approved),
+            decided_by=body.get("decided_by", "human"),
+            reason=body.get("reason", ""),
+        )
+
+        if not result:
+            return web.json_response({"error": "unknown request"}, status=404)
+
+        return web.json_response({
+            "request_id": result.request_id,
+            "status": result.status,
+        })
+
+    async def _handle_pending(self, request):
+        from aiohttp import web
+        return web.json_response({"pending": self.get_pending()})
+
+    def _build_app(self):
+        """Build the aiohttp application."""
+        from aiohttp import web
+
+        app = web.Application()
+        app.router.add_get("/health", self._handle_health)
+        app.router.add_post("/confirm", self._handle_confirm)
+        app.router.add_post("/respond", self._handle_respond)
+        app.router.add_get("/pending", self._handle_pending)
+        self._app = app
+        return app
+
+    async def start_async(self) -> None:
+        """Start the daemon as an async server."""
+        from aiohttp import web
+
+        app = self._build_app()
+        self._runner = web.AppRunner(app)
+        await self._runner.setup()
+        site = web.TCPSite(self._runner, self.host, self.port)
+        await site.start()
+        logger.info("Confirmation daemon listening on %s:%d", self.host, self.port)
+
+    async def stop_async(self) -> None:
+        """Stop the daemon."""
+        if self._runner:
+            await self._runner.cleanup()
+            self._runner = None
+
+    def start(self) -> None:
+        """Start daemon in a background thread (blocking caller)."""
+        def _run():
+            loop = asyncio.new_event_loop()
+            asyncio.set_event_loop(loop)
+            loop.run_until_complete(self.start_async())
+            loop.run_forever()
+
+        t = threading.Thread(target=_run, daemon=True, name="confirmation-daemon")
+        t.start()
+        logger.info("Confirmation daemon started in background thread")
+
+    def start_blocking(self) -> None:
+        """Start daemon and block (for standalone use)."""
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+        loop.run_until_complete(self.start_async())
+        try:
+            loop.run_forever()
+        except KeyboardInterrupt:
+            pass
+        finally:
+            loop.run_until_complete(self.stop_async())
+
+
+# =========================================================================
+# Convenience API for agent integration
+# =========================================================================
+
+# Global singleton — initialized by gateway or CLI at startup
+_daemon: Optional[ConfirmationDaemon] = None
+
+
+def get_daemon() -> Optional[ConfirmationDaemon]:
+    """Get the global confirmation daemon instance."""
+    return _daemon
+
+
+def init_daemon(
+    host: str = "127.0.0.1",
+    port: int = 6000,
+    notify_callback: Optional[Callable] = None,
+) -> ConfirmationDaemon:
+    """Initialize the global confirmation daemon."""
+    global _daemon
+    _daemon = ConfirmationDaemon(
+        host=host, port=port, notify_callback=notify_callback
+    )
+    return _daemon
+
+
+def request_confirmation(
+    action: str,
+    description: str,
+    payload: Optional[Dict[str, Any]] = None,
+    risk_level: Optional[str] = None,
+    session_key: str = "",
+    timeout: int = 300,
+) -> bool:
+    """Request human confirmation for a high-risk action.
+
+    This is the primary integration point for agent code. It:
+    1. Classifies the action risk level
+    2. Checks the whitelist
+    3. If confirmation needed, blocks until human responds
+    4. Returns True if approved, False if denied/expired
+
+    Args:
+        action: Action type (send_email, crypto_tx, etc.)
+        description: Human-readable description
+        payload: Action-specific data
+        risk_level: Override auto-classification
+        session_key: Session requesting approval
+        timeout: Seconds to wait for human response
+
+    Returns:
+        True if approved, False if denied or expired.
+    """
+    daemon = get_daemon()
+    if not daemon:
+        logger.warning(
+            "No confirmation daemon running — DENYING action %s by default. "
+            "Start daemon with init_daemon() or --confirmation-daemon flag.",
+            action,
+        )
+        return False
+
+    req = daemon.request(
+        action=action,
+        description=description,
+        payload=payload,
+        risk_level=risk_level,
+        session_key=session_key,
+        timeout=timeout,
+    )
+
+    # Auto-approved (whitelisted)
+    if req.status == ConfirmationStatus.AUTO_APPROVED.value:
+        return True
+
+    # Wait for human
+    result = daemon.wait_for_decision(req.request_id, timeout=timeout)
+    return result.status == ConfirmationStatus.APPROVED.value