harness-evolver 3.3.1 → 4.0.2

package/tools/adversarial_inject.py

@@ -0,0 +1,205 @@
+#!/usr/bin/env python3
+"""Inject adversarial examples into LangSmith dataset.
+
+Detects potential memorization by checking if agent outputs are suspiciously
+similar to reference outputs, then generates adversarial variations to test
+generalization.
+
+Usage:
+    python3 adversarial_inject.py \
+        --config .evolver.json \
+        --experiment v003a \
+        --output adversarial_report.json
+"""
+
+import argparse
+import json
+import os
+import platform
+import sys
+import random
+
+
+def ensure_langsmith_api_key():
+    """Load LANGSMITH_API_KEY from credentials file or .env if not in env."""
+    if os.environ.get("LANGSMITH_API_KEY"):
+        return True
+    if platform.system() == "Darwin":
+        creds_path = os.path.expanduser("~/Library/Application Support/langsmith-cli/credentials")
+    else:
+        creds_path = os.path.expanduser("~/.config/langsmith-cli/credentials")
+    if os.path.exists(creds_path):
+        try:
+            with open(creds_path) as f:
+                for line in f:
+                    line = line.strip()
+                    if line.startswith("LANGSMITH_API_KEY="):
+                        key = line.split("=", 1)[1].strip()
+                        if key:
+                            os.environ["LANGSMITH_API_KEY"] = key
+                            return True
+        except OSError:
+            pass
+    if os.path.exists(".env"):
+        try:
+            with open(".env") as f:
+                for line in f:
+                    line = line.strip()
+                    if line.startswith("LANGSMITH_API_KEY=") and not line.startswith("#"):
+                        key = line.split("=", 1)[1].strip().strip("'\"")
+                        if key:
+                            os.environ["LANGSMITH_API_KEY"] = key
+                            return True
+        except OSError:
+            pass
+    return False
+
+
+def detect_memorization(client, experiment_name, dataset_name):
+    """Check if agent outputs are suspiciously similar to reference outputs."""
+    suspicious = []
+    try:
+        runs = list(client.list_runs(project_name=experiment_name, is_root=True, limit=200))
+        examples = {str(e.id): e for e in client.list_examples(dataset_name=dataset_name, limit=500)}
+
+        for run in runs:
+            if not run.reference_example_id:
+                continue
+            example = examples.get(str(run.reference_example_id))
+            if not example or not example.outputs:
+                continue
+
+            run_output = str(run.outputs or "").lower().strip()
+            ref_output = str(example.outputs).lower().strip()
+
+            if not run_output or not ref_output:
+                continue
+
+            if run_output == ref_output:
+                suspicious.append({
+                    "example_id": str(run.reference_example_id),
+                    "match_type": "exact",
+                    "input": str(run.inputs)[:200],
+                })
+            elif len(run_output) > 50 and ref_output in run_output:
+                suspicious.append({
+                    "example_id": str(run.reference_example_id),
+                    "match_type": "contains_reference",
+                    "input": str(run.inputs)[:200],
+                })
+
+    except Exception as e:
+        print(f"Error checking memorization: {e}", file=sys.stderr)
+
+    return suspicious
+
+
+def generate_adversarial_inputs(client, dataset_name, num_inputs=5):
+    """Generate adversarial variations of existing examples.
+
+    Creates multiple variation types to test generalization:
+    - negation: inverts the question to test if the agent distinguishes
+    - constraint: adds a constraint that changes the expected answer
+    - ambiguous: makes the input ambiguous to test robustness
+    - partial: provides incomplete input to test graceful handling
+    """
+    examples = list(client.list_examples(dataset_name=dataset_name, limit=100))
+    if not examples:
+        return []
+
+    adversarial = []
+    sampled = random.sample(examples, min(num_inputs, len(examples)))
+
+    variation_types = [
+        ("negation", "What is NOT the case: {input}"),
+        ("constraint", "{input} Answer in exactly one sentence."),
+        ("ambiguous", "Someone asked something like: {input}"),
+        ("partial", "{partial_input}"),
+    ]
+
+    for example in sampled:
+        input_data = example.inputs or {}
+        input_text = str(input_data.get("input", input_data))
+
+        # Pick a variation type (rotate through them)
+        idx = sampled.index(example) % len(variation_types)
+        vtype, template = variation_types[idx]
+
+        if vtype == "partial":
+            # Use first half of the input
+            words = input_text.split()
+            partial = " ".join(words[:max(len(words) // 2, 3)])
+            varied_input = template.format(partial_input=partial)
+        else:
+            varied_input = template.format(input=input_text)
+
+        adversarial.append({
+            "inputs": {"input": varied_input},
+            "metadata": {
+                "source": "adversarial",
+                "original_example_id": str(example.id),
+                "variation_type": vtype,
+            },
+        })
+
+    return adversarial
+
+
+def inject_adversarial(client, dataset_id, adversarial_inputs):
+    """Add adversarial examples to dataset."""
+    added = 0
+    for adv in adversarial_inputs:
+        try:
+            client.create_example(
+                inputs=adv["inputs"],
+                dataset_id=dataset_id,
+                metadata=adv["metadata"],
+            )
+            added += 1
+        except Exception as e:
+            print(f"Failed to inject: {e}", file=sys.stderr)
+    return added
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Adversarial injection for evaluators")
+    parser.add_argument("--config", default=".evolver.json")
+    parser.add_argument("--experiment", required=True, help="Experiment to check for memorization")
+    parser.add_argument("--output", default=None, help="Output report path")
+    parser.add_argument("--inject", action="store_true", help="Actually inject adversarial examples")
+    parser.add_argument("--num-adversarial", type=int, default=5, help="Number of adversarial examples")
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        config = json.load(f)
+
+    ensure_langsmith_api_key()
+    from langsmith import Client
+    client = Client()
+
+    suspicious = detect_memorization(client, args.experiment, config["dataset"])
+    adversarial = generate_adversarial_inputs(client, config["dataset"], args.num_adversarial)
+
+    injected = 0
+    if args.inject and adversarial:
+        injected = inject_adversarial(client, config["dataset_id"], adversarial)
+
+    result = {
+        "memorization_suspects": len(suspicious),
+        "suspicious_examples": suspicious,
+        "adversarial_generated": len(adversarial),
+        "adversarial_injected": injected,
+    }
+
+    output = json.dumps(result, indent=2)
+    if args.output:
+        with open(args.output, "w") as f:
+            f.write(output)
+    print(output)
+
+    if suspicious:
+        print(f"\nWARNING: {len(suspicious)} examples show potential memorization!", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()

package/tools/consolidate.py

@@ -0,0 +1,235 @@
+#!/usr/bin/env python3
+"""Cross-iteration memory consolidation for Harness Evolver.
+
+Inspired by Claude Code's autoDream pattern. Analyzes evolution history
+to identify recurring patterns, successful strategies, and wasted approaches.
+Produces evolution_memory.md for proposer briefings.
+
+Usage:
+    python3 consolidate.py --config .evolver.json --output evolution_memory.md
+"""
+
+import argparse
+import json
+import os
+import sys
+from datetime import datetime, timezone
+
+
+def orient(config):
+    """Phase 1: Scan current state and history."""
+    history = config.get("history", [])
+    iterations = config.get("iterations", 0)
+    best_score = config.get("best_score", 0)
+    baseline_score = history[0]["score"] if history else 0
+
+    return {
+        "iterations": iterations,
+        "best_score": best_score,
+        "baseline_score": baseline_score,
+        "improvement": best_score - baseline_score,
+        "history": history,
+    }
+
+
+def gather(config, comparison_files):
+    """Phase 2: Extract signals from trace insights and comparisons."""
+    signals = {
+        "winning_strategies": [],
+        "losing_strategies": [],
+        "recurring_failures": {},
+        "score_deltas": [],
+    }
+
+    for comp_file in comparison_files:
+        if not os.path.exists(comp_file):
+            continue
+        try:
+            with open(comp_file) as f:
+                data = json.load(f)
+            comparison = data.get("comparison", data)
+
+            winner = comparison.get("winner", {})
+            if winner:
+                signals["winning_strategies"].append({
+                    "experiment": winner.get("experiment", ""),
+                    "score": winner.get("score", 0),
+                })
+
+            for candidate in comparison.get("all_candidates", []):
+                if candidate.get("experiment") != winner.get("experiment"):
+                    signals["losing_strategies"].append({
+                        "experiment": candidate.get("experiment", ""),
+                        "score": candidate.get("score", 0),
+                    })
+        except (json.JSONDecodeError, OSError):
+            continue
+
+    # Compute score deltas from history
+    history = config.get("history", [])
+    for i in range(1, len(history)):
+        signals["score_deltas"].append({
+            "version": history[i]["version"],
+            "delta": history[i]["score"] - history[i - 1]["score"],
+            "score": history[i]["score"],
+        })
+
+    # Read trace insights for recurring patterns
+    if os.path.exists("trace_insights.json"):
+        try:
+            with open("trace_insights.json") as f:
+                insights = json.load(f)
+            for issue in insights.get("top_issues", []):
+                pattern = issue.get("pattern", issue.get("description", "unknown"))
+                if pattern not in signals["recurring_failures"]:
+                    signals["recurring_failures"][pattern] = 0
+                signals["recurring_failures"][pattern] += 1
+        except (json.JSONDecodeError, OSError):
+            pass
+
+    return signals
+
+
+def consolidate(orientation, signals, existing_memory=None):
+    """Phase 3: Merge signals into consolidated memory."""
+    insights = []
+
+    # Strategy effectiveness
+    winning = signals.get("winning_strategies", [])
+    strategy_map = {"a": "exploit", "b": "explore", "c": "crossover", "d": "failure-targeted-1", "e": "failure-targeted-2"}
+    win_counts = {}
+    for w in winning:
+        exp = w.get("experiment", "")
+        if exp:
+            suffix = exp[-1]
+            name = strategy_map.get(suffix, suffix)
+            win_counts[name] = win_counts.get(name, 0) + 1
+
+    if win_counts:
+        best_strategy = max(win_counts, key=win_counts.get)
+        insights.append({
+            "type": "strategy_effectiveness",
+            "insight": f"Most winning strategy: {best_strategy} ({win_counts[best_strategy]} wins)",
+            "recurrence": win_counts[best_strategy],
+            "data": win_counts,
+        })
+
+    # Recurring failures (only promote if seen 2+ times)
+    recurring = {k: v for k, v in signals.get("recurring_failures", {}).items() if v >= 2}
+    for pattern, count in sorted(recurring.items(), key=lambda x: -x[1]):
+        insights.append({
+            "type": "recurring_failure",
+            "insight": f"Recurring failure ({count}x): {pattern}",
+            "recurrence": count,
+        })
+
+    # Score trajectory
+    deltas = signals.get("score_deltas", [])
+    if deltas:
+        positive = [d for d in deltas if d["delta"] > 0]
+        negative = [d for d in deltas if d["delta"] < 0]
+        stagnant = [d for d in deltas if abs(d["delta"]) < 0.01]
+        insights.append({
+            "type": "trajectory",
+            "insight": f"Score trajectory: {len(positive)} improvements, {len(negative)} regressions, {len(stagnant)} stagnant",
+            "recurrence": len(deltas),
+        })
+
+    # Merge with existing memory (update recurrence counts)
+    if existing_memory:
+        for existing in existing_memory.get("insights", []):
+            found = False
+            for new in insights:
+                if new["type"] == existing["type"] and new["insight"] == existing["insight"]:
+                    new["recurrence"] = max(new["recurrence"], existing.get("recurrence", 1)) + 1
+                    found = True
+                    break
+            if not found and existing.get("recurrence", 1) >= 2:
+                insights.append(existing)
+
+    return insights
+
+
+def prune(insights, max_insights=20):
+    """Phase 4: Cap size, remove stale entries."""
+    sorted_insights = sorted(insights, key=lambda x: -x.get("recurrence", 1))
+    return sorted_insights[:max_insights]
+
+
+def format_memory(orientation, insights):
+    """Format consolidated memory as markdown."""
+    lines = [
+        "# Evolution Memory",
+        "",
+        f"*Last updated: {datetime.now(timezone.utc).strftime('%Y-%m-%d %H:%M UTC')}*",
+        f"*Iterations: {orientation['iterations']} | Best: {orientation['best_score']:.3f} | Baseline: {orientation['baseline_score']:.3f} | Improvement: +{orientation['improvement']:.3f}*",
+        "",
+        "## Key Insights (promoted after 2+ recurrences)",
+        "",
+    ]
+
+    promoted = [i for i in insights if i.get("recurrence", 1) >= 2]
+    other = [i for i in insights if i.get("recurrence", 1) < 2]
+
+    if promoted:
+        for insight in promoted:
+            lines.append(f"- **[{insight['type']}]** {insight['insight']} (seen {insight['recurrence']}x)")
+    else:
+        lines.append("- No insights promoted yet (need 2+ recurrences)")
+
+    if other:
+        lines.append("")
+        lines.append("## Observations (1 recurrence, pending promotion)")
+        lines.append("")
+        for insight in other:
+            lines.append(f"- [{insight['type']}] {insight['insight']}")
+
+    lines.append("")
+    return "\n".join(lines)
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Cross-iteration memory consolidation")
+    parser.add_argument("--config", default=".evolver.json")
+    parser.add_argument("--output", default="evolution_memory.md", help="Output markdown path")
+    parser.add_argument("--output-json", default="evolution_memory.json", help="Output JSON path")
+    parser.add_argument("--comparison-files", nargs="*", default=[], help="Past comparison.json files")
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        config = json.load(f)
+
+    # Load existing memory if present
+    existing = None
+    if os.path.exists(args.output_json):
+        try:
+            with open(args.output_json) as f:
+                existing = json.load(f)
+        except (json.JSONDecodeError, OSError):
+            pass
+
+    # Four-phase consolidation
+    orientation = orient(config)
+    signals = gather(config, args.comparison_files or ["comparison.json"])
+    insights = consolidate(orientation, signals, existing)
+    insights = prune(insights)
+
+    # Write markdown
+    memory_md = format_memory(orientation, insights)
+    with open(args.output, "w") as f:
+        f.write(memory_md)
+
+    # Write JSON for programmatic access
+    memory_json = {
+        "updated_at": datetime.now(timezone.utc).isoformat(),
+        "orientation": orientation,
+        "insights": insights,
+    }
+    with open(args.output_json, "w") as f:
+        json.dump(memory_json, f, indent=2)
+
+    print(memory_md)
+
+
+if __name__ == "__main__":
+    main()

package/tools/iteration_gate.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+"""Three-gate iteration trigger for Harness Evolver.
+
+Evaluates whether the next evolution iteration should proceed based on:
+1. Score gate: skip if no meaningful delta or no clustered failures
+2. Cost gate: estimate token cost, stop if budget exceeded
+3. Convergence gate: detect statistical plateau
+
+Usage:
+    python3 iteration_gate.py --config .evolver.json --output gate_result.json
+"""
+
+import argparse
+import json
+import os
+import sys
+from datetime import datetime, timezone
+
+
+def score_gate(config, threshold=0.02):
+    """Check if there's meaningful room for improvement."""
+    history = config.get("history", [])
+    if len(history) < 2:
+        return {"pass": True, "reason": "Not enough history to evaluate"}
+
+    recent = [h["score"] for h in history[-3:]]
+    best = config.get("best_score", 0)
+    target = config.get("target_score")
+
+    if target and best >= target:
+        return {"pass": False, "reason": f"Target reached: {best:.3f} >= {target}"}
+
+    if len(recent) >= 3:
+        score_range = max(recent) - min(recent)
+        if score_range < threshold:
+            return {
+                "pass": False,
+                "reason": f"Plateau detected: last 3 scores within {score_range:.4f} (threshold: {threshold})",
+                "suggest": "architect",
+            }
+
+    return {"pass": True, "reason": f"Score delta exists: range={max(recent)-min(recent):.4f}"}
+
+
+def cost_gate(config, budget_tokens=None):
+    """Estimate cost of next iteration and check against budget."""
+    history = config.get("history", [])
+    iterations = config.get("iterations", 0)
+    estimated_cost = config.get("iteration_costs", {})
+
+    if not estimated_cost and iterations == 0:
+        return {"pass": True, "reason": "First iteration, no cost data yet"}
+
+    total_spent = sum(estimated_cost.get("per_iteration", [0]))
+    budget = budget_tokens or estimated_cost.get("budget_tokens")
+
+    if not budget:
+        return {"pass": True, "reason": "No budget configured"}
+
+    avg_cost = total_spent / max(iterations, 1)
+    remaining = budget - total_spent
+
+    if remaining < avg_cost * 0.5:
+        return {
+            "pass": False,
+            "reason": f"Budget nearly exhausted: {remaining:,} tokens remaining, avg iteration costs {avg_cost:,.0f}",
+        }
+
+    return {"pass": True, "reason": f"Budget OK: {remaining:,} tokens remaining"}
+
+
+def convergence_gate(config, min_improvement=0.005, lookback=5):
+    """Detect statistical convergence using diminishing returns."""
+    history = config.get("history", [])
+    if len(history) < 3:
+        return {"pass": True, "reason": "Not enough iterations for convergence analysis"}
+
+    recent = history[-lookback:] if len(history) >= lookback else history
+    deltas = []
+    for i in range(1, len(recent)):
+        deltas.append(recent[i]["score"] - recent[i - 1]["score"])
+
+    if not deltas:
+        return {"pass": True, "reason": "No deltas to analyze"}
+
+    avg_delta = sum(deltas) / len(deltas)
+    positive_deltas = [d for d in deltas if d > 0]
+    improvement_rate = len(positive_deltas) / len(deltas)
+
+    if avg_delta < min_improvement and improvement_rate < 0.4:
+        return {
+            "pass": False,
+            "reason": f"Converged: avg delta={avg_delta:.4f}, improvement rate={improvement_rate:.0%}",
+            "suggest": "architect" if improvement_rate < 0.2 else "continue_cautious",
+        }
+
+    return {
+        "pass": True,
+        "reason": f"Still improving: avg delta={avg_delta:.4f}, improvement rate={improvement_rate:.0%}",
+    }
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Three-gate iteration trigger")
+    parser.add_argument("--config", default=".evolver.json", help="Config path")
+    parser.add_argument("--output", default=None, help="Output JSON path")
+    parser.add_argument("--score-threshold", type=float, default=0.02, help="Score plateau threshold")
+    parser.add_argument("--budget-tokens", type=int, default=None, help="Token budget override")
+    args = parser.parse_args()
+
+    with open(args.config) as f:
+        config = json.load(f)
+
+    gates = {
+        "score": score_gate(config, args.score_threshold),
+        "cost": cost_gate(config, args.budget_tokens),
+        "convergence": convergence_gate(config),
+    }
+
+    all_pass = all(g["pass"] for g in gates.values())
+    suggestions = [g.get("suggest") for g in gates.values() if g.get("suggest")]
+
+    result = {
+        "proceed": all_pass,
+        "gates": gates,
+        "suggestions": suggestions,
+        "timestamp": datetime.now(timezone.utc).isoformat(),
+    }
+
+    output = json.dumps(result, indent=2)
+    if args.output:
+        with open(args.output, "w") as f:
+            f.write(output)
+    print(output)
+
+    sys.exit(0 if all_pass else 1)
+
+
+if __name__ == "__main__":
+    main()