feat(context): default supported Codex and benchmark delivery to on

apps/presentation/dashboard/src/features/personal-workspace/machine-configuration-settings.tsx

@@ -529,15 +529,17 @@ export function MachineConfigurationSettings({ section, onChanged }: {
             <div className="personal-capability-fields"><label className="is-boolean">
               <span>Codex</span>
               <input aria-label={locale === "zh-CN" ? "在 Codex 中使用差分上下文" : "Use differential context in Codex"}
-                type="checkbox" role="switch" checked={selectedCurrent?.mode === "on"} disabled={Boolean(busy) || !editorAvailable}
+                type="checkbox" role="switch" checked={(selectedCurrent?.mode ?? selected.default?.mode) === "on"} disabled={Boolean(busy) || !editorAvailable}
                 onChange={(event) => void setContextDelivery(event.target.checked ? "on" : "off")} />
             </label></div>
-            <p role="status">{selectedCurrent?.mode === "on"
-              ? locale === "zh-CN" ? "已为 Codex 保存开启偏好；宿主就绪与实际使用尚未验证。" : "Codex preference is on; host readiness and actual use are not verified."
-              : locale === "zh-CN" ? "Codex 未开启。" : "Off for Codex."}</p>
+            <p role="status">{(selectedCurrent?.mode ?? selected.default?.mode) === "on"
+              ? selectedCurrent
+                ? locale === "zh-CN" ? "已开启；原生初始化与实际使用尚未验证。" : "On; native setup and use unverified."
+                : locale === "zh-CN" ? "默认开启；原生初始化与实际使用尚未验证。" : "Default on; native setup and use unverified."
+              : locale === "zh-CN" ? "Codex 已关闭。" : "Off for Codex."}</p>
             <ContextHelp label={locale === "zh-CN" ? "Codex 支持范围与生效条件" : "Codex support and activation requirements"}><p>{locale === "zh-CN"
-              ? "目前仅支持 Codex App 心跳和 benchmark 心跳，无法确认或不支持的跑法报错。保存偏好不代表当前会话已采用差分，也不会改变其他应用或显式关闭的实验。宿主准备就绪且旧上下文完整时才使用差分，否则保留完整上下文。私有存储与会话隔离自动管理；每轮回执显示实际传输方式。"
-              : "Currently supports only Codex App and benchmark heartbeats; unknown or unsupported declarations return an error. Saving a preference does not establish use in the current session or change other applications or explicitly disabled runs. Delta requires a ready host and complete retained context; otherwise full context is preserved. Private storage and session isolation are automatic; each delivery reports its actual mode."}</p></ContextHelp>
+              ? "受支持的 POSIX Codex App／SSH 和 benchmark 心跳默认开启；已保存的关闭继续生效，benchmark trial 可单独显式关闭。其他应用和平台保留完整交付；显式开启遇到未知或冲突跑法会报错。保存偏好不代表当前会话已采用差分。宿主就绪且旧上下文完整时才使用差分，否则完整恢复；私有存储与会话隔离自动管理。"
+              : "Defaults on for supported POSIX Codex App/SSH and benchmark heartbeats. Saved off is preserved; benchmark trials can explicitly disable it independently. Other applications and platforms retain full delivery. Explicit on rejects unknown or conflicting modes. Saving a preference does not establish current-session adoption. Delta needs a ready host and complete retained context; otherwise full context is restored. Private storage and session isolation are automatic."}</p></ContextHelp>
           </section> : null}
           {editorAvailable && selected.capability_id !== "context_delivery" ? <>{editorMode === "json" || !selected.configuration_editor.fields.some((field) => field.key === "enabled" && field.input_kind === "boolean") ? <div className="personal-capability-editor-mode">
             <button disabled={Boolean(busy)} onClick={() => changeMode(editorMode === "guided" ? "json" : "guided")} type="button">

benchmark/README.md

@@ -878,7 +878,10 @@ must actually deliver a complete full packet and later deltas across worker
 wakes; every fresh receiver must begin with a complete current full packet.
 The static inference follows model-visible page cursors, and a separate byte-map
 oracle checks complete delivery and retained delta bases before candidate work.
-The default remains `off`. Neither mode measures paid inference, task quality or
+This joint qualifier deliberately retains its explicit `off` control by default;
+new benchmark heartbeat workers default on and can select `--context-delivery off`.
+Do not mutate a frozen trial to adopt that changed default.
+Neither qualifier mode measures paid inference, task quality or
 fees; reusing a point still does not count its source test as passed in this run.
 
 The public fixture runs six actual native work wakes, captures the pending

benchmark/edgebench/run.py

@@ -29,7 +29,7 @@
 from sforge.harness.run_agent import run_agent
 from sforge.harness.task_spec import make_task_spec
 
-from benchmark.runtime.codex import TASK_ENTRIES, turn_envelope_receipt
+from benchmark.runtime.codex import Execution, TASK_ENTRIES, turn_envelope_receipt
 from loopx.reasoning_effort import REASONING_EFFORTS
 from loopx.control_plane.work_items.stride_policy_codec import stride_strategy_options
 from benchmark.runtime.sforge import (
@@ -121,8 +121,8 @@ def main(argv=None):
                         help="New-run default: native public feedback; blind/best-only are explicit controls")
     parser.add_argument("--turn-envelope", action="store_true",
                         help="Opt-in short heartbeat context with same-invocation full captures")
-    parser.add_argument("--context-delivery", choices=("off", "on"), default="off",
-                        help="Experimental native Codex benchmark heartbeat reuse only; off by default")
+    parser.add_argument("--context-delivery", choices=("off", "on"),
+                        help="Native Codex context delivery: on by default for heartbeat workers; pass off to disable")
     parser.add_argument("--provider-request-admission", action="store_true",
                         help="Opt-in canonical admission before each heartbeat custom Responses request")
     parser.add_argument("--native-goal-continue", action="store_true",
@@ -197,6 +197,8 @@ def main(argv=None):
         parser.error("--task-entry loopx-planned requires a heartbeat worker")
     if args.turn_envelope and args.worker not in {"heartbeat-resume", "heartbeat-explore"}:
         parser.error("--turn-envelope requires a heartbeat worker")
+    if args.context_delivery is None:
+        args.context_delivery = Execution(mode="heartbeat" if args.worker.startswith("heartbeat-") else "plain").context_delivery
     if args.context_delivery != "off" and not args.worker.startswith("heartbeat-"):
         parser.error("--context-delivery requires a heartbeat worker")
     if args.provider_request_admission:

benchmark/runtime/RUNTIME.md

@@ -45,13 +45,17 @@ fail before execution. Set it back to `false` for the next run to roll back;
 do not change a running trial's treatment. Captures contain private Goal context.
 Transport tests do not establish lower token cost or better model decisions.
 
-For an explicitly selected native context-delivery treatment, add
-`context_delivery: "on"` to the same heartbeat agent `kwargs` (default `"off"`).
+New native heartbeat trials default to context delivery `on`. Add
+`context_delivery: "off"` to the same heartbeat agent `kwargs` to disable it.
+An explicit off is pinned into both native tools and worker preflight; it cannot
+inherit a machine-wide on. Plain, native-Goal and Turn execution remain off.
 Only the existing Codex heartbeat worker accepts `"on"`; booleans, other modes and
 scenario names are rejected. This is an experimental transport setting, separate
 from `turn_envelope`.
 The existing EdgeBench/SForge entry exposes the same selection through
-`--context-delivery on` on a heartbeat worker (`off` is the default).
+`--context-delivery off` to disable a heartbeat worker's on default. Existing
+running or frozen trials retain their recorded source and treatment; do not
+change them in place or infer a matched comparison from the new default.
 The worker preserves its existing `generic_cli` scheduler profile. Its execution
 owner requires `mode=heartbeat` and explicitly propagates `LOOPX_EXECUTION_MODE=heartbeat`
 and `LOOPX_CONTEXT_DELIVERY=on` into the native tool environment. Generic CLI

benchmark/runtime/codex.py

@@ -28,11 +28,13 @@ class Execution:
     validation_command: tuple[str, ...] = ()
     task_entry: str | None = None
     turn_envelope: bool = False
-    context_delivery: str = "off"
+    context_delivery: str | None = None
 
     def __post_init__(self) -> None:
         if self.context is None:
             object.__setattr__(self, "context", "resume" if self.mode in {"heartbeat", "turn"} else "fresh")
+        if self.context_delivery is None:
+            object.__setattr__(self, "context_delivery", "on" if self.mode == "heartbeat" else "off")
         if self.context_delivery not in ("off", "on") or (self.context_delivery != "off" and self.mode != "heartbeat"):
             raise ValueError("context_delivery supports only off or on with heartbeat execution")
         if not isinstance(self.turn_envelope, bool) or (self.turn_envelope and self.mode != "heartbeat"):
@@ -157,7 +159,8 @@ def codex_settings(
             # policy explicitly so model tools cannot silently lose it.
             "[shell_environment_policy.set]",
             'LOOPX_USAGE_PING = "0"',
-            *(['LOOPX_CONTEXT_DELIVERY = "on"', 'LOOPX_EXECUTION_MODE = "heartbeat"'] if execution.context_delivery != "off" else []),
+            f"LOOPX_CONTEXT_DELIVERY = {json.dumps(execution.context_delivery)}",
+            *(['LOOPX_EXECUTION_MODE = "heartbeat"'] if execution.context_delivery != "off" else []),
             "[features]",
             f"goals = {str(execution.native_goal).lower()}",
             "unified_exec = true",

benchmark/runtime/harbor.py

@@ -74,7 +74,7 @@ def __init__(
         task_entry=None,
         planning_timeout_sec=None,
         turn_envelope=False,
-        context_delivery="off",
+        context_delivery=None,
         provider_request_admission=False,
         native_goal_continue=False,
         task_objective=None,
@@ -634,7 +634,7 @@ def _worker_env(self, *, cwd: str) -> dict[str, str]:
                 **({"LOOPX_CHECKPOINT_REQUEST": self.checkpoint_request_path}
                    if self.checkpoint_request_path is not None else {}),
                 **({"LOOPX_TURN_ENVELOPE": "1"} if self.execution.turn_envelope else {}),
-                **({"LOOPX_CONTEXT_DELIVERY": "on"} if self.execution.context_delivery != "off" else {}),
+                "LOOPX_CONTEXT_DELIVERY": self.execution.context_delivery,
                 **({"LOOPX_PROVIDER_REQUEST_ADMISSION": "on"} if self.provider_request_admission else {}),
                 **({"LOOPX_NATIVE_GOAL_CONTINUE": "1"} if self.native_goal_continue else {}),
                 "LOOPX_ITERATION_CONTEXT": self.execution.context,

benchmark/runtime/sforge.py

@@ -93,7 +93,7 @@ def __init__(self, config, *, profile: str, cwd: str,
                  task_objective: str | None = None,
                  task_entry: str | None = None,
                  turn_envelope: bool = False,
-                 context_delivery: str = "off",
+                 context_delivery: str | None = None,
                  provider_request_admission: bool = False,
                  native_goal_continue: bool = False,
                  replan_after_turns: int | None = None,

benchmark/runtime/worker.py

@@ -450,6 +450,10 @@ def native_command(env, execution, stage, wake, session_wake):
     )
     if execution.context_delivery != "off":
         command[-1:-1] = json.loads(env.get("LOOPX_CODEX_CONTEXT_TRUST_ARGS", "[]"))
+        if env.get("LOOPX_CONTEXT_DELIVERY") == "off":
+            # Setup failure narrows this wake to full delivery, including the
+            # native shell whose frozen settings still record the requested on.
+            command[-1:-1] = ["-c", 'shell_environment_policy.set.LOOPX_CONTEXT_DELIVERY="off"']
     if stage == "plan":
         command[-1:-1] = ["-c", "features.goals=false"]
     return command
@@ -490,7 +494,7 @@ def run_once(env: dict[str, str]) -> dict:
         validation_command=json.loads(env.get("LOOPX_VALIDATION_COMMAND_JSON", "[]")),
         task_entry=env.get("LOOPX_TASK_ENTRY"),
         turn_envelope=env.get("LOOPX_TURN_ENVELOPE", "0") == "1",
-        context_delivery=env.get("LOOPX_CONTEXT_DELIVERY", "off"),
+        context_delivery=env.get("LOOPX_CONTEXT_DELIVERY"),
     )
     request_admission = False
     if env.get("LOOPX_PROVIDER_REQUEST_ADMISSION", "off") != "off":
@@ -512,7 +516,7 @@ def run_once(env: dict[str, str]) -> dict:
     home = Path(env["LOOPX_CODEX_HOME"])
     env = process_environment(home, base=env)
     # Preserve this trial's resolved off/on choice in both preflight and solver
-    # subprocesses; an omitted off must not inherit a machine-wide opt-in.
+    # subprocesses; an explicit off must not inherit a machine-wide default.
     env["LOOPX_CONTEXT_DELIVERY"] = execution.context_delivery
     # The resolved trial treatment must not inherit an ambient parent opt-in.
     if request_admission:
@@ -602,6 +606,7 @@ def run_once(env: dict[str, str]) -> dict:
                 # trial. Unavailable native hooks keep ordinary full IO.
                 env.pop("LOOPX_CODEX_CONTEXT_DIR", None)
                 env.pop("LOOPX_CODEX_CONTEXT_TRUST_ARGS", None)
+                env["LOOPX_CONTEXT_DELIVERY"] = "off"
                 receipt["context_transport"] = {"available": False, **_error_tail(exc)}
         body = "Finish the task."
         if stage == "plan":

benchmark/tests/test_resume_sessions.py

@@ -122,6 +122,8 @@ def resume_env(tmp_path, monkeypatch):
     skills.mkdir()
     env.update(
         LOOPX_EXECUTION_MODE="heartbeat",
+        # These legacy resume fixtures pin their original transport treatment.
+        LOOPX_CONTEXT_DELIVERY="off",
         LOOPX_ITERATION_CONTEXT="resume",
         LOOPX_RUNTIME_ROOT=str(tmp_path / "runtime"),
         LOOPX_REGISTRY=str(tmp_path / "registry.json"),

benchmark/tests/test_sforge_runtime.py

@@ -277,7 +277,7 @@ async def installed(self, environment):
     assert "turn_envelope" not in profile
     assert "turn_envelope_source" not in profile
     assert "LOOPX_TURN_ENVELOPE" not in env
-    assert env.get("LOOPX_CONTEXT_DELIVERY") == ("on" if context_delivery != "off" else None)
+    assert env["LOOPX_CONTEXT_DELIVERY"] == context_delivery
     assert profile.get("context_delivery") == (context_delivery if context_delivery != "off" else None)
     assert profile.get("checkpoint_capture_enabled") is (True if capture_path else None)
     with pytest.raises(ValueError, match="checkpoint_request_path"):
@@ -1414,7 +1414,7 @@ def stop_before_solver(**kwargs):
 
 
 @pytest.mark.parametrize("capture", [False, True])
-@pytest.mark.parametrize("context_delivery", ["off", "on"])
+@pytest.mark.parametrize("context_delivery", [None, "off", "on"])
 def test_native_checkpoint_storage_guidance_reaches_both_prompt_surfaces(tmp_path, monkeypatch, capture, context_delivery):
     pytest.importorskip("sforge")
     pytest.importorskip("harbor")
@@ -1433,9 +1433,10 @@ def test_native_checkpoint_storage_guidance_reaches_both_prompt_surfaces(tmp_pat
         game_mode=False, submit_paths=["solver.py"], work=SimpleNamespace(agent_query="Original acceptance."),
         judge=SimpleNamespace(selection="score_first", score_direction="maximize")))
     monkeypatch.setattr(run, "RecordingDockerBackend", lambda **k: SimpleNamespace(image_exists=lambda image: True))
+    expected_delivery = "on" if context_delivery is None else context_delivery
     def before_solver(**kwargs):
         worker = kwargs["agent"]
-        assert worker.context_delivery is context_delivery
+        assert worker.context_delivery == expected_delivery
         assert worker.task_objective == "Original acceptance."
         assert worker.feedback_prompt.startswith(worker.workspace_instructions)
         assert worker.feedback_prompt.endswith("---\n\nOriginal acceptance.\n")
@@ -1450,8 +1451,8 @@ def before_solver(**kwargs):
     args = ["--task", "fixture", "--tasks-dir", str(tmp_path), "--log-dir", str(tmp_path),
             "--run-id", "storage", "--worker", "heartbeat-explore", "--model", "fixture",
             "--effort", "xhigh", "--judge-url", "http://127.0.0.1:9999", "--feedback", "native"]
-    if context_delivery != "off":
-        args += ["--context-delivery", "on"]
+    if context_delivery is not None:
+        args += ["--context-delivery", context_delivery]
     if capture:
         declaration = tmp_path / "request.json"
         declaration.write_text(json.dumps({"destination": "/logs/agent/checkpoints",
@@ -1460,4 +1461,4 @@ def before_solver(**kwargs):
     with pytest.raises(RuntimeError, match="validated before model execution"):
         run.main(args)
     receipt = json.loads((tmp_path / "runs/storage/fixture/runtime-receipt.json").read_text())
-    assert receipt.get("context_delivery") == (context_delivery if context_delivery != "off" else None)
+    assert receipt.get("context_delivery") == (expected_delivery if expected_delivery != "off" else None)