feat: add training data quality filter (#687)

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/training_data_quality_filter.py

@@ -0,0 +1,286 @@
+#!/usr/bin/env python3
+"""
+Training Data Quality Filter
+
+Scores and removes low-quality training pairs from JSONL datasets.
+Supports two formats:
+  - ShareGPT session format: {"conversations": [...], ...}
+  - Scene/pair format: {"terse": "...", "rich": "..."} or {"lyric_line": "...", "scene": {...}}
+
+Scoring dimensions:
+  - Specificity: penalizes vague/generic content
+  - Length ratio: penalizes extreme input/output imbalances
+  - Code correctness: validates code blocks have matching fences
+
+Usage:
+  python3 scripts/training_data_quality_filter.py input.jsonl [--threshold 0.4] [--output filtered.jsonl]
+  python3 scripts/training_data_quality_filter.py --dir training-data/ [--threshold 0.4]
+"""
+
+import argparse
+import json
+import re
+import sys
+from pathlib import Path
+
+
+def score_specificity(text: str) -> float:
+    """Score 0-1 based on how specific vs generic the text is."""
+    if not text or len(text.strip()) < 10:
+        return 0.0
+
+    score = 0.5  # baseline
+
+    # Penalize very generic starters
+    generic_starters = [
+        "sure,", "of course", "i can help", "here is", "here are",
+        "certainly", "absolutely", "let me help", "great question",
+        "that\'s a great", "interesting question",
+    ]
+    lower = text.lower().strip()
+    for starter in generic_starters:
+        if lower.startswith(starter):
+            score -= 0.15
+            break
+
+    # Reward specific content indicators
+    if re.search(r"`[^`]+`", text):  # inline code
+        score += 0.1
+    if re.search(r"```[\s\S]*?```", text):  # code blocks
+        score += 0.15
+    if re.search(r"\d+\.\s", text):  # numbered lists
+        score += 0.05
+    if len(text.split()) > 50:  # substantial length
+        score += 0.1
+    if re.search(r"https?://", text):  # URLs/references
+        score += 0.05
+
+    # Penalize extremely short outputs
+    if len(text.split()) < 5:
+        score -= 0.2
+
+    # Penalize repetition (same sentence repeated)
+    sentences = re.split(r"[.!?]+", text)
+    sentences = [s.strip().lower() for s in sentences if s.strip()]
+    if sentences:
+        unique_ratio = len(set(sentences)) / len(sentences)
+        if unique_ratio < 0.7:
+            score -= 0.15
+
+    return max(0.0, min(1.0, score))
+
+
+def score_length_ratio(input_text: str, output_text: str) -> float:
+    """Score 0-1 based on input/output length balance."""
+    in_len = len(input_text.split())
+    out_len = len(output_text.split())
+
+    if in_len == 0 or out_len == 0:
+        return 0.0
+
+    ratio = out_len / in_len
+
+    # Ideal ratio: 0.5-5x (output can be shorter or longer, but not extreme)
+    if 0.5 <= ratio <= 5.0:
+        return 1.0
+    elif 0.2 <= ratio <= 10.0:
+        return 0.6
+    elif 0.1 <= ratio <= 20.0:
+        return 0.3
+    else:
+        return 0.1
+
+
+def score_code_correctness(text: str) -> float:
+    """Score 0-1 based on code block correctness."""
+    code_blocks = re.findall(r"```[\s\S]*?```", text)
+
+    if not code_blocks:
+        return 1.0  # no code = no code errors
+
+    for block in code_blocks:
+        # Check balanced fences
+        fence_count = block.count("```")
+        if fence_count % 2 != 0:
+            return 0.2
+
+        # Check for common errors
+        content = block.split("\n", 1)[-1] if "\n" in block else ""
+        if "SyntaxError" in content or "Traceback" in content:
+            return 0.3
+        if content.strip().endswith("...") and len(content.strip()) < 30:
+            return 0.4  # truncated code
+
+    return 1.0
+
+
+def score_pair(input_text: str, output_text: str) -> dict:
+    """Score a training pair on all dimensions."""
+    spec = score_specificity(output_text)
+    length = score_length_ratio(input_text, output_text)
+    code = score_code_correctness(output_text)
+
+    # Weighted composite
+    composite = (spec * 0.4) + (length * 0.3) + (code * 0.3)
+
+    return {
+        "specificity": round(spec, 3),
+        "length_ratio": round(length, 3),
+        "code_correctness": round(code, 3),
+        "composite": round(composite, 3),
+    }
+
+
+def extract_pairs(obj: dict) -> list:
+    """Extract (input, output) pairs from a JSONL object."""
+    pairs = []
+
+    # ShareGPT session format
+    if "conversations" in obj:
+        convs = obj["conversations"]
+        for i, msg in enumerate(convs):
+            if msg.get("from") in ("gpt", "assistant"):
+                # Find preceding human message
+                input_text = ""
+                for j in range(i - 1, -1, -1):
+                    if convs[j].get("from") == "human":
+                        input_text = convs[j].get("value", "")
+                        break
+                output_text = msg.get("value", "")
+                if input_text and output_text:
+                    pairs.append((input_text, output_text))
+
+    # Scene/pair format (terse/rich)
+    elif "terse" in obj and "rich" in obj:
+        pairs.append((obj["terse"], obj["rich"]))
+
+    # Scene description format
+    elif "lyric_line" in obj and "scene" in obj:
+        scene_text = json.dumps(obj["scene"]) if isinstance(obj["scene"], dict) else str(obj["scene"])
+        pairs.append((obj["lyric_line"], scene_text))
+
+    # Generic prompt/response
+    elif "prompt" in obj and "response" in obj:
+        pairs.append((obj["prompt"], obj["response"]))
+
+    # Generic input/output
+    elif "input" in obj and "output" in obj:
+        pairs.append((obj["input"], obj["output"]))
+
+    return pairs
+
+
+def filter_jsonl(input_path: str, threshold: float = 0.4, output_path: str = None) -> dict:
+    """Filter a JSONL file, removing low-quality pairs."""
+    path = Path(input_path)
+    if not path.exists():
+        return {"error": f"File not found: {input_path}"}
+
+    lines = path.read_text().strip().split("\n")
+    total = 0
+    kept = 0
+    removed = 0
+    scores_list = []
+    kept_lines = []
+
+    for line in lines:
+        line = line.strip()
+        if not line:
+            continue
+
+        try:
+            obj = json.loads(line)
+        except json.JSONDecodeError:
+            removed += 1
+            continue
+
+        pairs = extract_pairs(obj)
+        total += 1
+
+        if not pairs:
+            # No extractable pairs — keep as-is (might be metadata)
+            kept += 1
+            kept_lines.append(line)
+            continue
+
+        # Score all pairs in this object
+        pair_scores = [score_pair(inp, out) for inp, out in pairs]
+        avg_composite = sum(s["composite"] for s in pair_scores) / len(pair_scores)
+
+        scores_list.append(avg_composite)
+
+        if avg_composite >= threshold:
+            kept += 1
+            kept_lines.append(line)
+        else:
+            removed += 1
+
+    # Write output
+    if output_path:
+        Path(output_path).write_text("\n".join(kept_lines) + "\n")
+
+    return {
+        "file": input_path,
+        "total": total,
+        "kept": kept,
+        "removed": removed,
+        "removal_rate": f"{removed}/{total}" if total > 0 else "0/0",
+        "avg_score": round(sum(scores_list) / len(scores_list), 3) if scores_list else None,
+        "min_score": round(min(scores_list), 3) if scores_list else None,
+        "max_score": round(max(scores_list), 3) if scores_list else None,
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Filter low-quality training data pairs")
+    parser.add_argument("input", nargs="?", help="Input JSONL file")
+    parser.add_argument("--threshold", type=float, default=0.4, help="Minimum quality score (0-1)")
+    parser.add_argument("--output", "-o", help="Output file (default: input_filtered.jsonl)")
+    parser.add_argument("--dir", help="Process all .jsonl files in directory")
+    parser.add_argument("--dry-run", action="store_true", help="Score only, don\'t write output")
+    args = parser.parse_args()
+
+    if args.dir:
+        dirpath = Path(args.dir)
+        jsonl_files = sorted(dirpath.rglob("*.jsonl"))
+        if not jsonl_files:
+            print(f"No .jsonl files found in {args.dir}")
+            sys.exit(1)
+
+        print(f"Processing {len(jsonl_files)} files (threshold={args.threshold})\n")
+        print(f"{'File':<50} {'Total':>6} {'Kept':>6} {'Removed':>8} {'Avg':>6}")
+        print("-" * 82)
+
+        grand_total = grand_kept = grand_removed = 0
+        for fpath in jsonl_files:
+            out = str(fpath).replace(".jsonl", "_filtered.jsonl") if not args.dry_run else None
+            result = filter_jsonl(str(fpath), args.threshold, out)
+            if "error" in result:
+                print(f"{str(fpath):<50} ERROR: {result['error']}")
+                continue
+            print(f"{fpath.name:<50} {result['total']:>6} {result['kept']:>6} {result['removed']:>8} {result['avg_score']:>6.3f}")
+            grand_total += result["total"]
+            grand_kept += result["kept"]
+            grand_removed += result["removed"]
+
+        print("-" * 82)
+        print(f"{'TOTAL':<50} {grand_total:>6} {grand_kept:>6} {grand_removed:>8}")
+
+    elif args.input:
+        out = args.output or args.input.replace(".jsonl", "_filtered.jsonl")
+        if args.dry_run:
+            out = None
+        result = filter_jsonl(args.input, args.threshold, out)
+        if "error" in result:
+            print(f"Error: {result['error']}")
+            sys.exit(1)
+        print(json.dumps(result, indent=2))
+        if out:
+            print(f"\nFiltered output written to: {out}")
+    else:
+        parser.print_help()
+        sys.exit(1)
+
+
+if __name__ == "__main__":
+    main()

scripts/visual_pr_reviewer.py

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