feat: add training data quality filter script (#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_quality_filter.py

@@ -0,0 +1,266 @@
+#!/usr/bin/env python3
+"""
+[QUALITY] Training Data Quality Filter
+Part of the Timmy Foundation tooling.
+
+Scores and filters JSONL training pairs on specificity, length ratio,
+and code correctness. Removes low-quality pairs and reports results.
+
+Usage:
+    python3 scripts/training_quality_filter.py input.jsonl -o filtered.jsonl
+    python3 scripts/training_quality_filter.py input.jsonl --threshold 0.4
+    cat input.jsonl | python3 scripts/training_quality_filter.py -
+"""
+
+import sys
+import json
+import argparse
+import re
+from typing import Dict, Any, Tuple
+
+DEFAULT_THRESHOLD = 0.35
+MIN_TERSE_LEN = 3
+MIN_RICH_LEN = 10
+
+
+def score_specificity(terse: str, rich: str) -> float:
+    """Score how specific the rich response is vs the terse prompt.
+    
+    Higher score = more specific, actionable detail in the rich version.
+    """
+    if not terse or not rich:
+        return 0.0
+    
+    # Ratio of unique words (higher = more varied/specific language)
+    rich_words = rich.lower().split()
+    terse_words = terse.lower().split()
+    
+    if len(rich_words) < 3:
+        return 0.1
+    
+    unique_ratio = len(set(rich_words)) / len(rich_words)
+    
+    # Check for concrete details: numbers, file paths, commands, code refs
+    concrete_patterns = [
+        r"\b\d+\b",              # numbers
+        r"[/\\]\w+",            # file paths
+        r"`[^`]+`",               # inline code
+        r"\b(fix|add|remove|update|create|delete|check|run|use)\b",  # action verbs
+    ]
+    concrete_count = sum(
+        len(re.findall(p, rich, re.IGNORECASE)) for p in concrete_patterns
+    )
+    concrete_score = min(concrete_count / 5.0, 1.0)
+    
+    # Length expansion ratio (rich should be meaningfully longer than terse)
+    expansion = len(rich_words) / max(len(terse_words), 1)
+    expansion_score = min(expansion / 5.0, 1.0)
+    
+    return round(0.3 * unique_ratio + 0.4 * concrete_score + 0.3 * expansion_score, 3)
+
+
+def score_length_ratio(terse: str, rich: str) -> float:
+    """Score the length ratio between terse and rich.
+    
+    Too short rich = low quality. Too long = possibly padded.
+    Sweet spot: 3-15x expansion.
+    """
+    if not terse or not rich:
+        return 0.0
+    
+    t_len = len(terse.split())
+    r_len = len(rich.split())
+    
+    if t_len < MIN_TERSE_LEN or r_len < MIN_RICH_LEN:
+        return 0.1
+    
+    ratio = r_len / max(t_len, 1)
+    
+    if ratio < 1.5:
+        return 0.2  # barely expanded
+    elif ratio < 3.0:
+        return 0.5  # some expansion
+    elif ratio <= 15.0:
+        return 1.0  # good expansion
+    elif ratio <= 30.0:
+        return 0.7  # possibly padded
+    else:
+        return 0.4  # very padded
+
+
+def score_code_correctness(terse: str, rich: str) -> float:
+    """Score code blocks in the rich response for basic correctness.
+    
+    Checks for matching brackets, valid-looking syntax patterns.
+    """
+    if not rich:
+        return 0.5  # no code = neutral
+    
+    code_blocks = re.findall(r"```(?:\w*)\n(.*?)```", rich, re.DOTALL)
+    if not code_blocks:
+        return 0.5  # no code blocks = neutral
+    
+    scores = []
+    for block in code_blocks:
+        block_score = 1.0
+        
+        # Check bracket balance
+        for open_c, close_c in [("(", ")"), ("[", "]"), ("{", "}")]:
+            if block.count(open_c) != block.count(close_c):
+                block_score -= 0.3
+        
+        # Check for common syntax errors
+        if re.search(r"def \w+[^:]*\n(?!\s)", block):
+            block_score -= 0.2  # missing colon or body
+        
+        # Minimum viable code length
+        if len(block.strip()) < 10:
+            block_score -= 0.3
+        
+        scores.append(max(block_score, 0.0))
+    
+    return round(sum(scores) / len(scores), 3) if scores else 0.5
+
+
+def score_pair(pair: Dict[str, Any]) -> Tuple[float, Dict[str, float]]:
+    """Score a single training pair. Returns (total_score, breakdown)."""
+    terse = pair.get("terse", "") or pair.get("prompt", "") or ""
+    rich = pair.get("rich", "") or pair.get("response", "") or ""
+    
+    spec = score_specificity(terse, rich)
+    length = score_length_ratio(terse, rich)
+    code = score_code_correctness(terse, rich)
+    
+    # Weighted total
+    total = round(0.4 * spec + 0.3 * length + 0.3 * code, 3)
+    
+    return total, {"specificity": spec, "length_ratio": length, "code_correctness": code}
+
+
+def filter_pairs(input_path: str, output_path: str, threshold: float, 
+                 report: bool = False) -> Dict[str, Any]:
+    """Filter JSONL training pairs by quality score."""
+    kept = []
+    removed = []
+    errors = 0
+    
+    source = sys.stdin if input_path == "-" else open(input_path, "r")
+    
+    try:
+        for line_num, line in enumerate(source, 1):
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                pair = json.loads(line)
+            except json.JSONDecodeError:
+                errors += 1
+                continue
+            
+            score, breakdown = score_pair(pair)
+            entry = {**pair, "_quality_score": score, "_quality_breakdown": breakdown}
+            
+            if score >= threshold:
+                kept.append(entry)
+            else:
+                removed.append(entry)
+    finally:
+        if source is not sys.stdin:
+            source.close()
+    
+    # Write filtered output
+    if output_path:
+        out = sys.stdout if output_path == "-" else open(output_path, "w")
+        try:
+            for pair in kept:
+                # Strip internal scoring fields before output
+                clean = {k: v for k, v in pair.items() if not k.startswith("_quality")}
+                out.write(json.dumps(clean, ensure_ascii=False) + "\n")
+        finally:
+            if out is not sys.stdin:
+                out.close()
+    
+    result = {
+        "total": len(kept) + len(removed),
+        "kept": len(kept),
+        "filtered_out": len(removed),
+        "errors": errors,
+        "threshold": threshold,
+        "filter_rate": round(len(removed) / max(len(kept) + len(removed), 1) * 100, 1),
+    }
+    
+    if report and removed:
+        # Show worst offenders
+        removed_sorted = sorted(removed, key=lambda x: x["_quality_score"])
+        result["worst_5"] = [
+            {
+                "score": e["_quality_score"],
+                "terse": (e.get("terse", "") or e.get("prompt", ""))[:80],
+                "breakdown": e["_quality_breakdown"],
+            }
+            for e in removed_sorted[:5]
+        ]
+    
+    return result
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Filter training data pairs by quality")
+    parser.add_argument("input", help="Input JSONL file (use - for stdin)")
+    parser.add_argument("-o", "--output", default="-", help="Output JSONL file (default: stdout)")
+    parser.add_argument("-t", "--threshold", type=float, default=DEFAULT_THRESHOLD,
+                        help=f"Quality threshold (0.0-1.0, default: {DEFAULT_THRESHOLD})")
+    parser.add_argument("--report", action="store_true", help="Show quality report")
+    parser.add_argument("--dry-run", action="store_true", help="Score only, dont filter")
+    
+    args = parser.parse_args()
+    
+    if args.dry_run:
+        # Just score and report, no filtering
+        source = sys.stdin if args.input == "-" else open(args.input, "r")
+        scores = []
+        try:
+            for line in source:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    pair = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                score, breakdown = score_pair(pair)
+                scores.append(score)
+        finally:
+            if source is not sys.stdin:
+                source.close()
+        
+        if scores:
+            avg = sum(scores) / len(scores)
+            below = sum(1 for s in scores if s < args.threshold)
+            print(f"Total pairs: {len(scores)}")
+            print(f"Average score: {avg:.3f}")
+            print(f"Below threshold ({args.threshold}): {below} ({below/len(scores)*100:.1f}%)")
+            print(f"Min: {min(scores):.3f}  Max: {max(scores):.3f}  Median: {sorted(scores)[len(scores)//2]:.3f}")
+        return
+    
+    result = filter_pairs(args.input, args.output, args.threshold, report=args.report)
+    
+    print(f"Training Data Quality Filter", file=sys.stderr)
+    print(f"{'='*40}", file=sys.stderr)
+    print(f"Total pairs:     {result['total']}", file=sys.stderr)
+    print(f"Kept:            {result['kept']}", file=sys.stderr)
+    print(f"Filtered out:    {result['filtered_out']} ({result['filter_rate']}%)", file=sys.stderr)
+    print(f"Errors:          {result['errors']}", file=sys.stderr)
+    print(f"Threshold:       {result['threshold']}", file=sys.stderr)
+    
+    if args.report and "worst_5" in result:
+        print(f"\nWorst 5 pairs:", file=sys.stderr)
+        for w in result["worst_5"]:
+            terse_preview = w["terse"][:60]
+            print(f"  [{w['score']:.3f}] {terse_preview}...", file=sys.stderr)
+            bd = w["breakdown"]
+            print(f"         spec={bd['specificity']} length={bd['length_ratio']} code={bd['code_correctness']}", file=sys.stderr)
+
+
+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