scripts/preflight_study3.py

"""Outcome-blind runtime and executor preflight for frozen E09."""

from __future__ import annotations

from dataclasses import asdict
from hashlib import sha256
import importlib.metadata
import json
import platform
from pathlib import Path
import shutil
import subprocess
import sys
import tempfile
import time
from typing import Any

from agent_harness.lm_studio import LMStudioClient
from agent_harness.lm_studio_management import LMStudioResidencyManager, LMStudioServer
from agent_harness.pilot import research_code_revision
from agent_harness.protocol_experiment import ProtocolWorkspace
from agent_harness.specs import (
    load_edit_interfaces,
    load_experiments,
    load_models,
    load_repositories,
    load_task_split,
    load_tasks,
    validate_configuration_tree,
)
from agent_harness.study2_experiment import tokenizer_for


PACKAGES = ("numpy", "pytest", "scipy", "statsmodels", "tokenizers")


def _command(arguments: list[str], cwd: Path | None = None) -> dict[str, Any]:
    result = subprocess.run(
        arguments,
        cwd=cwd,
        text=True,
        capture_output=True,
        check=False,
        timeout=180,
    )
    return {
        "command": arguments,
        "returncode": result.returncode,
        "stdout": result.stdout,
        "stderr": result.stderr,
    }


def _tool_probe(client: LMStudioClient, model_key: str) -> dict[str, Any]:
    response = client.chat_completions(
        model_key,
        [
            {
                "role": "system",
                "content": "This is an outcome-blind runtime preflight. Call the required tool once.",
            },
            {
                "role": "user",
                "content": "Call preflight_echo with marker MODEL_TOOL_OK. Do not answer in prose.",
            },
        ],
        tools=[
            {
                "type": "function",
                "function": {
                    "name": "preflight_echo",
                    "description": "Return the exact requested runtime marker.",
                    "parameters": {
                        "type": "object",
                        "properties": {"marker": {"type": "string"}},
                        "required": ["marker"],
                        "additionalProperties": False,
                    },
                },
            }
        ],
        max_tokens=2_048,
        seed=0,
    )
    try:
        message = response["choices"][0]["message"]
        function = message["tool_calls"][0]["function"]
        raw = function["arguments"]
        arguments = raw if isinstance(raw, dict) else json.loads(raw)
    except (KeyError, IndexError, TypeError, json.JSONDecodeError) as exc:
        raise RuntimeError("model did not return a valid preflight tool call") from exc
    if function.get("name") != "preflight_echo" or arguments != {
        "marker": "MODEL_TOOL_OK"
    }:
        raise RuntimeError(f"unexpected tool preflight payload: {function}")
    serialized = json.dumps(response, sort_keys=True, separators=(",", ":"))
    return {
        "tool_name": function["name"],
        "arguments": arguments,
        "finish_reason": response["choices"][0].get("finish_reason"),
        "usage": response.get("usage", {}),
        "response_sha256": sha256(serialized.encode()).hexdigest(),
    }


def _executor_conformance(root: Path) -> dict[str, Any]:
    task = load_tasks(root)["TASK_CR_001"]
    interfaces = load_edit_interfaces(root)
    with tempfile.TemporaryDirectory(prefix="agent-harness-e09-preflight-") as temporary:
        tree = Path(temporary)
        target = tree / "example.go"
        target.write_text("package example\n\nconst value = 1\n", encoding="utf-8")
        for arguments in (["git", "init", "-q"], ["git", "add", "example.go"]):
            result = _command(arguments, tree)
            if result["returncode"]:
                raise RuntimeError(f"executor preflight git setup failed: {result}")
        records: dict[str, Any] = {}

        patch_workspace = ProtocolWorkspace(tree, ("example.go",), task, 2)
        patch = """--- a/example.go
+++ b/example.go
@@ -1,3 +1,3 @@
 package example

-const value = 1
+const value = 2
"""
        patch_result = patch_workspace.apply_patch(patch)
        if not patch_result["accepted"]:
            raise RuntimeError(f"P001 executor conformance failed: {patch_result}")
        records["P001"] = {
            "interface": asdict(interfaces["P001"]),
            "accepted": True,
            "final_patch_sha256": sha256(
                patch_workspace.final_patch().encode()
            ).hexdigest(),
        }

        target.write_text("package example\n\nconst value = 1\n", encoding="utf-8")
        replace_workspace = ProtocolWorkspace(tree, ("example.go",), task, 2)
        replace_result = replace_workspace.replace_text(
            "example.go", "value = 1", "value = 2"
        )
        records["P002"] = {
            "interface": asdict(interfaces["P002"]),
            "accepted": replace_result["accepted"],
            "final_patch_sha256": sha256(
                replace_workspace.final_patch().encode()
            ).hexdigest(),
        }

        target.write_text("package example\n\nconst value = 1\n", encoding="utf-8")
        write_workspace = ProtocolWorkspace(tree, ("example.go",), task, 2)
        write_result = write_workspace.write_file(
            "example.go", "package example\n\nconst value = 2\n"
        )
        records["P003"] = {
            "interface": asdict(interfaces["P003"]),
            "accepted": write_result["accepted"],
            "final_patch_sha256": sha256(
                write_workspace.final_patch().encode()
            ).hexdigest(),
        }
    if not all(record["accepted"] for record in records.values()):
        raise RuntimeError(f"an edit executor failed deterministic conformance: {records}")
    return records


def run(root: Path) -> dict[str, Any]:
    revision = research_code_revision(root)
    errors, warnings = validate_configuration_tree(root)
    if errors or warnings:
        raise RuntimeError(f"configuration failed: errors={errors}, warnings={warnings}")
    if (root / "results" / "raw" / "E09").exists():
        raise RuntimeError("E09 raw directory exists before outcome-blind preflight")
    audit = json.loads((root / "docs" / "STUDY3_DESIGN_AUDIT.json").read_text())
    if audit.get("planned_cells") != 540 or not audit.get("outcome_blind"):
        raise RuntimeError("Study 3 design audit is not a valid 540-cell freeze")
    experiment = load_experiments(root)["E09"]
    split = load_task_split(root / "tasks" / "splits" / "study3_protocol.txt")
    if experiment.cells_per_task() * len(split) != 540:
        raise RuntimeError("E09 execution plan is not exactly 540 cells")

    disk = shutil.disk_usage(root)
    if disk.free < 20 * 1024**3:
        raise RuntimeError(f"less than 20 GiB free before Study 3: {disk.free} bytes")
    models = load_models(root)
    repositories = load_repositories(root)
    server = LMStudioServer(port=1234)
    server_state = server.ensure_running()
    residency = LMStudioResidencyManager(
        models["M002"].base_url,
        models["M002"].api_token_env,
        timeout_seconds=1_800,
    )
    report: dict[str, Any] = {
        "schema_version": 1,
        "study": "Study 3 / E09",
        "outcome_blind": True,
        "research_code_revision": revision,
        "design_sha256": audit["design_sha256"],
        "planned_cells": 540,
        "started_unix": time.time(),
        "platform": platform.platform(),
        "python": sys.version,
        "torch_used": False,
        "disk": {"total": disk.total, "used": disk.used, "free": disk.free},
        "lms_version": _command([str(server.cli_path), "--version"]),
        "server_start": server_state,
        "dependencies": {
            package: importlib.metadata.version(package) for package in PACKAGES
        },
        "executor_conformance": _executor_conformance(root),
        "repositories": {},
        "models": {},
    }
    try:
        residency.unload_all()
        for repository_id, repository in sorted(repositories.items()):
            observed = _command(
                ["git", "rev-parse", "HEAD"], cwd=root / repository.local_path
            )
            if observed["returncode"] or observed["stdout"].strip() != repository.pinned_head:
                raise RuntimeError(f"repository head mismatch: {repository_id}: {observed}")
            report["repositories"][repository_id] = {
                "spec": asdict(repository),
                "observed_head": observed["stdout"].strip(),
            }

        for model_id in experiment.model_ids:
            model = models[model_id]
            transition = residency.ensure_exclusive(
                model.expected_inference_key, model.context_length
            )
            client = LMStudioClient(model, timeout_seconds=1_800)
            discovery, resolved = client.resolve()
            tokenizer = tokenizer_for(model)
            report["models"][model_id] = {
                "spec": asdict(model),
                "config_hash": model.config_hash,
                "transition": transition.to_dict(),
                "resolved": resolved.to_dict(),
                "discovery_errors": discovery.endpoint_errors,
                "tokenizer_path": str(tokenizer.path),
                "tokenizer_sha256": tokenizer.sha256,
                "tool_probe": _tool_probe(client, resolved.inference_key),
                "unload": residency.unload_all().to_dict(),
            }
        report["passed"] = True
        return report
    finally:
        cleanup_errors: list[str] = []
        try:
            report["final_unload"] = residency.unload_all().to_dict()
        except Exception as cleanup_error:
            cleanup_errors.append(f"unload_all: {cleanup_error}")
        try:
            status = server.status()
            report["server_stop"] = (
                server.stop()
                if status["running"]
                else {"action": "already_stopped", "status": status}
            )
        except Exception as cleanup_error:
            cleanup_errors.append(f"server_stop: {cleanup_error}")
        report["cleanup_errors"] = cleanup_errors
        report["finished_unix"] = time.time()


def main() -> None:
    root = Path(__file__).resolve().parents[1]
    output = root / "results" / "reports" / "study3_preflight.json"
    output.parent.mkdir(parents=True, exist_ok=True)
    report: dict[str, Any] = {}
    try:
        report = run(root)
    except Exception as exc:
        report = {**report, "passed": False, "error": repr(exc)}
        output.write_text(
            json.dumps(report, indent=2, sort_keys=True, default=str) + "\n",
            encoding="utf-8",
        )
        raise
    output.write_text(
        json.dumps(report, indent=2, sort_keys=True, default=str) + "\n",
        encoding="utf-8",
    )
    print(json.dumps({"passed": True, "report": str(output)}, indent=2))


if __name__ == "__main__":
    main()