feat: add model fallback verification script (#514)

- Tests model switches with verification prompts
- Validates context window meets 64K minimum
- Checks primary model and fallback chain
- Supports OpenRouter, Anthropic, Nous, Kimi, Ollama
- Exits with error if no viable model found

Closes #514

bin/model-fallback-verify.py

@@ -0,0 +1,443 @@
+#!/usr/bin/env python3
+"""
+Model Fallback Verification Script
+Issue #514: [Robustness] Model fallback verification — test before trusting
+
+Tests model switches with verification prompts, validates context windows,
+and ensures at least one viable model is available before starting loops.
+
+Usage:
+  python3 model-fallback-verify.py              # Run full verification
+  python3 model-fallback-verify.py check <model> # Test specific model
+  python3 model-fallback-verify.py context <model> # Check context window
+  python3 model-fallback-verify.py list          # List available models
+"""
+
+import os, sys, json, yaml, urllib.request
+from datetime import datetime, timezone
+from pathlib import Path
+
+# Configuration
+HERMES_HOME = Path(os.environ.get("HERMES_HOME", Path.home() / ".hermes"))
+CONFIG_FILE = HERMES_HOME / "config.yaml"
+LOG_DIR = HERMES_HOME / "logs"
+LOG_FILE = LOG_DIR / "model-verify.log"
+MIN_CONTEXT_WINDOW = 64 * 1024  # 64K tokens minimum
+
+# Provider endpoints
+PROVIDER_CONFIGS = {
+    "openrouter": {
+        "base_url": "https://openrouter.ai/api/v1",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+    "anthropic": {
+        "base_url": "https://api.anthropic.com/v1",
+        "headers": lambda api_key: {"x-api-key": api_key, "anthropic-version": "2023-06-01"},
+        "chat_url": "/messages",
+    },
+    "nous": {
+        "base_url": "https://inference.nousresearch.com/v1",
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+    "kimi-coding": {
+        "base_url": "https://api.kimi.com/coding/v1",
+        "headers": lambda api_key: {"x-api-key": api_key, "x-api-provider": "kimi-coding"},
+        "chat_url": "/chat/completions",
+    },
+    "custom": {
+        "base_url": None,
+        "headers": lambda api_key: {"Authorization": "Bearer " + api_key},
+        "chat_url": "/chat/completions",
+    },
+}
+
+# Known context windows for common models
+KNOWN_CONTEXT_WINDOWS = {
+    "claude-opus-4-6": 200000,
+    "claude-sonnet-4": 200000,
+    "claude-3.5-sonnet": 200000,
+    "gpt-4o": 128000,
+    "gpt-4": 128000,
+    "gpt-3.5-turbo": 16385,
+    "qwen3:30b": 32768,
+    "qwen2.5:7b": 32768,
+    "hermes4:14b": 32768,
+    "gemma3:1b": 8192,
+    "gemma4": 32768,
+    "phi3:3.8b": 128000,
+    "kimi-k2.5": 128000,
+    "google/gemini-2.5-pro": 1048576,
+    "xiaomi/mimo-v2-pro": 131072,
+    "deepseek/deepseek-r1": 131072,
+    "deepseek/deepseek-chat-v3-0324": 131072,
+}
+
+def log(msg):
+    """Log message to file and optionally to console."""
+    timestamp = datetime.now(timezone.utc).strftime("%Y-%m-%d %H:%M:%S")
+    log_entry = "[" + timestamp + "] " + msg
+    
+    LOG_DIR.mkdir(parents=True, exist_ok=True)
+    with open(LOG_FILE, "a") as f:
+        f.write(log_entry + "\n")
+    
+    if "--quiet" not in sys.argv:
+        print(log_entry)
+
+def load_config():
+    """Load Hermes config.yaml."""
+    if not CONFIG_FILE.exists():
+        return None
+    
+    with open(CONFIG_FILE) as f:
+        return yaml.safe_load(f)
+
+def get_provider_api_key(provider):
+    """Get API key for a provider from .env or environment."""
+    env_file = HERMES_HOME / ".env"
+    if env_file.exists():
+        with open(env_file) as f:
+            for line in f:
+                line = line.strip()
+                if line.startswith(provider.upper() + "_API_KEY="):
+                    return line.split("=", 1)[1].strip().strip("'\"")
+    
+    return os.environ.get(provider.upper() + "_API_KEY")
+
+def get_ollama_models():
+    """Get list of available Ollama models."""
+    ollama_host = os.environ.get("OLLAMA_HOST", "localhost:11434")
+    try:
+        resp = urllib.request.urlopen("http://" + ollama_host + "/api/tags", timeout=5)
+        data = json.loads(resp.read())
+        return [m["name"] for m in data.get("models", [])]
+    except:
+        return []
+
+def test_model(model, provider, api_key=None, base_url=None):
+    """
+    Test a model with a verification prompt.
+    Returns (success, response, error_message)
+    """
+    if provider == "ollama" or ":" in model:
+        # Local Ollama model
+        ollama_host = os.environ.get("OLLAMA_HOST", "localhost:11434")
+        try:
+            body = json.dumps({
+                "model": model,
+                "prompt": "Say exactly VERIFIED and nothing else.",
+                "stream": False,
+                "options": {"num_predict": 10}
+            }).encode()
+            req = urllib.request.Request(
+                "http://" + ollama_host + "/api/generate",
+                data=body,
+                headers={"Content-Type": "application/json"}
+            )
+            resp = urllib.request.urlopen(req, timeout=30)
+            result = json.loads(resp.read())
+            response_text = result.get("response", "").strip()
+            if "VERIFIED" in response_text.upper():
+                return True, response_text, None
+            return False, response_text, "Unexpected response: " + response_text[:100]
+        except Exception as e:
+            return False, "", "Ollama error: " + str(e)[:200]
+    
+    # Cloud provider
+    config = PROVIDER_CONFIGS.get(provider)
+    if not config:
+        return False, "", "Unknown provider: " + provider
+    
+    url = base_url or config["base_url"]
+    if not url:
+        return False, "", "No base URL for provider: " + provider
+    
+    headers = config["headers"](api_key or "")
+    headers["Content-Type"] = "application/json"
+    
+    try:
+        body = json.dumps({
+            "model": model,
+            "max_tokens": 20,
+            "messages": [{"role": "user", "content": "Say exactly VERIFIED and nothing else."}]
+        }).encode()
+        
+        req = urllib.request.Request(
+            url + config["chat_url"],
+            data=body,
+            headers=headers
+        )
+        resp = urllib.request.urlopen(req, timeout=30)
+        result = json.loads(resp.read())
+        
+        if provider == "anthropic":
+            content = result.get("content", [{}])[0].get("text", "")
+        else:
+            choices = result.get("choices", [{}])
+            content = choices[0].get("message", {}).get("content", "") if choices else ""
+        
+        if "VERIFIED" in content.upper():
+            return True, content, None
+        return False, content, "Unexpected response: " + content[:100]
+    
+    except urllib.error.HTTPError as e:
+        error_body = e.read().decode() if e.fp else str(e)
+        if e.code == 404:
+            return False, "", "Model not found (404): " + error_body[:200]
+        elif e.code == 429:
+            return True, "", "Rate limited but model exists"
+        elif e.code >= 500:
+            return False, "", "Server error (" + str(e.code) + "): " + error_body[:200]
+        else:
+            return False, "", "HTTP " + str(e.code) + ": " + error_body[:200]
+    except Exception as e:
+        return False, "", "Request error: " + str(e)[:200]
+
+def get_context_window(model, provider):
+    """
+    Get the context window size for a model.
+    Returns (window_size, source)
+    """
+    if model in KNOWN_CONTEXT_WINDOWS:
+        return KNOWN_CONTEXT_WINDOWS[model], "known"
+    
+    model_lower = model.lower()
+    if "claude" in model_lower:
+        return 200000, "inferred (claude)"
+    elif "gpt-4" in model_lower:
+        return 128000, "inferred (gpt-4)"
+    elif "gemini" in model_lower:
+        return 1048576, "inferred (gemini)"
+    elif "qwen" in model_lower:
+        return 32768, "inferred (qwen)"
+    elif "gemma" in model_lower:
+        return 8192, "inferred (gemma)"
+    elif "phi" in model_lower:
+        return 128000, "inferred (phi)"
+    
+    return 32768, "default"
+
+def verify_model(model, provider, api_key=None, base_url=None):
+    """
+    Full verification of a model: test prompt + context window.
+    Returns dict with verification results.
+    """
+    result = {
+        "model": model,
+        "provider": provider,
+        "tested": False,
+        "responded": False,
+        "response": "",
+        "error": None,
+        "context_window": 0,
+        "context_source": "unknown",
+        "meets_minimum": False,
+        "viable": False,
+    }
+    
+    success, response, error = test_model(model, provider, api_key, base_url)
+    result["tested"] = True
+    result["responded"] = success
+    result["response"] = response[:200] if response else ""
+    result["error"] = error
+    
+    window, source = get_context_window(model, provider)
+    result["context_window"] = window
+    result["context_source"] = source
+    result["meets_minimum"] = window >= MIN_CONTEXT_WINDOW
+    
+    result["viable"] = success and result["meets_minimum"]
+    
+    return result
+
+def get_fallback_chain(config):
+    """Get the fallback chain from config or defaults."""
+    chain = []
+    
+    model_config = config.get("model", {})
+    if isinstance(model_config, dict):
+        primary = model_config.get("default", "")
+        provider = model_config.get("provider", "")
+        if primary and provider:
+            chain.append({"model": primary, "provider": provider, "role": "primary"})
+    elif model_config:
+        chain.append({"model": str(model_config), "provider": "unknown", "role": "primary"})
+    
+    auxiliary = config.get("auxiliary", {})
+    for aux_name, aux_config in auxiliary.items():
+        if isinstance(aux_config, dict):
+            aux_model = aux_config.get("model", "")
+            aux_provider = aux_config.get("provider", "")
+            if aux_model and aux_provider and aux_provider != "auto":
+                chain.append({"model": aux_model, "provider": aux_provider, "role": "auxiliary:" + aux_name})
+    
+    ollama_models = get_ollama_models()
+    for model in ollama_models[:3]:
+        if not any(c["model"] == model for c in chain):
+            chain.append({"model": model, "provider": "ollama", "role": "local-fallback"})
+    
+    return chain
+
+def run_verification():
+    """Run full model fallback verification."""
+    log("=== Model Fallback Verification ===")
+    
+    config = load_config()
+    if not config:
+        log("ERROR: No config.yaml found")
+        return {"success": False, "error": "No config file"}
+    
+    chain = get_fallback_chain(config)
+    if not chain:
+        log("ERROR: No models configured")
+        return {"success": False, "error": "No models in chain"}
+    
+    results = []
+    viable_models = []
+    
+    for entry in chain:
+        model = entry["model"]
+        provider = entry["provider"]
+        role = entry["role"]
+        
+        api_key = get_provider_api_key(provider) if provider != "ollama" else None
+        
+        base_url = None
+        if provider == "custom":
+            provider_config = config.get("auxiliary", {}).get("vision", {})
+            base_url = provider_config.get("base_url")
+        
+        log("Testing [" + role + "] " + model + " (" + provider + ")...")
+        result = verify_model(model, provider, api_key, base_url)
+        result["role"] = role
+        results.append(result)
+        
+        status = "PASS" if result["viable"] else "FAIL"
+        details = []
+        if not result["responded"]:
+            details.append("no response: " + str(result["error"]))
+        if not result["meets_minimum"]:
+            details.append("context " + str(result["context_window"]) + " < " + str(MIN_CONTEXT_WINDOW))
+        
+        log("  [" + status + "] " + model + " - " + (", ".join(details) if details else "verified"))
+        
+        if result["viable"]:
+            viable_models.append(result)
+    
+    log("=== Results: " + str(len(viable_models)) + "/" + str(len(results)) + " models viable ===")
+    
+    if not viable_models:
+        log("CRITICAL: No viable models found!")
+        for r in results:
+            log("  - " + r["model"] + " (" + r["provider"] + "): responded=" + str(r["responded"]) + ", context=" + str(r["context_window"]))
+        return {"success": False, "results": results, "viable": []}
+    
+    log("Viable models (in priority order):")
+    for i, r in enumerate(viable_models, 1):
+        log("  " + str(i) + ". " + r["model"] + " (" + r["provider"] + ") - context: " + str(r["context_window"]) + " tokens [" + r["role"] + "]")
+    
+    return {
+        "success": True,
+        "results": results,
+        "viable": viable_models,
+        "primary": viable_models[0] if viable_models else None,
+    }
+
+def check_single_model(model):
+    """Check a specific model."""
+    if ":" in model:
+        provider = "ollama"
+    elif "/" in model:
+        provider = "openrouter"
+    else:
+        provider = "unknown"
+    
+    config = load_config() or {}
+    api_key = get_provider_api_key(provider) if provider != "ollama" else None
+    
+    result = verify_model(model, provider, api_key)
+    
+    if result["viable"]:
+        print("PASS: " + model)
+        print("  Context window: " + str(result["context_window"]) + " tokens")
+        print("  Response: " + result["response"][:100])
+    else:
+        print("FAIL: " + model)
+        if result["error"]:
+            print("  Error: " + str(result["error"]))
+        if not result["meets_minimum"]:
+            print("  Context window: " + str(result["context_window"]) + " < " + str(MIN_CONTEXT_WINDOW) + " minimum")
+    
+    return result["viable"]
+
+def check_context_window(model):
+    """Check context window for a model."""
+    if ":" in model:
+        provider = "ollama"
+    elif "/" in model:
+        provider = "openrouter"
+    else:
+        provider = "unknown"
+    
+    window, source = get_context_window(model, provider)
+    meets = window >= MIN_CONTEXT_WINDOW
+    
+    print("Model: " + model)
+    print("Provider: " + provider)
+    print("Context window: " + str(window) + " tokens (" + source + ")")
+    print("Minimum (" + str(MIN_CONTEXT_WINDOW) + "): " + ("PASS" if meets else "FAIL"))
+    
+    return meets
+
+def list_models():
+    """List all available models."""
+    config = load_config() or {}
+    chain = get_fallback_chain(config)
+    
+    print("Configured models:")
+    for entry in chain:
+        print("  " + entry["model"].ljust(30) + " " + entry["provider"].ljust(15) + " [" + entry["role"] + "]")
+    
+    ollama = get_ollama_models()
+    if ollama:
+        print("")
+        print("Ollama models:")
+        for m in ollama:
+            print("  " + m)
+
+def main():
+    if len(sys.argv) < 2:
+        result = run_verification()
+        sys.exit(0 if result["success"] else 1)
+    
+    cmd = sys.argv[1]
+    
+    if cmd == "check" and len(sys.argv) > 2:
+        model = sys.argv[2]
+        success = check_single_model(model)
+        sys.exit(0 if success else 1)
+    
+    elif cmd == "context" and len(sys.argv) > 2:
+        model = sys.argv[2]
+        meets = check_context_window(model)
+        sys.exit(0 if meets else 1)
+    
+    elif cmd == "list":
+        list_models()
+    
+    elif cmd == "test":
+        result = run_verification()
+        sys.exit(0 if result["success"] else 1)
+    
+    else:
+        print("Usage:")
+        print("  model-fallback-verify.py              Run full verification")
+        print("  model-fallback-verify.py check <model> Test specific model")
+        print("  model-fallback-verify.py context <model> Check context window")
+        print("  model-fallback-verify.py list          List available models")
+        sys.exit(1)
+
+if __name__ == "__main__":
+    main()

bin/nostr-agent-demo.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Full Nostr agent-to-agent communication demo - FINAL WORKING
 """

bin/soul_eval_gate.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 """
 Soul Eval Gate — The Conscience of the Training Pipeline
 

cron/pipeline-scheduler.yml

@@ -0,0 +1,9 @@
+- name: Nightly Pipeline Scheduler
+  schedule: '*/30 18-23,0-8 * * *' # Every 30 min, off-peak hours only
+  tasks:
+    - name: Check and start pipelines
+      shell: "bash scripts/nightly-pipeline-scheduler.sh"
+      env:
+        PIPELINE_TOKEN_LIMIT: "500000"
+        PIPELINE_PEAK_START: "9"
+        PIPELINE_PEAK_END: "18"

scripts/captcha_bypass_handler.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/diagram_meaning_extractor.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision
 

scripts/nightly-pipeline-scheduler.md

@@ -0,0 +1,50 @@
+# Nightly Pipeline Scheduler
+
+Auto-starts batch pipelines when inference is available.
+
+## What It Does
+
+1. Checks inference provider health (OpenRouter, Ollama, RunPod)
+2. Checks if it's off-peak hours (configurable, default: after 6PM)
+3. Checks interactive session load (don't fight with live users)
+4. Checks daily token budget (configurable limit)
+5. Starts the highest-priority incomplete pipeline
+
+## Pipeline Priority Order
+
+| Priority | Pipeline | Deps | Max Tokens |
+|----------|----------|------|------------|
+| 1 | playground-factory | none | 100,000 |
+| 2 | training-factory | none | 150,000 |
+| 3 | knowledge-mine | training-factory running | 80,000 |
+| 4 | adversary | knowledge-mine running | 50,000 |
+| 5 | codebase-genome | none | 120,000 |
+
+## Usage
+
+```bash
+# Normal run (used by cron)
+./scripts/nightly-pipeline-scheduler.sh
+
+# Dry run (show what would start)
+./scripts/nightly-pipeline-scheduler.sh --dry-run
+
+# Status report
+./scripts/nightly-pipeline-scheduler.sh --status
+
+# Force start during peak hours
+./scripts/nightly-pipeline-scheduler.sh --force
+```
+
+## Configuration
+
+Set via environment variables:
+- `PIPELINE_TOKEN_LIMIT`: Daily token budget (default: 500,000)
+- `PIPELINE_PEAK_START`: Peak hours start (default: 9)
+- `PIPELINE_PEAK_END`: Peak hours end (default: 18)
+- `HERMES_HOME`: Hermes home directory (default: ~/.hermes)
+
+## Cron
+
+Runs every 30 minutes. Off-peak only (unless --force).
+See `cron/pipeline-scheduler.yml`.

scripts/nightly-pipeline-scheduler.sh

@@ -0,0 +1,383 @@
+#!/usr/bin/env bash
+# nightly-pipeline-scheduler.sh — Auto-start batch pipelines when inference is available.
+#
+# Checks provider health, pipeline progress, token budget, and interactive load.
+# Starts the highest-priority incomplete pipeline that can run.
+#
+# Usage:
+#   ./scripts/nightly-pipeline-scheduler.sh          # Normal run
+#   ./scripts/nightly-pipeline-scheduler.sh --dry-run # Show what would start
+#   ./scripts/nightly-pipeline-scheduler.sh --status  # Pipeline status report
+
+set -euo pipefail
+
+# --- Configuration ---
+HERMES_HOME="${HERMES_HOME:-$HOME/.hermes}"
+BUDGET_FILE="${HERMES_HOME}/pipeline_budget.json"
+STATE_FILE="${HERMES_HOME}/pipeline_state.json"
+LOG_FILE="${HERMES_HOME}/logs/pipeline-scheduler.log"
+TOKEN_DAILY_LIMIT="${PIPELINE_TOKEN_LIMIT:-500000}"
+PEAK_HOURS_START="${PIPELINE_PEAK_START:-9}"
+PEAK_HOURS_END="${PIPELINE_PEAK_END:-18}"
+
+# Pipeline definitions (priority order)
+# Each pipeline: name, script, max_tokens, dependencies
+PIPELINES=(
+    "playground-factory|scripts/pipeline_playground_factory.sh|100000|none"
+    "training-factory|scripts/pipeline_training_factory.sh|150000|none"
+    "knowledge-mine|scripts/pipeline_knowledge_mine.sh|80000|training-factory"
+    "adversary|scripts/pipeline_adversary.sh|50000|knowledge-mine"
+    "codebase-genome|scripts/pipeline_codebase_genome.sh|120000|none"
+)
+
+# --- Colors ---
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[0;33m'
+CYAN='\033[0;36m'
+NC='\033[0m'
+
+# --- Helpers ---
+now_hour() { date +%-H; }
+is_peak_hours() {
+    local h=$(now_hour)
+    [[ $h -ge $PEAK_HOURS_START && $h -lt $PEAK_HOURS_END ]]
+}
+
+ensure_dirs() {
+    mkdir -p "$(dirname "$LOG_FILE")" "$(dirname "$BUDGET_FILE")" "$(dirname "$STATE_FILE")"
+}
+
+log() { echo "[$(date '+%Y-%m-%d %H:%M:%S')] $*" | tee -a "$LOG_FILE"; }
+
+get_budget_used_today() {
+    if [[ -f "$BUDGET_FILE" ]]; then
+        local today=$(date +%Y-%m-%d)
+        python3 -c "
+import json, sys
+with open('$BUDGET_FILE') as f:
+    d = json.load(f)
+print(d.get('daily', {}).get('$today', {}).get('tokens_used', 0))
+" 2>/dev/null || echo 0
+    else
+        echo 0
+    fi
+}
+
+get_budget_remaining() {
+    local used=$(get_budget_used_today)
+    echo $((TOKEN_DAILY_LIMIT - used))
+}
+
+update_budget() {
+    local pipeline="$1"
+    local tokens="$2"
+    local today=$(date +%Y-%m-%d)
+    python3 -c "
+import json, os
+path = '$BUDGET_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+daily = d.setdefault('daily', {})
+day = daily.setdefault('$today', {'tokens_used': 0, 'pipelines': {}})
+day['tokens_used'] = day.get('tokens_used', 0) + $tokens
+day['pipelines']['$pipeline'] = day['pipelines'].get('$pipeline', 0) + $tokens
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+}
+
+get_pipeline_state() {
+    if [[ -f "$STATE_FILE" ]]; then
+        cat "$STATE_FILE"
+    else
+        echo "{}"
+    fi
+}
+
+set_pipeline_state() {
+    local pipeline="$1"
+    local state="$2"  # running, complete, failed, skipped
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+d['$pipeline'] = {'state': '$state', 'updated': '$(date -Iseconds)'}
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+}
+
+is_pipeline_complete() {
+    local pipeline="$1"
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('false')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    state = d.get('$pipeline', {}).get('state', 'not_started')
+    print('true' if state == 'complete' else 'false')
+" 2>/dev/null || echo false
+}
+
+is_pipeline_running() {
+    local pipeline="$1"
+    python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('false')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    state = d.get('$pipeline', {}).get('state', 'not_started')
+    print('true' if state == 'running' else 'false')
+" 2>/dev/null || echo false
+}
+
+check_dependency() {
+    local dep="$1"
+    if [[ "$dep" == "none" ]]; then
+        return 0
+    fi
+    # For knowledge-mine: training-factory must be running or complete
+    if [[ "$dep" == "training-factory" ]]; then
+        local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('training-factory', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+        [[ "$state" == "running" || "$state" == "complete" ]]
+        return $?
+    fi
+    # For adversary: knowledge-mine must be at least 50% done
+    # Simplified: check if it's running (we'd need progress tracking for 50%)
+    if [[ "$dep" == "knowledge-mine" ]]; then
+        local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('knowledge-mine', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+        [[ "$state" == "running" || "$state" == "complete" ]]
+        return $?
+    fi
+    return 0
+}
+
+check_inference_available() {
+    # Check if any inference provider is responding
+    # 1. Check OpenRouter
+    local or_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+        --connect-timeout 5 "https://openrouter.ai/api/v1/models" 2>/dev/null || echo "000")
+
+    # 2. Check local Ollama
+    local ollama_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+        --connect-timeout 5 "http://localhost:11434/api/tags" 2>/dev/null || echo "000")
+
+    # 3. Check RunPod (if configured)
+    local runpod_ok="000"
+    if [[ -n "${RUNPOD_ENDPOINT:-}" ]]; then
+        runpod_ok=$(curl -s -o /dev/null -w "%{http_code}" \
+            --connect-timeout 5 "$RUNPOD_ENDPOINT/health" 2>/dev/null || echo "000")
+    fi
+
+    if [[ "$or_ok" == "200" || "$ollama_ok" == "200" || "$runpod_ok" == "200" ]]; then
+        return 0
+    fi
+    return 1
+}
+
+check_interactive_load() {
+    # Check if there are active interactive sessions (don't fight with live users)
+    # Look for tmux panes with active hermes sessions
+    local active=$(tmux list-panes -a -F '#{pane_pid} #{pane_current_command}' 2>/dev/null \
+        | grep -c "hermes\|python3" || echo 0)
+
+    # If more than 3 interactive sessions, skip pipeline start
+    if [[ $active -gt 3 ]]; then
+        return 1
+    fi
+    return 0
+}
+
+start_pipeline() {
+    local name="$1"
+    local script="$2"
+    local max_tokens="$3"
+    local budget_remaining="$4"
+    local mode="${5:-run}"
+
+    if [[ "$budget_remaining" -lt "$max_tokens" ]]; then
+        log "SKIP $name: insufficient budget ($budget_remaining < $max_tokens tokens)"
+        return 1
+    fi
+
+    if [[ ! -f "$script" ]]; then
+        log "SKIP $name: script not found ($script)"
+        return 1
+    fi
+
+    if [[ "$mode" == "dry-run" ]]; then
+        log "DRY-RUN: Would start $name (budget: $budget_remaining, needs: $max_tokens)"
+        return 0
+    fi
+
+    log "START $name (budget: $budget_remaining, max_tokens: $max_tokens)"
+    set_pipeline_state "$name" "running"
+
+    # Run in background, capture output
+    local log_path="${HERMES_HOME}/logs/pipeline-${name}.log"
+    bash "$script" --max-tokens "$max_tokens" >> "$log_path" 2>&1 &
+    local pid=$!
+
+    # Wait a moment to check if it started OK
+    sleep 2
+    if kill -0 $pid 2>/dev/null; then
+        log "RUNNING $name (PID: $pid, log: $log_path)"
+        # Record the PID
+        python3 -c "
+import json, os
+path = '$STATE_FILE'
+d = {}
+if os.path.exists(path):
+    with open(path) as f:
+        d = json.load(f)
+d['$name']['pid'] = $pid
+with open(path, 'w') as f:
+    json.dump(d, f, indent=2)
+"
+        return 0
+    else
+        log "FAIL $name: script exited immediately"
+        set_pipeline_state "$name" "failed"
+        return 1
+    fi
+}
+
+# --- Main ---
+main() {
+    local mode="${1:-run}"
+    ensure_dirs
+
+    log "=== Pipeline Scheduler ($mode) ==="
+
+    # Check 1: Is inference available?
+    if ! check_inference_available; then
+        log "No inference provider available. Skipping all pipelines."
+        exit 0
+    fi
+    log "Inference: AVAILABLE"
+
+    # Check 2: Is it peak hours?
+    if is_peak_hours && [[ "$mode" != "--force" ]]; then
+        local h=$(now_hour)
+        log "Peak hours ($h:00). Skipping pipeline start. Use --force to override."
+        exit 0
+    fi
+    log "Off-peak: OK"
+
+    # Check 3: Interactive load
+    if ! check_interactive_load && [[ "$mode" != "--force" ]]; then
+        log "High interactive load. Skipping pipeline start."
+        exit 0
+    fi
+    log "Interactive load: OK"
+
+    # Check 4: Token budget
+    local budget=$(get_budget_remaining)
+    log "Token budget remaining: $budget / $TOKEN_DAILY_LIMIT"
+
+    if [[ $budget -le 0 ]]; then
+        log "Daily token budget exhausted. Stopping."
+        exit 0
+    fi
+
+    # Check 5: Pipeline status
+    if [[ "$mode" == "--status" ]]; then
+        echo -e "${CYAN}Pipeline Status:${NC}"
+        echo "────────────────────────────────────────────────────"
+        for entry in "${PIPELINES[@]}"; do
+            IFS='|' read -r name script max_tokens dep <<< "$entry"
+            local state=$(python3 -c "
+import json, os
+path = '$STATE_FILE'
+if not os.path.exists(path):
+    print('not_started')
+else:
+    with open(path) as f:
+        d = json.load(f)
+    print(d.get('$name', {}).get('state', 'not_started'))
+" 2>/dev/null || echo "not_started")
+
+            local color=$NC
+            case "$state" in
+                running)  color=$YELLOW ;;
+                complete) color=$GREEN ;;
+                failed)   color=$RED ;;
+            esac
+            printf "  %-25s %b%s%b (max: %s tokens, dep: %s)\n" "$name" "$color" "$state" "$NC" "$max_tokens" "$dep"
+        done
+        echo "────────────────────────────────────────────────────"
+        echo "  Budget: $budget / $TOKEN_DAILY_LIMIT tokens remaining"
+        echo "  Peak hours: $PEAK_HOURS_START:00 - $PEAK_HOURS_END:00"
+        exit 0
+    fi
+
+    # Find and start the highest-priority incomplete pipeline
+    local started=0
+    for entry in "${PIPELINES[@]}"; do
+        IFS='|' read -r name script max_tokens dep <<< "$entry"
+
+        # Skip if already running or complete
+        if [[ "$(is_pipeline_running $name)" == "true" ]]; then
+            log "SKIP $name: already running"
+            continue
+        fi
+        if [[ "$(is_pipeline_complete $name)" == "true" ]]; then
+            log "SKIP $name: already complete"
+            continue
+        fi
+
+        # Check dependency
+        if ! check_dependency "$dep"; then
+            log "SKIP $name: dependency $dep not met"
+            continue
+        fi
+
+        # Try to start
+        if start_pipeline "$name" "$script" "$max_tokens" "$budget" "$mode"; then
+            started=1
+            # Only start one pipeline per run (let it claim tokens before next check)
+            # Exception: playground-factory and training-factory can run in parallel
+            if [[ "$name" != "playground-factory" && "$name" != "training-factory" ]]; then
+                break
+            fi
+        fi
+    done
+
+    if [[ $started -eq 0 ]]; then
+        log "No pipelines to start (all complete, running, or blocked)."
+    fi
+
+    log "=== Pipeline Scheduler done ==="
+}
+
+main "$@"

scripts/visual_pr_reviewer.py

@@ -1,3 +1,4 @@
+#!/usr/bin/env python3
 import json
 from hermes_tools import browser_navigate, browser_vision