cawdex 1.35.74 → 1.35.76

package/resources/hal/cawdex_agent/main.py

@@ -1,550 +1,550 @@
-"""HAL custom-agent adapter for Cawdex.
-
-HAL expects a module-level run(input, **kwargs) function. This adapter keeps
-Cawdex framework-agnostic by launching the installed CLI in headless
-benchmark mode, then returning the artifact shape expected by common HAL tasks.
-"""
-
-from __future__ import annotations
-
-import json
-import os
-import re
-import shlex
-import subprocess
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Any
-
-
-SECRET_REPLACEMENTS = [
-    (re.compile(r"sk-or-v1-[A-Za-z0-9_-]+"), "sk-or-v1-[REDACTED]"),
-    (re.compile(r"sk-[A-Za-z0-9_-]{16,}"), "sk-[REDACTED]"),
-    (re.compile(r"hf_[A-Za-z0-9]{16,}"), "hf_[REDACTED]"),
-    (re.compile(r"KGAT_[A-Za-z0-9]{16,}"), "KGAT_[REDACTED]"),
-    (re.compile(r"npm_[A-Za-z0-9]{16,}"), "npm_[REDACTED]"),
-]
-
-ORACLE_FIELD_RE = re.compile(
-    r"(^|_)(patch|test_patch|solution|answer|gold|fail_to_pass|pass_to_pass)($|_)",
-    re.IGNORECASE,
-)
-
-SAFE_FIELD_ORDER = [
-    "instance_id",
-    "task_id",
-    "repo",
-    "base_commit",
-    "version",
-    "created_at",
-    "problem_statement",
-    "hints_text",
-    "description",
-    "description_no_samples",
-    "samples",
-    "num_tests",
-    "num_samples",
-    "problem_link",
-    "problem_level",
-    "cp_id",
-    "problem_id",
-    "runtime_limit",
-    "memory_limit",
-    "runtime_limit_sentences",
-    "memory_limit_sentences",
-    "task_inst",
-    "dataset_path",
-    "dataset_folder_tree",
-    "dataset_preview",
-    "output_fname",
-    "domain_knowledge",
-]
-
-
-@dataclass
-class AgentRun:
-    returncode: int
-    stdout: str
-    stderr: str
-    trace_dir: Path
-
-
-def _redact(text: Any) -> str:
-    value = str(text or "")
-    for pattern, replacement in SECRET_REPLACEMENTS:
-        value = pattern.sub(replacement, value)
-    return value
-
-
-def _truncate(text: str, limit: int = 50000) -> str:
-    clean = _redact(text)
-    if len(clean) <= limit:
-        return clean
-    omitted = len(clean) - limit
-    return clean[:limit] + f"\n...[truncated {omitted} chars]"
-
-
-def _safe_task_id(task_id: Any) -> str:
-    raw = str(task_id or "task")
-    safe = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip("-")
-    return safe or "task"
-
-
-def _include_oracle_fields() -> bool:
-    return os.environ.get("CAWDEX_HAL_INCLUDE_ORACLE_FIELDS", "").lower() in {"1", "true", "yes", "on"}
-
-
-def _safe_task_view(task: dict[str, Any]) -> tuple[dict[str, Any], list[str]]:
-    if _include_oracle_fields():
-        return task, []
-
-    allowed: dict[str, Any] = {}
-    omitted: list[str] = []
-    ordered_keys = [key for key in SAFE_FIELD_ORDER if key in task]
-    ordered_keys.extend(sorted(key for key in task if key not in ordered_keys))
-    for key in ordered_keys:
-        if ORACLE_FIELD_RE.search(key):
-            omitted.append(key)
-            continue
-        allowed[key] = task[key]
-    return allowed, omitted
-
-
-def _is_patch_task(task: dict[str, Any]) -> bool:
-    return bool(
-        task.get("problem_statement")
-        and (task.get("repo") or task.get("base_commit") or task.get("instance_id"))
-    )
-
-
-def _is_science_agent_task(task: dict[str, Any]) -> bool:
-    return bool(
-        task.get("task_inst")
-        and (task.get("dataset_path") or task.get("output_fname") or task.get("dataset_folder_tree"))
-    )
-
-
-def _is_appworld_task(task: dict[str, Any]) -> bool:
-    keys = set(task.keys())
-    return bool(task.get("task_id") and keys.issubset({"task_id", "instance_id"}))
-
-
-def _is_usaco_task(task: dict[str, Any]) -> bool:
-    return bool(
-        task.get("description")
-        and (task.get("samples") or task.get("cp_id") or task.get("problem_id") or task.get("problem_link"))
-    )
-
-
-def _profile_for_task(task: dict[str, Any]) -> str:
-    task_text = json.dumps(task, ensure_ascii=False).lower()
-    if _is_appworld_task(task) or "appworld" in task_text or "app-world" in task_text:
-        return "appworld"
-    if (
-        "browsecomp" in task_text
-        or "browsecomp+" in task_text
-        or "browse-comp" in task_text
-        or "deep research" in task_text
-        or "web research" in task_text
-    ):
-        return "browsecomp"
-    if (
-        "tau2" in task_text
-        or "tau 2" in task_text
-        or "tau-bench" in task_text
-        or "tau_bench" in task_text
-        or "taubench" in task_text
-        or "customer support" in task_text
-    ):
-        return "tau2"
-    if (
-        "roadmapbench" in task_text
-        or "roadmap-bench" in task_text
-        or "long-horizon" in task_text
-        or "long horizon" in task_text
-        or "version upgrade" in task_text
-        or "multi-target" in task_text
-    ):
-        return "roadmapbench"
-    if (
-        "swe-cycle" in task_text
-        or "swecycle" in task_text
-        or "swe cycle" in task_text
-        or "swe-judge" in task_text
-        or "swejudge" in task_text
-        or "fullcycle" in task_text
-        or "codeimpl" in task_text
-        or "testgen" in task_text
-        or "run_script" in task_text
-        or "parsing_script" in task_text
-        or "selected_test_files_to_run" in task_text
-        or "environment_setup_commit" in task_text
-        or "before_repo_set_cmd" in task_text
-        or "bare repository" in task_text
-    ):
-        return "swe-cycle"
-    if (
-        "swe-ci" in task_text
-        or "sweci" in task_text
-        or "swe ci" in task_text
-        or "run_tests" in task_text
-        or "define_requirements" in task_text
-        or "modify_code" in task_text
-        or "test gap" in task_text
-        or "current_sha" in task_text
-        or "target_sha" in task_text
-        or ("codebase maintenance" in task_text and "continuous integration" in task_text)
-    ):
-        return "swe-ci"
-    if (
-        "swe-prbench" in task_text
-        or "swe prbench" in task_text
-        or "swe-pr" in task_text
-        or "prbench" in task_text
-        or "pull request review" in task_text
-        or "code review quality" in task_text
-        or "human_review_comments" in task_text
-        or "diff_patch" in task_text
-        or "type2_contextual" in task_text
-    ):
-        return "swe-prbench"
-    if (
-        "tml-bench" in task_text
-        or "tmlbench" in task_text
-        or "tabular ml" in task_text
-        or "kaggle-style" in task_text
-        or "kaggle style" in task_text
-        or "sample_submission" in task_text
-        or "private holdout" in task_text
-        or ("train.csv" in task_text and "test.csv" in task_text and "submission" in task_text)
-    ):
-        return "tml-bench"
-    if (
-        "pi-bench" in task_text
-        or "pibench" in task_text
-        or "proactive personal assistant" in task_text
-        or "proactive assistant" in task_text
-        or "hidden intent" in task_text
-        or "latent intent" in task_text
-        or "user profile" in task_text
-        or "message history" in task_text
-        or "current app" in task_text
-        or "proactivity score" in task_text
-        or "completion score" in task_text
-    ):
-        return "pi-bench"
-    if (
-        "saasbench" in task_text
-        or "saas-bench" in task_text
-        or "enterprise saas" in task_text
-        or "validation nodes" in task_text
-        or "tenant" in task_text
-        or "migration" in task_text
-    ):
-        return "saasbench"
-    if (
-        "swe-bench mobile" in task_text
-        or "swebench mobile" in task_text
-        or "xcode" in task_text
-        or "swift" in task_text
-        or "objective-c" in task_text
-        or "figma" in task_text
-        or "simulator" in task_text
-    ):
-        return "swe-bench-mobile"
-    if (
-        "swe-webdevbench" in task_text
-        or "swe-webdev-bench" in task_text
-        or "webdevbench" in task_text
-        or "webdev-bench" in task_text
-        or "vibe coding" in task_text
-        or "virtual software agency" in task_text
-        or "canary requirement" in task_text
-        or "frontend-backend" in task_text
-        or "production readiness" in task_text
-    ):
-        return "webdevbench"
-    if _is_patch_task(task):
-        return "swe-bench"
-    if (
-        "terminalworld" in task_text
-        or "terminal-world" in task_text
-        or "tw_" in task_text
-        or "asciinema" in task_text
-    ):
-        return "terminalworld"
-    if "terminal-bench" in task_text or "terminalbench" in task_text:
-        return "terminal-bench"
-    if (
-        "wildclaw" in task_text
-        or "openclaw" in task_text
-        or "browsecomp" in task_text
-        or "ossworld" in task_text
-        or "bfcl" in task_text
-        or "webwalkerqa" in task_text
-    ):
-        return "wildclaw"
-    if (
-        "arc-agi" in task_text
-        or "arc_agi" in task_text
-        or "arc prize" in task_text
-        or "arc-prize" in task_text
-        or "kaggle arc" in task_text
-    ):
-        return "arc-agi"
-    if (
-        "specbench" in task_text
-        or "spec-bench" in task_text
-        or "spec compliance" in task_text
-        or "visible tests" in task_text
-        or "held-out" in task_text
-        or "holdout" in task_text
-    ):
-        return "specbench"
-    if (
-        "reward hacking benchmark" in task_text
-        or "reward-hacking" in task_text
-        or "reward_hacking" in task_text
-        or "rhb" in task_text
-        or "evaluator tamper" in task_text
-    ):
-        return "reward-hacking"
-    return "generic"
-
-
-def _build_prompt(task_id: str, task: dict[str, Any]) -> str:
-    profile = _profile_for_task(task)
-    safe_task, omitted = _safe_task_view(task)
-    body = json.dumps(safe_task, ensure_ascii=False, indent=2, sort_keys=True)
-
-    lines = [
-        f"/benchmark {profile} HAL task {task_id}",
-        "",
-        "You are running inside the Holistic Agent Leaderboard harness.",
-        "Use Cawdex benchmark discipline: inspect local files, patch only what is needed, run targeted verification, and preserve trace evidence.",
-    ]
-    if profile == "swe-bench":
-        lines.extend([
-            "This is a SWE-bench-style patch task. Modify the checked-out repository; the HAL adapter will collect the git patch after the run.",
-            "Do not edit tests or harness files unless the task explicitly asks for that.",
-        ])
-    elif profile == "terminalworld":
-        lines.append("This is a TerminalWorld-style terminal workflow. Treat instruction.md/task text as the contract, avoid solve.sh/reference material, produce required persistent artifacts, and verify files/services in the environment.")
-    elif _is_science_agent_task(task):
-        lines.append("This is a ScienceAgentBench-style task. Produce a concise solution trajectory and any required output/program artifact in the final response.")
-    elif profile == "appworld" or _is_appworld_task(task):
-        lines.append("This is an AppWorld-style environment task. Interact with the environment as needed, then complete the task through the environment API.")
-    elif profile == "browsecomp":
-        lines.append("This is a BrowseComp+-style research task. Use source-grounded browsing/retrieval evidence, cross-check claims, and return the answer with auditable attribution.")
-    elif profile == "tau2":
-        lines.append("This is a tau2/Tau-Bench-style policy workflow. Follow the domain policy, use only available action schemas, and verify tool observations before completing.")
-    elif profile == "webdevbench":
-        lines.append("This is a SWE-WebDevBench-style full-stack app-agency task. Preserve canary business requirements, verify frontend-backend coupling, and collect production/security evidence when feasible.")
-    elif profile == "swe-cycle":
-        lines.append("This is a SWE-Cycle/SWE-Judge-style issue-resolution lifecycle task. Track environment setup, code implementation, verification-test generation when required, and post-edit static/dynamic judge evidence.")
-    elif profile == "swe-ci":
-        lines.append("This is a SWE-CI-style repository evolution task. Track current/target commits, test gaps, requirement derivation, and CI-loop validation across run_tests -> define_requirements -> modify_code.")
-    elif profile == "swe-prbench":
-        lines.append("This is a SWE-PRBench-style pull request review task. Inspect PR metadata and diff first, expand context only for concrete suspected issues, and return severity-rated review findings with file/line evidence instead of patching unless explicitly requested.")
-    elif profile == "tml-bench":
-        lines.append("This is a TML-Bench/Kaggle-style tabular ML task. Build the data contract first, avoid hidden-label leakage, train an honest baseline, and produce a sample_submission-compatible artifact with validation evidence.")
-    elif profile == "pi-bench":
-        lines.append("This is a Pi-Bench-style proactive personal assistant task. Build the user/workspace/app context contract, infer hidden intents carefully, ask one focused clarification when needed, and verify observable state after proactive actions.")
-    elif _is_usaco_task(task):
-        lines.append("This is a USACO-style programming task. Produce the final code solution in the final response.")
-    else:
-        lines.append("Return the final task response clearly; the HAL adapter will store it in the task response field.")
-
-    if omitted:
-        lines.append("Oracle-like task fields omitted from the prompt by default: " + ", ".join(sorted(omitted)) + ".")
-
-    lines.extend(["", "## HAL task data", _truncate(body)])
-    return "\n".join(lines)
-
-
-def _base_command() -> list[str]:
-    command = os.environ.get("CAWDEX_HAL_COMMAND") or os.environ.get("CAWDEX_HAL_COMMAND", "cawdex")
-    parts = shlex.split(command, posix=os.name != "nt")
-    return parts or ["cawdex"]
-
-
-def _append_flag(args: list[str], flag: str, value: Any) -> None:
-    if value is None:
-        return
-    text = str(value).strip()
-    if not text:
-        return
-    args.extend([flag, text])
-
-
-def _run_cawdex(task_id: str, prompt: str, kwargs: dict[str, Any]) -> AgentRun:
-    trace_root = Path(os.environ.get("CAWDEX_HAL_TRACE_DIR", ".cawdex/hal-trace"))
-    trace_dir = trace_root / _safe_task_id(task_id)
-    trace_dir.mkdir(parents=True, exist_ok=True)
-
-    env = os.environ.copy()
-    env.setdefault("CAWDEX_ENV_CONFIG", "1")
-    env.setdefault("CAWDEX_THEME", "minimal")
-    env.setdefault("CAWDEX_SHOW_THINKING", "0")
-    env.setdefault("CAWDEX_MEMORY", "0")
-    env.setdefault("CAWDEX_BASH_TIMEOUT_MS", "300000")
-
-    args = _base_command()
-    args.extend([
-        "--prompt",
-        prompt,
-        "--perm",
-        "yolo",
-        "--benchmark-trace-dir",
-        str(trace_dir),
-    ])
-    _append_flag(args, "--model", kwargs.get("model_name") or kwargs.get("model"))
-    _append_flag(args, "--provider", kwargs.get("provider"))
-    _append_flag(args, "--max-turns", kwargs.get("max_turns"))
-    _append_flag(args, "--max-tokens", kwargs.get("max_tokens"))
-    _append_flag(args, "--temperature", kwargs.get("temperature"))
-    _append_flag(args, "--output-format", kwargs.get("output_format"))
-
-    timeout = int(os.environ.get("CAWDEX_HAL_TIMEOUT_SEC", "1800"))
-    try:
-        completed = subprocess.run(
-            args,
-            cwd=os.getcwd(),
-            env=env,
-            text=True,
-            capture_output=True,
-            timeout=timeout,
-            check=False,
-        )
-        stdout = _redact(completed.stdout)
-        stderr = _redact(completed.stderr)
-        returncode = completed.returncode
-    except subprocess.TimeoutExpired as exc:
-        stdout = _redact(exc.stdout)
-        stderr = _redact(exc.stderr) + f"\nCawdex timed out after {timeout}s"
-        returncode = 124
-
-    (trace_dir / "hal-stdout.txt").write_text(stdout, encoding="utf-8")
-    (trace_dir / "hal-stderr.txt").write_text(stderr, encoding="utf-8")
-    return AgentRun(returncode=returncode, stdout=stdout, stderr=stderr, trace_dir=trace_dir)
-
-
-def _run_git(args: list[str], cwd: Path | None = None) -> str:
-    try:
-        completed = subprocess.run(
-            ["git", *args],
-            cwd=str(cwd) if cwd else None,
-            text=True,
-            capture_output=True,
-            check=False,
-            timeout=60,
-        )
-    except Exception:
-        return ""
-    if completed.returncode not in {0, 1}:
-        return ""
-    return _redact(completed.stdout)
-
-
-def _latest_trace_patch(trace_dir: Path) -> str:
-    patches = sorted(
-        trace_dir.rglob("worktree.patch"),
-        key=lambda path: path.stat().st_mtime,
-        reverse=True,
-    )
-    for patch in patches:
-        try:
-            text = _redact(patch.read_text(encoding="utf-8", errors="replace"))
-            if text.strip():
-                return text
-        except OSError:
-            continue
-    return ""
-
-
-def _collect_git_patch(trace_dir: Path) -> str:
-    trace_patch = _latest_trace_patch(trace_dir)
-    if trace_patch:
-        return trace_patch
-
-    parts = [
-        _run_git(["diff", "--binary", "--no-ext-diff"]),
-        _run_git(["diff", "--cached", "--binary", "--no-ext-diff"]),
-    ]
-    if os.name != "nt":
-        raw_untracked = _run_git(["ls-files", "--others", "--exclude-standard", "-z"])
-        for filename in raw_untracked.split("\0"):
-            if filename:
-                parts.append(_run_git(["diff", "--no-index", "--binary", "--no-ext-diff", "--", "/dev/null", filename]))
-    return "".join(part for part in parts if part)
-
-
-def _latest_summary(trace_dir: Path) -> dict[str, Any]:
-    summaries = sorted(
-        trace_dir.rglob("summary.json"),
-        key=lambda path: path.stat().st_mtime,
-        reverse=True,
-    )
-    for summary in summaries:
-        try:
-            return json.loads(summary.read_text(encoding="utf-8"))
-        except Exception:
-            continue
-    return {}
-
-
-def _response_text(run_result: AgentRun) -> str:
-    summary = _latest_summary(run_result.trace_dir)
-    final = summary.get("finalAssistant")
-    if isinstance(final, str) and final.strip():
-        return _truncate(final, 20000)
-    combined = "\n".join(part for part in [run_result.stdout, run_result.stderr] if part)
-    return _truncate(combined, 20000)
-
-
-def _submission_for_task(task: dict[str, Any], run_result: AgentRun) -> Any:
-    response = _response_text(run_result)
-    if _is_science_agent_task(task):
-        return response
-    if _is_appworld_task(task):
-        return "Completed" if run_result.returncode == 0 else response
-
-    updated = dict(task)
-    updated["response"] = response
-    return updated
-
-
-def run(input: dict[str, dict[str, Any]], **kwargs: Any) -> dict[str, Any]:
-    """Run Cawdex for HAL.
-
-    Patch-style tasks return {task_id: patch}. ScienceAgentBench-style tasks
-    return a trajectory string. AppWorld-style tasks return "Completed" after
-    a successful run. Other text/code tasks return the original task dict with
-    a response field, matching HAL's USACO-style pattern.
-    """
-    if not isinstance(input, dict):
-        raise TypeError("Cawdex HAL adapter expects input to be a dictionary")
-
-    patch_task_ids = [
-        str(task_id)
-        for task_id, task in input.items()
-        if isinstance(task, dict) and _is_patch_task(task)
-    ]
-    if len(patch_task_ids) > 1:
-        raise ValueError("Cawdex HAL adapter expects one patch-style task per checked-out worktree")
-
-    output: dict[str, Any] = {}
-    for task_id, task in input.items():
-        if not isinstance(task, dict):
-            output[str(task_id)] = task
-            continue
-
-        prompt = _build_prompt(str(task_id), task)
-        run_result = _run_cawdex(str(task_id), prompt, kwargs)
-
-        if _is_patch_task(task):
-            output[str(task_id)] = _collect_git_patch(run_result.trace_dir)
-        else:
-            output[str(task_id)] = _submission_for_task(task, run_result)
-
-    return output
+"""HAL custom-agent adapter for Cawdex.
+
+HAL expects a module-level run(input, **kwargs) function. This adapter keeps
+Cawdex framework-agnostic by launching the installed CLI in headless
+benchmark mode, then returning the artifact shape expected by common HAL tasks.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import shlex
+import subprocess
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any
+
+
+SECRET_REPLACEMENTS = [
+    (re.compile(r"sk-or-v1-[A-Za-z0-9_-]+"), "sk-or-v1-[REDACTED]"),
+    (re.compile(r"sk-[A-Za-z0-9_-]{16,}"), "sk-[REDACTED]"),
+    (re.compile(r"hf_[A-Za-z0-9]{16,}"), "hf_[REDACTED]"),
+    (re.compile(r"KGAT_[A-Za-z0-9]{16,}"), "KGAT_[REDACTED]"),
+    (re.compile(r"npm_[A-Za-z0-9]{16,}"), "npm_[REDACTED]"),
+]
+
+ORACLE_FIELD_RE = re.compile(
+    r"(^|_)(patch|test_patch|solution|answer|gold|fail_to_pass|pass_to_pass)($|_)",
+    re.IGNORECASE,
+)
+
+SAFE_FIELD_ORDER = [
+    "instance_id",
+    "task_id",
+    "repo",
+    "base_commit",
+    "version",
+    "created_at",
+    "problem_statement",
+    "hints_text",
+    "description",
+    "description_no_samples",
+    "samples",
+    "num_tests",
+    "num_samples",
+    "problem_link",
+    "problem_level",
+    "cp_id",
+    "problem_id",
+    "runtime_limit",
+    "memory_limit",
+    "runtime_limit_sentences",
+    "memory_limit_sentences",
+    "task_inst",
+    "dataset_path",
+    "dataset_folder_tree",
+    "dataset_preview",
+    "output_fname",
+    "domain_knowledge",
+]
+
+
+@dataclass
+class AgentRun:
+    returncode: int
+    stdout: str
+    stderr: str
+    trace_dir: Path
+
+
+def _redact(text: Any) -> str:
+    value = str(text or "")
+    for pattern, replacement in SECRET_REPLACEMENTS:
+        value = pattern.sub(replacement, value)
+    return value
+
+
+def _truncate(text: str, limit: int = 50000) -> str:
+    clean = _redact(text)
+    if len(clean) <= limit:
+        return clean
+    omitted = len(clean) - limit
+    return clean[:limit] + f"\n...[truncated {omitted} chars]"
+
+
+def _safe_task_id(task_id: Any) -> str:
+    raw = str(task_id or "task")
+    safe = re.sub(r"[^A-Za-z0-9_.-]+", "-", raw).strip("-")
+    return safe or "task"
+
+
+def _include_oracle_fields() -> bool:
+    return os.environ.get("CAWDEX_HAL_INCLUDE_ORACLE_FIELDS", "").lower() in {"1", "true", "yes", "on"}
+
+
+def _safe_task_view(task: dict[str, Any]) -> tuple[dict[str, Any], list[str]]:
+    if _include_oracle_fields():
+        return task, []
+
+    allowed: dict[str, Any] = {}
+    omitted: list[str] = []
+    ordered_keys = [key for key in SAFE_FIELD_ORDER if key in task]
+    ordered_keys.extend(sorted(key for key in task if key not in ordered_keys))
+    for key in ordered_keys:
+        if ORACLE_FIELD_RE.search(key):
+            omitted.append(key)
+            continue
+        allowed[key] = task[key]
+    return allowed, omitted
+
+
+def _is_patch_task(task: dict[str, Any]) -> bool:
+    return bool(
+        task.get("problem_statement")
+        and (task.get("repo") or task.get("base_commit") or task.get("instance_id"))
+    )
+
+
+def _is_science_agent_task(task: dict[str, Any]) -> bool:
+    return bool(
+        task.get("task_inst")
+        and (task.get("dataset_path") or task.get("output_fname") or task.get("dataset_folder_tree"))
+    )
+
+
+def _is_appworld_task(task: dict[str, Any]) -> bool:
+    keys = set(task.keys())
+    return bool(task.get("task_id") and keys.issubset({"task_id", "instance_id"}))
+
+
+def _is_usaco_task(task: dict[str, Any]) -> bool:
+    return bool(
+        task.get("description")
+        and (task.get("samples") or task.get("cp_id") or task.get("problem_id") or task.get("problem_link"))
+    )
+
+
+def _profile_for_task(task: dict[str, Any]) -> str:
+    task_text = json.dumps(task, ensure_ascii=False).lower()
+    if _is_appworld_task(task) or "appworld" in task_text or "app-world" in task_text:
+        return "appworld"
+    if (
+        "browsecomp" in task_text
+        or "browsecomp+" in task_text
+        or "browse-comp" in task_text
+        or "deep research" in task_text
+        or "web research" in task_text
+    ):
+        return "browsecomp"
+    if (
+        "tau2" in task_text
+        or "tau 2" in task_text
+        or "tau-bench" in task_text
+        or "tau_bench" in task_text
+        or "taubench" in task_text
+        or "customer support" in task_text
+    ):
+        return "tau2"
+    if (
+        "roadmapbench" in task_text
+        or "roadmap-bench" in task_text
+        or "long-horizon" in task_text
+        or "long horizon" in task_text
+        or "version upgrade" in task_text
+        or "multi-target" in task_text
+    ):
+        return "roadmapbench"
+    if (
+        "swe-cycle" in task_text
+        or "swecycle" in task_text
+        or "swe cycle" in task_text
+        or "swe-judge" in task_text
+        or "swejudge" in task_text
+        or "fullcycle" in task_text
+        or "codeimpl" in task_text
+        or "testgen" in task_text
+        or "run_script" in task_text
+        or "parsing_script" in task_text
+        or "selected_test_files_to_run" in task_text
+        or "environment_setup_commit" in task_text
+        or "before_repo_set_cmd" in task_text
+        or "bare repository" in task_text
+    ):
+        return "swe-cycle"
+    if (
+        "swe-ci" in task_text
+        or "sweci" in task_text
+        or "swe ci" in task_text
+        or "run_tests" in task_text
+        or "define_requirements" in task_text
+        or "modify_code" in task_text
+        or "test gap" in task_text
+        or "current_sha" in task_text
+        or "target_sha" in task_text
+        or ("codebase maintenance" in task_text and "continuous integration" in task_text)
+    ):
+        return "swe-ci"
+    if (
+        "swe-prbench" in task_text
+        or "swe prbench" in task_text
+        or "swe-pr" in task_text
+        or "prbench" in task_text
+        or "pull request review" in task_text
+        or "code review quality" in task_text
+        or "human_review_comments" in task_text
+        or "diff_patch" in task_text
+        or "type2_contextual" in task_text
+    ):
+        return "swe-prbench"
+    if (
+        "tml-bench" in task_text
+        or "tmlbench" in task_text
+        or "tabular ml" in task_text
+        or "kaggle-style" in task_text
+        or "kaggle style" in task_text
+        or "sample_submission" in task_text
+        or "private holdout" in task_text
+        or ("train.csv" in task_text and "test.csv" in task_text and "submission" in task_text)
+    ):
+        return "tml-bench"
+    if (
+        "pi-bench" in task_text
+        or "pibench" in task_text
+        or "proactive personal assistant" in task_text
+        or "proactive assistant" in task_text
+        or "hidden intent" in task_text
+        or "latent intent" in task_text
+        or "user profile" in task_text
+        or "message history" in task_text
+        or "current app" in task_text
+        or "proactivity score" in task_text
+        or "completion score" in task_text
+    ):
+        return "pi-bench"
+    if (
+        "saasbench" in task_text
+        or "saas-bench" in task_text
+        or "enterprise saas" in task_text
+        or "validation nodes" in task_text
+        or "tenant" in task_text
+        or "migration" in task_text
+    ):
+        return "saasbench"
+    if (
+        "swe-bench mobile" in task_text
+        or "swebench mobile" in task_text
+        or "xcode" in task_text
+        or "swift" in task_text
+        or "objective-c" in task_text
+        or "figma" in task_text
+        or "simulator" in task_text
+    ):
+        return "swe-bench-mobile"
+    if (
+        "swe-webdevbench" in task_text
+        or "swe-webdev-bench" in task_text
+        or "webdevbench" in task_text
+        or "webdev-bench" in task_text
+        or "vibe coding" in task_text
+        or "virtual software agency" in task_text
+        or "canary requirement" in task_text
+        or "frontend-backend" in task_text
+        or "production readiness" in task_text
+    ):
+        return "webdevbench"
+    if _is_patch_task(task):
+        return "swe-bench"
+    if (
+        "terminalworld" in task_text
+        or "terminal-world" in task_text
+        or "tw_" in task_text
+        or "asciinema" in task_text
+    ):
+        return "terminalworld"
+    if "terminal-bench" in task_text or "terminalbench" in task_text:
+        return "terminal-bench"
+    if (
+        "wildclaw" in task_text
+        or "openclaw" in task_text
+        or "browsecomp" in task_text
+        or "ossworld" in task_text
+        or "bfcl" in task_text
+        or "webwalkerqa" in task_text
+    ):
+        return "wildclaw"
+    if (
+        "arc-agi" in task_text
+        or "arc_agi" in task_text
+        or "arc prize" in task_text
+        or "arc-prize" in task_text
+        or "kaggle arc" in task_text
+    ):
+        return "arc-agi"
+    if (
+        "specbench" in task_text
+        or "spec-bench" in task_text
+        or "spec compliance" in task_text
+        or "visible tests" in task_text
+        or "held-out" in task_text
+        or "holdout" in task_text
+    ):
+        return "specbench"
+    if (
+        "reward hacking benchmark" in task_text
+        or "reward-hacking" in task_text
+        or "reward_hacking" in task_text
+        or "rhb" in task_text
+        or "evaluator tamper" in task_text
+    ):
+        return "reward-hacking"
+    return "generic"
+
+
+def _build_prompt(task_id: str, task: dict[str, Any]) -> str:
+    profile = _profile_for_task(task)
+    safe_task, omitted = _safe_task_view(task)
+    body = json.dumps(safe_task, ensure_ascii=False, indent=2, sort_keys=True)
+
+    lines = [
+        f"/benchmark {profile} HAL task {task_id}",
+        "",
+        "You are running inside the Holistic Agent Leaderboard harness.",
+        "Use Cawdex benchmark discipline: inspect local files, patch only what is needed, run targeted verification, and preserve trace evidence.",
+    ]
+    if profile == "swe-bench":
+        lines.extend([
+            "This is a SWE-bench-style patch task. Modify the checked-out repository; the HAL adapter will collect the git patch after the run.",
+            "Do not edit tests or harness files unless the task explicitly asks for that.",
+        ])
+    elif profile == "terminalworld":
+        lines.append("This is a TerminalWorld-style terminal workflow. Treat instruction.md/task text as the contract, avoid solve.sh/reference material, produce required persistent artifacts, and verify files/services in the environment.")
+    elif _is_science_agent_task(task):
+        lines.append("This is a ScienceAgentBench-style task. Produce a concise solution trajectory and any required output/program artifact in the final response.")
+    elif profile == "appworld" or _is_appworld_task(task):
+        lines.append("This is an AppWorld-style environment task. Interact with the environment as needed, then complete the task through the environment API.")
+    elif profile == "browsecomp":
+        lines.append("This is a BrowseComp+-style research task. Use source-grounded browsing/retrieval evidence, cross-check claims, and return the answer with auditable attribution.")
+    elif profile == "tau2":
+        lines.append("This is a tau2/Tau-Bench-style policy workflow. Follow the domain policy, use only available action schemas, and verify tool observations before completing.")
+    elif profile == "webdevbench":
+        lines.append("This is a SWE-WebDevBench-style full-stack app-agency task. Preserve canary business requirements, verify frontend-backend coupling, and collect production/security evidence when feasible.")
+    elif profile == "swe-cycle":
+        lines.append("This is a SWE-Cycle/SWE-Judge-style issue-resolution lifecycle task. Track environment setup, code implementation, verification-test generation when required, and post-edit static/dynamic judge evidence.")
+    elif profile == "swe-ci":
+        lines.append("This is a SWE-CI-style repository evolution task. Track current/target commits, test gaps, requirement derivation, and CI-loop validation across run_tests -> define_requirements -> modify_code.")
+    elif profile == "swe-prbench":
+        lines.append("This is a SWE-PRBench-style pull request review task. Inspect PR metadata and diff first, expand context only for concrete suspected issues, and return severity-rated review findings with file/line evidence instead of patching unless explicitly requested.")
+    elif profile == "tml-bench":
+        lines.append("This is a TML-Bench/Kaggle-style tabular ML task. Build the data contract first, avoid hidden-label leakage, train an honest baseline, and produce a sample_submission-compatible artifact with validation evidence.")
+    elif profile == "pi-bench":
+        lines.append("This is a Pi-Bench-style proactive personal assistant task. Build the user/workspace/app context contract, infer hidden intents carefully, ask one focused clarification when needed, and verify observable state after proactive actions.")
+    elif _is_usaco_task(task):
+        lines.append("This is a USACO-style programming task. Produce the final code solution in the final response.")
+    else:
+        lines.append("Return the final task response clearly; the HAL adapter will store it in the task response field.")
+
+    if omitted:
+        lines.append("Oracle-like task fields omitted from the prompt by default: " + ", ".join(sorted(omitted)) + ".")
+
+    lines.extend(["", "## HAL task data", _truncate(body)])
+    return "\n".join(lines)
+
+
+def _base_command() -> list[str]:
+    command = os.environ.get("CAWDEX_HAL_COMMAND") or os.environ.get("CAWDEX_HAL_COMMAND", "cawdex")
+    parts = shlex.split(command, posix=os.name != "nt")
+    return parts or ["cawdex"]
+
+
+def _append_flag(args: list[str], flag: str, value: Any) -> None:
+    if value is None:
+        return
+    text = str(value).strip()
+    if not text:
+        return
+    args.extend([flag, text])
+
+
+def _run_cawdex(task_id: str, prompt: str, kwargs: dict[str, Any]) -> AgentRun:
+    trace_root = Path(os.environ.get("CAWDEX_HAL_TRACE_DIR", ".cawdex/hal-trace"))
+    trace_dir = trace_root / _safe_task_id(task_id)
+    trace_dir.mkdir(parents=True, exist_ok=True)
+
+    env = os.environ.copy()
+    env.setdefault("CAWDEX_ENV_CONFIG", "1")
+    env.setdefault("CAWDEX_THEME", "minimal")
+    env.setdefault("CAWDEX_SHOW_THINKING", "0")
+    env.setdefault("CAWDEX_MEMORY", "0")
+    env.setdefault("CAWDEX_BASH_TIMEOUT_MS", "300000")
+
+    args = _base_command()
+    args.extend([
+        "--prompt",
+        prompt,
+        "--perm",
+        "yolo",
+        "--benchmark-trace-dir",
+        str(trace_dir),
+    ])
+    _append_flag(args, "--model", kwargs.get("model_name") or kwargs.get("model"))
+    _append_flag(args, "--provider", kwargs.get("provider"))
+    _append_flag(args, "--max-turns", kwargs.get("max_turns"))
+    _append_flag(args, "--max-tokens", kwargs.get("max_tokens"))
+    _append_flag(args, "--temperature", kwargs.get("temperature"))
+    _append_flag(args, "--output-format", kwargs.get("output_format"))
+
+    timeout = int(os.environ.get("CAWDEX_HAL_TIMEOUT_SEC", "1800"))
+    try:
+        completed = subprocess.run(
+            args,
+            cwd=os.getcwd(),
+            env=env,
+            text=True,
+            capture_output=True,
+            timeout=timeout,
+            check=False,
+        )
+        stdout = _redact(completed.stdout)
+        stderr = _redact(completed.stderr)
+        returncode = completed.returncode
+    except subprocess.TimeoutExpired as exc:
+        stdout = _redact(exc.stdout)
+        stderr = _redact(exc.stderr) + f"\nCawdex timed out after {timeout}s"
+        returncode = 124
+
+    (trace_dir / "hal-stdout.txt").write_text(stdout, encoding="utf-8")
+    (trace_dir / "hal-stderr.txt").write_text(stderr, encoding="utf-8")
+    return AgentRun(returncode=returncode, stdout=stdout, stderr=stderr, trace_dir=trace_dir)
+
+
+def _run_git(args: list[str], cwd: Path | None = None) -> str:
+    try:
+        completed = subprocess.run(
+            ["git", *args],
+            cwd=str(cwd) if cwd else None,
+            text=True,
+            capture_output=True,
+            check=False,
+            timeout=60,
+        )
+    except Exception:
+        return ""
+    if completed.returncode not in {0, 1}:
+        return ""
+    return _redact(completed.stdout)
+
+
+def _latest_trace_patch(trace_dir: Path) -> str:
+    patches = sorted(
+        trace_dir.rglob("worktree.patch"),
+        key=lambda path: path.stat().st_mtime,
+        reverse=True,
+    )
+    for patch in patches:
+        try:
+            text = _redact(patch.read_text(encoding="utf-8", errors="replace"))
+            if text.strip():
+                return text
+        except OSError:
+            continue
+    return ""
+
+
+def _collect_git_patch(trace_dir: Path) -> str:
+    trace_patch = _latest_trace_patch(trace_dir)
+    if trace_patch:
+        return trace_patch
+
+    parts = [
+        _run_git(["diff", "--binary", "--no-ext-diff"]),
+        _run_git(["diff", "--cached", "--binary", "--no-ext-diff"]),
+    ]
+    if os.name != "nt":
+        raw_untracked = _run_git(["ls-files", "--others", "--exclude-standard", "-z"])
+        for filename in raw_untracked.split("\0"):
+            if filename:
+                parts.append(_run_git(["diff", "--no-index", "--binary", "--no-ext-diff", "--", "/dev/null", filename]))
+    return "".join(part for part in parts if part)
+
+
+def _latest_summary(trace_dir: Path) -> dict[str, Any]:
+    summaries = sorted(
+        trace_dir.rglob("summary.json"),
+        key=lambda path: path.stat().st_mtime,
+        reverse=True,
+    )
+    for summary in summaries:
+        try:
+            return json.loads(summary.read_text(encoding="utf-8"))
+        except Exception:
+            continue
+    return {}
+
+
+def _response_text(run_result: AgentRun) -> str:
+    summary = _latest_summary(run_result.trace_dir)
+    final = summary.get("finalAssistant")
+    if isinstance(final, str) and final.strip():
+        return _truncate(final, 20000)
+    combined = "\n".join(part for part in [run_result.stdout, run_result.stderr] if part)
+    return _truncate(combined, 20000)
+
+
+def _submission_for_task(task: dict[str, Any], run_result: AgentRun) -> Any:
+    response = _response_text(run_result)
+    if _is_science_agent_task(task):
+        return response
+    if _is_appworld_task(task):
+        return "Completed" if run_result.returncode == 0 else response
+
+    updated = dict(task)
+    updated["response"] = response
+    return updated
+
+
+def run(input: dict[str, dict[str, Any]], **kwargs: Any) -> dict[str, Any]:
+    """Run Cawdex for HAL.
+
+    Patch-style tasks return {task_id: patch}. ScienceAgentBench-style tasks
+    return a trajectory string. AppWorld-style tasks return "Completed" after
+    a successful run. Other text/code tasks return the original task dict with
+    a response field, matching HAL's USACO-style pattern.
+    """
+    if not isinstance(input, dict):
+        raise TypeError("Cawdex HAL adapter expects input to be a dictionary")
+
+    patch_task_ids = [
+        str(task_id)
+        for task_id, task in input.items()
+        if isinstance(task, dict) and _is_patch_task(task)
+    ]
+    if len(patch_task_ids) > 1:
+        raise ValueError("Cawdex HAL adapter expects one patch-style task per checked-out worktree")
+
+    output: dict[str, Any] = {}
+    for task_id, task in input.items():
+        if not isinstance(task, dict):
+            output[str(task_id)] = task
+            continue
+
+        prompt = _build_prompt(str(task_id), task)
+        run_result = _run_cawdex(str(task_id), prompt, kwargs)
+
+        if _is_patch_task(task):
+            output[str(task_id)] = _collect_git_patch(run_result.trace_dir)
+        else:
+            output[str(task_id)] = _submission_for_task(task, run_result)
+
+    return output