Analyze Project Skill Security Audit

299300    for path in python_files:301        text = path.read_text(encoding="utf-8", errors="ignore")302        rel = path.relative_to(repo).as_posix()303        lower = text.lower()304305        if "attention" in lower or "transformer" in lower:306            saw_attention = True307        if any(token in lower for token in ["positional", "position_embedding", "position encoding", "pos_embed"]):308            saw_position = True309        if "sigmoid" in lower and lower.count("sigmoid") >= 2:310            findings.append(f"{rel}: repeated `sigmoid` usage detected; review for duplicated post-processing.")311        if "relu" in lower and "sigmoid" in lower:312            findings.append(f"{rel}: both `relu` and `sigmoid` appear in the same file; check activation order and intent.")313        if ".eval()" in lower and "dropout" in lower:314            findings.append(f"{rel}: review whether dropout-sensitive evaluation behavior is intentional.")315        if "optimizer" in lower and "requires_grad" not in lower and "param_groups" not in lower:

1291            )1292        else:1293            run_data = maybe_run_command(1294                repo_path,1295                chosen["documented_command"],1296                args.timeout,1297                args.user_language,1298                args.shell_mode,1299                runtime_root,1300                model_adapter,1301                args.monitor_gpu,1302            )1303

486487    try:488        argv = build_command(command, shell_mode)489        environment = dict(os.environ if child_env is None else child_env)490        spec["requested_argv"] = list(argv)491        atomic_write_json(run_dir / "spec.json", spec)492        if shell_mode == "direct":493            argv = resolve_direct_argv(argv, repo, environment)494        spec["argv"] = list(argv)495        atomic_write_json(run_dir / "spec.json", spec)496        creationflags = subprocess.CREATE_NEW_PROCESS_GROUP if os.name == "nt" else 0497        process = subprocess.Popen(498            argv,499            env=environment,500            cwd=repo,501            stdout=subprocess.PIPE,502            stderr=subprocess.PIPE,503            text=True,504            encoding="utf-8",505            errors="replace",506            bufsize=1,507            creationflags=creationflags,508            start_new_session=os.name != "nt",509        )510    except (FileNotFoundError, ShellSyntaxRequired, OSError, ValueError) as exc:

11661167    scan_data = run_json(scan_script, ["--repo", str(repo_path), "--json"])1168    readme_path = scan_data.get("readme_path")1169    command_data: Dict[str, Any] = {"commands": [], "counts": {}, "warnings": []}1170    if readme_path:1171        command_data = run_json(extract_script, ["--readme", readme_path, "--json"])1172        command_data = delegate_to_docs(readme_path, extract_script, command_data)1173

1271        )1272    elif args.run_selected:1273        if chosen["selected_goal"] == "training":1274            run_data = maybe_run_training(1275                repo_path=repo_path,1276                command=chosen["documented_command"],1277                train_script=train_execute_script,1278                lane=args.lane,1279                user_language=args.user_language,1280                full_training_authorized=args.full_training_authorized,1281                train_timeout=args.train_timeout,1282                dataset_hint=dataset_hint,1283                checkpoint_hint=checkpoint_hint,1284                resume_from=args.resume_from,1285                max_train_steps=args.max_train_steps,1286                shell_mode=args.shell_mode,1287                runtime_root=runtime_root,1288                model_profile_json=args.model_profile_json,1289                required_model_capabilities=args.require_model_capability,1290                gpu_monitor_enabled=not args.no_gpu_monitor,1291            )1292        else:1293            run_data = maybe_run_command(1294                repo_path,1295                chosen["documented_command"],1296                args.timeout,1297                args.user_language,1298                args.shell_mode,1299                runtime_root,1300                model_adapter,1301                args.monitor_gpu,1302            )1303

494                        value = {"error": str(exc)}495                    event("tool_result", tool=name, result=value)496                    state["tool_results"].append({"type": "tool_result", "tool_use_id": call["id"],497                        "content": json.dumps(value, ensure_ascii=False), "is_error": "error" in value})498                    state["pending"].pop(0)499                    tools_this_turn += 1500                    save()501                    if pause_after_tools and tools_this_turn >= pause_after_tools and state["status"] == "running":502                        state["status"] = "paused"503                        event("paused", reason="Explicit test/session checkpoint")504                    continue505                if state["tool_results"]:506                    state["messages"].append({"role": "user", "content": state.pop("tool_results")})507                    state["tool_results"] = []508                request_bytes = len(json.dumps([SYSTEM, TOOLS, state["messages"]]).encode())509                used = sum(state["usage"].values())510                reserve = request_bytes + budget["max_output_tokens"] + 1024511                if state["model_calls"] >= budget["max_model_calls"] or used + reserve > budget["max_total_tokens"]:512                    block("Model call/token reservation budget reached")513                    break514                state["model_calls"] += 1515                state["model_pending"] = True516                state["usage_complete"] = False517                save()518                event("model_request", call=state["model_calls"], reserved_tokens=reserve)519                response = provider.complete(state["messages"], SYSTEM, TOOLS, budget["max_output_tokens"], min(60, remaining))520                if not isinstance(response, dict):

109110`endpoint` optionally names the final HTTPS endpoint. Without it, the client111uses `ANTHROPIC_BASE_URL` or the official endpoint. For an already configured112Bearer gateway, set `metadata.auth_scheme` to `bearer` and name its credential113environment variable. Redirects are refused so credentials are not forwarded.114

569    args = parser.parse_args()570    profile = load_model_profile(Path(args.model_profile))571    if profile["provider"] != "anthropic":572        parser.error("P1 supports the Anthropic Messages protocol; other adapters remain metadata-only")573    task = json.loads(Path(args.task).read_text(encoding="utf-8-sig"))574    state = run(task, Path(args.repo), Path(args.output), profile, AnthropicProvider(profile),575                resume=args.resume, pause_after_tools=args.pause_after_tools, source_adjacent_readme=args.source_adjacent_readme)576    print(json.dumps({**{k: state[k] for k in ["status", "model_calls", "tool_calls", "usage", "verification"]},

246247    for rel in candidate_paths[:24]:248        path = repo / rel249        if not path.exists() or path.suffix.lower() != ".py":250            continue251        try:252            tree = ast.parse(path.read_text(encoding="utf-8", errors="ignore"))253        except SyntaxError:254            continue255        for node in ast.walk(tree):256            if isinstance(node, ast.ClassDef):257                symbol_hints.append(f"{rel}:{node.name}")258                has_init = any(isinstance(item, ast.FunctionDef) and item.name == "__init__" for item in node.body)259                has_forward = any(isinstance(item, ast.FunctionDef) and item.name == "forward" for item in node.body)260                if has_init:261                    constructor_candidates.append(f"{rel}:{node.name}")262                if has_forward:263                    forward_candidates.append(f"{rel}:{node.name}.forward")264            elif isinstance(node, ast.FunctionDef):265                symbol_hints.append(f"{rel}:{node.name}")266                if node.name in {"forward", "__call__", "predict"}:

308            saw_position = True309        if "sigmoid" in lower and lower.count("sigmoid") >= 2:310            findings.append(f"{rel}: repeated `sigmoid` usage detected; review for duplicated post-processing.")311        if "relu" in lower and "sigmoid" in lower:312            findings.append(f"{rel}: both `relu` and `sigmoid` appear in the same file; check activation order and intent.")313        if ".eval()" in lower and "dropout" in lower:314            findings.append(f"{rel}: review whether dropout-sensitive evaluation behavior is intentional.")315        if "optimizer" in lower and "requires_grad" not in lower and "param_groups" not in lower:316            findings.append(f"{rel}: verify optimizer parameter coverage if custom freezing is expected.")317

587588    risks = [589        "# Suspicious Patterns",590        "",591    ]592    patterns = data["suspicious_patterns"]593    if patterns:594        risks.extend(f"- {item}" for item in patterns)595    else:596        risks.append("- No high-signal suspicious patterns were detected by the lightweight heuristic pass.")597    risks.append("")598    (output_dir / "RISKS.md").write_text("\n".join(risks), encoding="utf-8")599

116117    entry = {118        "ts": int(time.time()),119        "kind": kind,120        "skill": sanitize(skill) or "unknown",121        "summary": clean_summary,122        "detail": clean_detail,123        "fingerprint": sanitize(fingerprint) or "",124    }125    path = lessons_path()126    path.parent.mkdir(parents=True, exist_ok=True)127    with path.open("a", encoding="utf-8") as handle:128        handle.write(json.dumps(entry, ensure_ascii=False) + "\n")129    return path130

13901391    context["lesson_recorded"] = maybe_record_lesson(repo_path, context) if args.run_selected else None1392

122- Load `references/deep-learning-experiment-principles.md` when dataset, split, metric, checkpoint, training, or evaluation details matter.123- Consult `~/.rigorpilot/PERSONAL_RIGOR.md` if present, under `references/continuous-learning-policy.md` (advisory only; core wins).124- Failed and later-resolved runs are auto-recorded as lessons via `shared/scripts/lessons_store.py` (`RIGORPILOT_LESSONS=0` disables).125- Load `references/research-safety-principles.md` before protocol-sensitive

4041def lessons_home() -> Path:42    root = os.environ.get("RIGORPILOT_HOME")43    return Path(root).expanduser() if root else Path.home() / ".rigorpilot"444546def lessons_enabled() -> bool:47    return os.environ.get("RIGORPILOT_LESSONS", "1") != "0"48

6667def sanitize(text: str) -> Optional[str]:68    cleaned = " ".join(str(text or "").split())[:MAX_FIELD_CHARS]69    if not cleaned:70        return None71    if SECRET_RE.search(cleaned):72        return None73    return cleaned74

61    try:62        if status in {"partial", "blocked"}:63            path = store.record_lesson(64                kind="failure-fix",65                skill="ai-research-reproduction",66                summary=f"[{status}] {context.get('main_blocker', 'unrecorded blocker')}",67                detail=str(context.get("documented_command") or ""),68                fingerprint=fingerprint,69            )70            return str(path) if path else None

587588    risks = [589        "# Suspicious Patterns",590        "",591    ]592    patterns = data["suspicious_patterns"]593    if patterns:594        risks.extend(f"- {item}" for item in patterns)595    else:596        risks.append("- No high-signal suspicious patterns were detected by the lightweight heuristic pass.")597    risks.append("")598    (output_dir / "RISKS.md").write_text("\n".join(risks), encoding="utf-8")599

630    }631    (output_dir / "status.json").write_text(json.dumps(status, indent=2, ensure_ascii=False), encoding="utf-8")632633634def main() -> int:635    parser = argparse.ArgumentParser(description="Analyze a deep learning research repository conservatively.")636    parser.add_argument("--repo", required=True, help="Path to the target repository.")637    parser.add_argument("--output-dir", default="analysis_outputs", help="Directory for analysis outputs.")638    parser.add_argument("--analysis-context-json", default="", help="Optional analysis context JSON or YAML path.")639    parser.add_argument("--json", action="store_true", help="Emit JSON to stdout instead of writing files.")640    args = parser.parse_args()641642    repo = Path(args.repo).resolve()643    context = load_context(args.analysis_context_json)644    data = analyze_repo(repo, context)645    if args.json:646        print(json.dumps(data, indent=2, ensure_ascii=False))647        return 0648649    write_outputs(Path(args.output_dir).resolve(), data)650    print(json.dumps(data, indent=2, ensure_ascii=False))651    return 0

2829## Clear boundaries3031- This skill is read-mostly.32- It may run lightweight static inspection helpers.33- It does not patch repository code.34- It does not own final reproduction outputs.35- It should mark suspicious patterns as heuristics, not confirmed bugs.3637## Output expectations3839- `analysis_outputs/SUMMARY.md`40- `analysis_outputs/RISKS.md`41- `analysis_outputs/status.json`42

585    ]586    (output_dir / "SUMMARY.md").write_text("\n".join(summary), encoding="utf-8")587588    risks = [589        "# Suspicious Patterns",590        "",591    ]592    patterns = data["suspicious_patterns"]593    if patterns:594        risks.extend(f"- {item}" for item in patterns)595    else:596        risks.append("- No high-signal suspicious patterns were detected by the lightweight heuristic pass.")597    risks.append("")598    (output_dir / "RISKS.md").write_text("\n".join(risks), encoding="utf-8")599