agentengine-sdk-python 0.8.4__py3-none-any.whl → 0.8.5__py3-none-any.whl

ksadk/skills/runtime/backends/local.py

@@ -2,41 +2,69 @@ from __future__ import annotations
 
 import json
 import os
+import shutil
 import subprocess
 import sys
 import tempfile
 import time
+from dataclasses import replace
 from pathlib import Path
+from types import TracebackType
+from uuid import uuid4
 
+from ksadk._process import starts_new_process_group, terminate_process_group
+from ksadk.sandbox.local_controls import LocalControlSettings
+from ksadk.skills.events import (
+    SKILL_EVENT_FILE_ENV,
+    SkillEvent,
+    SkillInvocationPlan,
+    parse_sandbox_skill_event_lines,
+)
 from ksadk.skills.package_store import SkillPackage
-from ksadk.skills.runtime.artifact_delivery import export_artifacts, import_artifacts
+from ksadk.skills.runtime.artifact_delivery import (
+    MAX_FILES,
+    ArtifactDeliveryError,
+    export_artifacts,
+    import_artifacts,
+)
 from ksadk.skills.runtime.base import (
     SandboxInputFile,
     SkillRuntimeError,
     SkillRuntimeResult,
     format_skill_names_env,
     normalize_skill_names,
-    parse_output_files,
+    parse_workflow_result,
+    sandbox_runtime_env,
 )
 from ksadk.skills.runtime.pinned import stage_packages
 
-
-def _coerce_output(value: str | bytes | None) -> str:
-    if isinstance(value, bytes):
-        return value.decode("utf-8", errors="replace")
-    return value or ""
+_MAX_EVENT_FILE_BYTES = 1024 * 1024
+_MAX_PROCESS_OUTPUT_BYTES = 4 * 1024 * 1024
 
 
 class LocalProcessSkillRuntimeBackend:
-    def __init__(self, agent_path: str | Path, timeout: int = 900):
+    """Run the Skill runtime in a host process with request-scoped storage.
+
+    ``artifact_directory`` is a caller-owned parent for verified artifact
+    snapshots. When omitted, each non-empty result is delivered to a temporary
+    directory that the caller must remove after consuming ``output_files``.
+    """
+
+    def __init__(
+        self,
+        agent_path: str | Path,
+        timeout: int = 900,
+        artifact_directory: Path | None = None,
+        controls: LocalControlSettings | None = None,
+    ):
         self.agent_path = Path(agent_path)
         self.timeout = timeout
+        self.artifact_directory = artifact_directory
+        self.controls = controls or LocalControlSettings.from_trusted_host_env()
 
     @classmethod
     def from_env(cls) -> "LocalProcessSkillRuntimeBackend":
-        agent_path = os.environ.get("KSADK_SKILL_RUNTIME_AGENT_PATH") or str(
-            Path(__file__).resolve().parents[1] / "agent.py"
-        )
+        agent_path = os.environ.get("KSADK_SKILL_RUNTIME_AGENT_PATH") or str(_bundled_agent_path())
         timeout = int(os.environ.get("KSADK_SKILL_RUNTIME_TIMEOUT", "900"))
         return cls(agent_path=agent_path, timeout=timeout)
 
@@ -49,124 +77,598 @@ class LocalProcessSkillRuntimeBackend:
         skill_names: list[str] | None = None,
         env: dict[str, str] | None = None,
         input_files: list[SandboxInputFile] | None = None,
+        invocation_plan: SkillInvocationPlan | None = None,
         pinned_packages: list[SkillPackage] | None = None,
         timeout: int = 900,
     ) -> SkillRuntimeResult:
+        del input_files  # Local Skill input staging is not part of the current contract.
         started = time.monotonic()
-        if pinned_packages is not None and self.agent_path.resolve() != (
-            Path(__file__).resolve().parents[1] / "agent.py"
-        ):
-            raise ValueError("Pinned Skill execution requires the bundled runtime agent")
-        runtime_env = (
-            {
-                key: value
-                for key, value in os.environ.items()
-                if key in {"PATH", "LANG", "LC_ALL", "TMPDIR", "SYSTEMROOT"}
-            }
-            if pinned_packages is not None
-            else os.environ.copy()
-        )
-        runtime_env.update(env or {})
-        runtime_env["KSADK_SKILL_SPACE_IDS"] = ",".join(skill_space_ids)
-        runtime_env["SKILL_SPACE_ID"] = skill_space_ids[0] if skill_space_ids else ""
-        if pinned_packages is None and (
-            public_spaces := os.environ.get("KSADK_PUBLIC_SKILL_SPACE_IDS")
-        ):
-            runtime_env["KSADK_PUBLIC_SKILL_SPACE_IDS"] = public_spaces
-        selected_skill_names = format_skill_names_env(skill_names)
-        if selected_skill_names:
-            runtime_env["KSADK_SELECTED_SKILL_NAMES"] = selected_skill_names
-        else:
-            runtime_env.pop("KSADK_SELECTED_SKILL_NAMES", None)
+        runtime_id = f"local:{session_id}:{uuid4().hex}"
+        lifecycle_attributes = {
+            "backend": "local_process",
+            "cleanup_scope": "request_directory",
+        }
+        sandbox: dict[str, object] = {
+            "backend": "local_process",
+            "runtime_id": runtime_id,
+            "creation_status": "unavailable",
+            "instance_status": "not_created",
+            "cleanup_status": "not_started",
+            "cleanup_error": None,
+            "cleanup_scope": "request_directory",
+        }
+        skill_events: list[SkillEvent] = []
+        request_root: Path | None = None
+        event_path: Path | None = None
+        process: subprocess.Popen[bytes] | None = None
+        result: SkillRuntimeResult | None = None
+        primary_error: BaseException | None = None
+        primary_traceback: TracebackType | None = None
+        process_cleanup_error: str | None = None
+        failure_stage = "create"
+
         try:
-            with tempfile.TemporaryDirectory(prefix="ksadk-skill-runtime-") as tmp_dir:
-                request_path = Path(tmp_dir) / "workflow-request.json"
-                request = {
-                    "workflow_prompt": workflow_prompt,
-                    "skill_names": normalize_skill_names(skill_names),
+            if pinned_packages is not None and self.agent_path.resolve() != _bundled_agent_path():
+                raise ValueError("Pinned Skill execution requires the bundled runtime agent")
+
+            request_parent = _request_parent(env)
+            request_root = Path(
+                tempfile.mkdtemp(prefix="ksadk-skill-request-", dir=request_parent)
+            ).resolve()
+            work_dir = request_root / "work"
+            package_dir = request_root / "packages"
+            process_tmp_dir = request_root / "tmp"
+            for directory in (work_dir, package_dir, process_tmp_dir):
+                directory.mkdir(mode=0o700)
+
+            runtime_env = _runtime_environment(
+                env,
+                pinned=pinned_packages is not None,
+                extra_env_names=self.controls.extra_env_names,
+            )
+            # Apply controls after caller environment. A Skill invocation cannot
+            # disable or weaken the host-selected launch policy.
+            runtime_env.update(self.controls.runtime_environment())
+            runtime_env.update(
+                {
+                    "KSADK_SKILL_SPACE_IDS": ",".join(skill_space_ids),
+                    "SKILL_SPACE_ID": skill_space_ids[0] if skill_space_ids else "",
+                    "KSADK_SKILL_WORKDIR": str(work_dir),
+                    "TMPDIR": str(process_tmp_dir),
                 }
-                if pinned_packages is not None:
-                    entries = stage_packages(pinned_packages, Path(tmp_dir))
-                    request["pinned_packages"] = [entry.model_dump() for entry in entries]
-                    request["pinned_protocol_version"] = 1
-                    if not runtime_env.get("KSADK_SKILL_WORKDIR"):
-                        # Artifacts outlive the temporary archive delivery directory.
-                        runtime_env["KSADK_SKILL_WORKDIR"] = tempfile.mkdtemp(
-                            prefix="ksadk-pinned-artifacts-"
+            )
+            if pinned_packages is None and (
+                public_spaces := os.environ.get("KSADK_PUBLIC_SKILL_SPACE_IDS")
+            ):
+                runtime_env["KSADK_PUBLIC_SKILL_SPACE_IDS"] = public_spaces
+            selected_skill_names = format_skill_names_env(skill_names)
+            if selected_skill_names:
+                runtime_env["KSADK_SELECTED_SKILL_NAMES"] = selected_skill_names
+            else:
+                runtime_env.pop("KSADK_SELECTED_SKILL_NAMES", None)
+
+            request_path = package_dir / "workflow-request.json"
+            event_path = request_root / "skill-events.jsonl"
+            runtime_env[SKILL_EVENT_FILE_ENV] = str(event_path)
+            request_payload: dict[str, object] = {
+                "workflow_prompt": workflow_prompt,
+                "skill_names": normalize_skill_names(skill_names),
+            }
+            if invocation_plan is not None:
+                request_payload["invocation_plan"] = _request_invocation_plan(invocation_plan)
+            failure_stage = "initialize"
+            if pinned_packages is not None:
+                entries = stage_packages(pinned_packages, package_dir)
+                request_payload["pinned_packages"] = [entry.model_dump() for entry in entries]
+                request_payload["pinned_protocol_version"] = 1
+            request_path.write_text(
+                json.dumps(request_payload, ensure_ascii=False), encoding="utf-8"
+            )
+
+            stdout_path = request_root / "stdout.log"
+            stderr_path = request_root / "stderr.log"
+            effective_timeout = timeout or self.timeout
+            command = _agent_command(self.agent_path, request_path, pinned_packages is not None)
+            failure_stage = "create"
+            try:
+                with (
+                    stdout_path.open("wb") as stdout_stream,
+                    stderr_path.open("wb") as stderr_stream,
+                ):
+                    process = subprocess.Popen(
+                        command,
+                        cwd=str(work_dir),
+                        env=runtime_env,
+                        stdout=stdout_stream,
+                        stderr=stderr_stream,
+                        start_new_session=starts_new_process_group(),
+                    )
+                    sandbox.update(creation_status="completed", instance_status="created")
+                    skill_events.append(
+                        SkillEvent.create(
+                            "sandbox.session.created",
+                            status="completed",
+                            runtime_id=runtime_id,
+                            attributes=lifecycle_attributes,
                         )
-                    # Use the same canonical root in the child and host validator
-                    # (macOS temporary directories may be reached through /var).
-                    runtime_env["KSADK_SKILL_WORKDIR"] = str(
-                        Path(runtime_env["KSADK_SKILL_WORKDIR"]).resolve()
                     )
-                request_path.write_text(
-                    json.dumps(
-                        request,
-                        ensure_ascii=False,
-                    ),
-                    encoding="utf-8",
+                    failure_stage = "execute"
+                    timed_out = False
+                    try:
+                        process.wait(timeout=effective_timeout)
+                    except subprocess.TimeoutExpired:
+                        timed_out = True
+                    finally:
+                        process_cleanup_error = terminate_process_group(process)
+                    exit_code = process.returncode
+                    process = None
+            except (OSError, ValueError) as exc:
+                sandbox.update(creation_status="failed", instance_status="unknown")
+                result = SkillRuntimeResult(
+                    runtime_id=runtime_id,
+                    exit_code=None,
+                    duration_ms=int((time.monotonic() - started) * 1000),
+                    error_type=type(exc).__name__,
+                    error_message=str(exc),
                 )
-                completed = subprocess.run(
-                    [
-                        sys.executable,
-                        *(["-I"] if pinned_packages is not None else []),
-                        "-u",
-                        str(self.agent_path),
-                        "--request-file",
-                        str(request_path),
-                    ],
-                    text=True,
-                    capture_output=True,
-                    timeout=timeout or self.timeout,
-                    env=runtime_env,
-                    check=False,
+            else:
+                stdout = _read_output(stdout_path)
+                stderr = _read_output(stderr_path)
+                if timed_out:
+                    if process_cleanup_error:
+                        sandbox.update(
+                            artifact_collection_status="failed",
+                            artifact_collection_error="process_cleanup_failed",
+                        )
+                        recovered_stdout, recovered_files = stdout, []
+                    else:
+                        recovered_stdout, recovered_files = self._recover_artifacts(
+                            stdout, work_dir, request_root, sandbox
+                        )
+                    parsed = parse_workflow_result(recovered_stdout)
+                    _record_control_observation(recovered_stdout, sandbox)
+                    result = SkillRuntimeResult(
+                        runtime_id=runtime_id,
+                        exit_code=None,
+                        stdout=recovered_stdout,
+                        stderr=stderr,
+                        duration_ms=int((time.monotonic() - started) * 1000),
+                        timed_out=True,
+                        output_files=recovered_files,
+                        output_text=parsed.output_text,
+                        output_text_truncated=parsed.output_text_truncated,
+                        workflow_status=parsed.workflow_status,
+                        executed_skill=parsed.executed_skill,
+                        instructions=parsed.instructions,
+                        error_type="TimeoutExpired",
+                        error_message=f"Skill workflow timed out after {effective_timeout}s",
+                    )
+                elif exit_code == 0 and process_cleanup_error:
+                    failure_stage = "cleanup"
+                    result = SkillRuntimeResult(
+                        runtime_id=runtime_id,
+                        exit_code=exit_code,
+                        stdout=stdout,
+                        stderr=stderr,
+                        duration_ms=int((time.monotonic() - started) * 1000),
+                        error_type="ProcessCleanupError",
+                        error_message="Local Skill runtime process group could not be stopped",
+                    )
+                else:
+                    strict_collection = exit_code == 0
+                    failure_stage = "collect" if strict_collection else "execute"
+                    if process_cleanup_error:
+                        sandbox.update(
+                            artifact_collection_status="failed",
+                            artifact_collection_error="process_cleanup_failed",
+                        )
+                        delivered_stdout, output_files = stdout, []
+                    else:
+                        delivered_stdout, output_files = self._deliver_artifacts(
+                            stdout,
+                            work_dir,
+                            request_root,
+                            sandbox,
+                            strict=strict_collection,
+                        )
+                    parsed = parse_workflow_result(delivered_stdout)
+                    _record_control_observation(delivered_stdout, sandbox)
+                    result = SkillRuntimeResult(
+                        runtime_id=runtime_id,
+                        exit_code=exit_code,
+                        stdout=delivered_stdout,
+                        stderr=stderr,
+                        duration_ms=int((time.monotonic() - started) * 1000),
+                        output_files=output_files,
+                        output_text=parsed.output_text,
+                        output_text_truncated=parsed.output_text_truncated,
+                        workflow_status=parsed.workflow_status,
+                        executed_skill=parsed.executed_skill,
+                        instructions=parsed.instructions,
+                        error_type=parsed.command_error_type or None,
+                        error_message=parsed.error or None,
+                    )
+                    if exit_code == 0:
+                        failure_stage = ""
+        except BaseException as exc:
+            primary_error = exc
+            primary_traceback = exc.__traceback__
+        finally:
+            if process is not None:
+                process_cleanup_error = terminate_process_group(process) or process_cleanup_error
+            if event_path is not None:
+                try:
+                    skill_events.extend(
+                        replace(event, runtime_id=event.runtime_id or runtime_id)
+                        for event in _read_request_events(
+                            event_path,
+                            expected_invocations=_expected_invocations(invocation_plan),
+                        )
+                    )
+                except Exception:
+                    pass
+
+            cleanup_error = process_cleanup_error
+            if request_root is not None:
+                try:
+                    _remove_request_directory(request_root)
+                except Exception:
+                    cleanup_error = cleanup_error or "request_directory_cleanup_failed"
+            if cleanup_error:
+                sandbox.update(cleanup_status="failed", cleanup_error="cleanup_failed")
+                if not failure_stage:
+                    failure_stage = "cleanup"
+                skill_events.append(
+                    SkillEvent.create(
+                        "sandbox.session.cleanup_failed",
+                        status="failed",
+                        runtime_id=runtime_id,
+                        error_category="cleanup_failed",
+                        attributes=lifecycle_attributes,
+                    )
                 )
-                stdout = completed.stdout
-                output_files = parse_output_files(stdout)
-                if pinned_packages is not None:
-                    payloads = [
-                        json.loads(line.split("=", 1)[1])
-                        for line in stdout.splitlines()
-                        if line.startswith("workflow_result=")
-                    ]
-                    if len(payloads) != 1 or not isinstance(payloads[0], dict):
-                        raise SkillRuntimeError("Runtime did not return a unique workflow result")
-                    payload = payloads[0]
-                    paths = payload.get("output_files")
-                    if not isinstance(paths, list) or any(not isinstance(p, str) for p in paths):
-                        raise SkillRuntimeError("Runtime returned invalid artifact paths")
-                    # Snapshot only bounded regular files from the admitted workspace.
-                    # A workflow's stdout is not authority to read arbitrary host files.
-                    bundle_path = Path(tmp_dir) / "artifacts.zip"
-                    receipt = export_artifacts(
-                        paths, Path(runtime_env["KSADK_SKILL_WORKDIR"]), bundle_path
+            elif request_root is not None:
+                sandbox.update(cleanup_status="completed", cleanup_error=None)
+                skill_events.append(
+                    SkillEvent.create(
+                        "sandbox.session.cleaned_up",
+                        status="completed",
+                        runtime_id=runtime_id,
+                        attributes=lifecycle_attributes,
                     )
-                    output_files = import_artifacts(bundle_path.read_bytes(), receipt)
-                    payload["output_files"] = output_files
-                    payload["artifacts"] = output_files
-                    payload["artifact_bundle"] = receipt.model_dump()
-                    stdout = "\n".join(
-                        "workflow_result=" + json.dumps(payload, ensure_ascii=False, sort_keys=True)
-                        if line.startswith("workflow_result=")
-                        else line
-                        for line in stdout.splitlines()
-                    ) + "\n"
-            return SkillRuntimeResult(
-                runtime_id=f"local:{session_id}",
-                exit_code=completed.returncode,
-                stdout=stdout,
-                stderr=completed.stderr,
-                duration_ms=int((time.monotonic() - started) * 1000),
-                output_files=output_files,
+                )
+            else:
+                sandbox.update(cleanup_status="not_started", cleanup_error=None)
+            if failure_stage:
+                sandbox["failure_stage"] = failure_stage
+
+        if primary_error is not None:
+            _attach_exception_diagnostics(primary_error, sandbox, skill_events)
+            raise primary_error.with_traceback(primary_traceback)
+        assert result is not None
+        return replace(result, skill_events=skill_events, sandbox=sandbox)
+
+    def _deliver_artifacts(
+        self,
+        stdout: str,
+        work_dir: Path,
+        request_root: Path,
+        sandbox: dict[str, object],
+        *,
+        strict: bool,
+    ) -> tuple[str, list[str]]:
+        try:
+            payload = _workflow_payload(stdout)
+            paths = payload.get("output_files")
+            if not isinstance(paths, list) or any(not isinstance(path, str) for path in paths):
+                raise SkillRuntimeError("Runtime returned invalid artifact paths")
+            bundle_path = request_root / "artifacts.zip"
+            receipt = export_artifacts(paths, work_dir, bundle_path)
+            output_files = import_artifacts(
+                bundle_path.read_bytes(), receipt, parent=self.artifact_directory
+            )
+            payload["output_files"] = output_files
+            payload["artifacts"] = output_files
+            payload["artifact_bundle"] = receipt.model_dump()
+            sandbox["artifact_collection_status"] = "completed"
+            return _rewrite_workflow_payload(stdout, payload), output_files
+        except Exception as exc:
+            sandbox.update(
+                artifact_collection_status="failed",
+                artifact_collection_error=type(exc).__name__,
+            )
+            if strict:
+                raise
+            return stdout, []
+
+    def _recover_artifacts(
+        self,
+        stdout: str,
+        work_dir: Path,
+        request_root: Path,
+        sandbox: dict[str, object],
+    ) -> tuple[str, list[str]]:
+        delivered_stdout, delivered_files = self._deliver_artifacts(
+            stdout, work_dir, request_root, sandbox, strict=False
+        )
+        if sandbox.get("artifact_collection_status") == "completed":
+            sandbox["artifact_collection_status"] = "recovered_partial"
+            return delivered_stdout, delivered_files
+        try:
+            known_paths = _known_output_artifacts(work_dir / "artifacts")
+            if not known_paths:
+                return stdout, []
+            bundle_path = request_root / "recovered-artifacts.zip"
+            receipt = export_artifacts(known_paths, work_dir, bundle_path)
+            output_files = import_artifacts(
+                bundle_path.read_bytes(), receipt, parent=self.artifact_directory
+            )
+            sandbox.update(
+                artifact_collection_status="recovered_partial",
+                artifact_collection_source="request_output_directory",
             )
-        except subprocess.TimeoutExpired as exc:
-            return SkillRuntimeResult(
-                runtime_id=f"local:{session_id}",
-                exit_code=None,
-                stdout=_coerce_output(exc.stdout),
-                stderr=_coerce_output(exc.stderr),
-                duration_ms=int((time.monotonic() - started) * 1000),
-                timed_out=True,
-                error_type="TimeoutExpired",
-                error_message=f"Skill workflow timed out after {timeout or self.timeout}s",
+            sandbox.pop("artifact_collection_error", None)
+            return stdout, output_files
+        except Exception as exc:
+            sandbox.update(
+                artifact_collection_status="failed",
+                artifact_collection_error=type(exc).__name__,
             )
+            return stdout, []
+
+
+def _runtime_environment(
+    env: dict[str, str] | None,
+    *,
+    pinned: bool,
+    extra_env_names: tuple[str, ...] = (),
+) -> dict[str, str]:
+    base_allowed = {
+        "PATH",
+        "LANG",
+        "LC_ALL",
+        "TMPDIR",
+        "SYSTEMROOT",
+        "KSADK_SKILL_RUNTIME_TIMEOUT",
+        "KSADK_SKILL_OUTPUT_TEXT_MAX_BYTES",
+        "KSADK_SKILL_ARTIFACT_PROJECT",
+    }
+    service_allowed = {
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "NO_PROXY",
+        "KSADK_SKILL_SERVICE_URL",
+        "KSADK_SKILL_SERVICE_ENDPOINT",
+        "KSADK_SKILL_SERVICE_SCHEME",
+        "KSADK_SKILL_SERVICE_ACCOUNT_ID",
+        "KSADK_SKILL_SERVICE_REGION",
+        "KSADK_SKILL_SERVICE_TOKEN",
+        "KSADK_SKILL_MANIFEST_TIMEOUT",
+        "KSADK_SKILL_CACHE_DIR",
+        "KSADK_LOCAL_SKILLS_DIR",
+        "KSADK_SKILL_ALLOW_HASH_MISMATCH",
+        "KSADK_PUBLIC_SKILL_ALLOWLIST",
+        "AGENTENGINE_PRE_CONTROL_REGION",
+        "KSYUN_REGION",
+    }
+    allowed = base_allowed | (set() if pinned else service_allowed)
+    allowed.update(extra_env_names)
+    runtime_env = {key: value for key, value in os.environ.items() if key in allowed}
+    # Explicit call configuration belongs to the trusted runtime. It does not
+    # become Skill script environment unless separately allowlisted.
+    runtime_env.update(
+        {
+            key: value
+            for key, value in (env or {}).items()
+            if not key.startswith("KSADK_LOCAL_PROCESS_")
+        }
+    )
+    return sandbox_runtime_env(runtime_env)
+
+
+def _request_parent(env: dict[str, str] | None) -> str | None:
+    configured = (env or {}).get("KSADK_SKILL_WORKDIR") or os.environ.get("KSADK_SKILL_WORKDIR")
+    if not configured:
+        return None
+    parent = Path(configured).expanduser().resolve()
+    parent.mkdir(parents=True, exist_ok=True)
+    return str(parent)
+
+
+def _bundled_agent_path() -> Path:
+    return (Path(__file__).resolve().parents[1] / "agent.py").resolve()
+
+
+def _trusted_sdk_root() -> Path:
+    # local.py is <trusted-root>/ksadk/skills/runtime/backends/local.py. The
+    # imported backend module, never cwd/PYTHONPATH, is the authority for this root.
+    return Path(__file__).resolve().parents[4]
+
+
+def _agent_command(agent_path: Path, request_path: Path, pinned: bool) -> list[str]:
+    if not pinned:
+        return [
+            sys.executable,
+            "-u",
+            str(agent_path),
+            "--request-file",
+            str(request_path),
+        ]
+    bootstrap = (
+        "import sys;"
+        "sys.path.insert(0,sys.argv.pop(1));"
+        "from ksadk.skills.runtime.agent import main;"
+        "raise SystemExit(main(sys.argv[1:]))"
+    )
+    return [
+        sys.executable,
+        "-I",
+        "-u",
+        "-c",
+        bootstrap,
+        str(_trusted_sdk_root()),
+        "--request-file",
+        str(request_path),
+    ]
+
+
+def _workflow_payload(stdout: str) -> dict[str, object]:
+    raw_payloads = [
+        line.split("=", 1)[1] for line in stdout.splitlines() if line.startswith("workflow_result=")
+    ]
+    if len(raw_payloads) != 1:
+        raise SkillRuntimeError("Runtime did not return a unique workflow result")
+    try:
+        payload = json.loads(raw_payloads[0])
+    except json.JSONDecodeError:
+        raise SkillRuntimeError("Runtime returned a malformed workflow result") from None
+    if not isinstance(payload, dict):
+        raise SkillRuntimeError("Runtime did not return a unique workflow result")
+    return payload
+
+
+def _record_control_observation(stdout: str, sandbox: dict[str, object]) -> None:
+    try:
+        payload = _workflow_payload(stdout)
+    except SkillRuntimeError:
+        return
+    commands = payload.get("commands")
+    if not isinstance(commands, list):
+        return
+    observations: list[dict[str, object]] = []
+    for command in commands:
+        if not isinstance(command, dict):
+            continue
+        controls = command.get("controls")
+        observation: dict[str, object] = {}
+        if isinstance(controls, dict):
+            status = controls.get("status")
+            if status in {"applied", "partial", "failed", "blocked"}:
+                observation["status"] = status
+            applied = controls.get("applied")
+            if isinstance(applied, dict):
+                observation["applied_limits"] = sorted(
+                    name
+                    for name, value in applied.items()
+                    if isinstance(name, str) and type(value) is int
+                )
+            unsupported = controls.get("unsupported")
+            if isinstance(unsupported, list):
+                observation["unsupported_limits"] = [
+                    item for item in unsupported if isinstance(item, str)
+                ][:10]
+            static_check = controls.get("static_check")
+            if isinstance(static_check, str):
+                observation["static_check"] = static_check
+        for key in (
+            "error_type",
+            "blocked_reason",
+            "resource_limit",
+            "timed_out",
+            "output_limit_exceeded",
+            "stdout_truncated",
+            "stderr_truncated",
+            "cleanup_status",
+        ):
+            value = command.get(key)
+            if isinstance(value, (str, bool)):
+                observation[key] = value
+        if observation:
+            observations.append(observation)
+    if observations:
+        sandbox["controls"] = observations
+
+
+def _rewrite_workflow_payload(stdout: str, payload: dict[str, object]) -> str:
+    return (
+        "\n".join(
+            "workflow_result=" + json.dumps(payload, ensure_ascii=False, sort_keys=True)
+            if line.startswith("workflow_result=")
+            else line
+            for line in stdout.splitlines()
+        )
+        + "\n"
+    )
+
+
+def _read_output(path: Path) -> str:
+    if not path.exists():
+        return ""
+    with path.open("rb") as source:
+        size = path.stat().st_size
+        if size > _MAX_PROCESS_OUTPUT_BYTES:
+            source.seek(size - _MAX_PROCESS_OUTPUT_BYTES)
+        content = source.read(_MAX_PROCESS_OUTPUT_BYTES)
+    return content.decode("utf-8", errors="replace")
+
+
+def _known_output_artifacts(output_dir: Path) -> list[str]:
+    """Find only a bounded set in the runtime's declared output directory."""
+
+    if not output_dir.is_dir():
+        return []
+    paths: list[str] = []
+    visited = 0
+    for root, directories, files in os.walk(output_dir, followlinks=False):
+        directories[:] = [name for name in directories if not (Path(root) / name).is_symlink()]
+        visited += len(directories) + len(files)
+        if visited > MAX_FILES * 4:
+            raise ArtifactDeliveryError("Artifact recovery scan limit exceeded")
+        for name in files:
+            paths.append(str(Path(root) / name))
+            if len(paths) > MAX_FILES:
+                raise ArtifactDeliveryError("Too many artifacts")
+    return sorted(paths)
+
+
+def _read_request_events(
+    path: Path, *, expected_invocations: dict[str, object] | None
+) -> list[SkillEvent]:
+    try:
+        with path.open("rb") as source:
+            content = source.read(_MAX_EVENT_FILE_BYTES + 1)
+        if len(content) > _MAX_EVENT_FILE_BYTES:
+            return []
+        return parse_sandbox_skill_event_lines(
+            content.decode("utf-8", errors="replace"),
+            expected_invocations=expected_invocations,
+        )
+    except OSError:
+        return []
+
+
+def _remove_request_directory(path: Path) -> None:
+    shutil.rmtree(path)
+
+
+def _attach_exception_diagnostics(
+    exc: BaseException, sandbox: dict[str, object], skill_events: list[SkillEvent]
+) -> None:
+    try:
+        setattr(exc, "sandbox", dict(sandbox))
+        setattr(exc, "skill_events", list(skill_events))
+    except Exception:
+        pass
+    if sandbox.get("cleanup_status") == "failed":
+        try:
+            exc.add_note("Local Skill runtime request cleanup failed; see .sandbox diagnostics")
+        except (AttributeError, TypeError):
+            pass
+
+
+def _request_invocation_plan(plan: SkillInvocationPlan | None) -> list[dict[str, str]]:
+    if plan is None:
+        return []
+    return [
+        {
+            "skill_id": entry.skill_ref.skill_id,
+            "skill_invocation_id": entry.skill_invocation_id,
+        }
+        for entry in plan.entries
+    ]
+
+
+def _expected_invocations(plan: SkillInvocationPlan | None) -> dict[str, object] | None:
+    if plan is None:
+        return None
+    return {entry.skill_invocation_id: entry.skill_ref for entry in plan.entries}