From f2e2fd0f8aa890e6a8108a9521c3f9b5228ca4de Mon Sep 17 00:00:00 2001 From: LoopX Agent <337587101+loopx-agent@users.noreply.github.com> Date: Wed, 18 Nov 2026 11:13:29 +0800 Subject: [PATCH 1/2] refactor(testing): use complete C1 source in host prompt qualification Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com> --- .../contract-and-context.md | 11 ++++ .../contract-and-context.zh-CN.md | 7 +++ .../testing/action_portfolio_scenarios.py | 7 +-- .../testing/host_prompt_behavior.py | 9 ++- .../test_host_prompt_behavior.py | 56 +++++++++++++++++++ 5 files changed, 82 insertions(+), 8 deletions(-) diff --git a/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.md b/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.md index 335ae7f4f4..cd77c8c55a 100644 --- a/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.md +++ b/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.md @@ -155,6 +155,17 @@ complete contract. They do **not** establish a current authority-loss bug: the recommendation is diagnostic, the unknown field is synthetic, and the complete source commitment still distinguishes both inputs. No model benefit is measured. +The release-only host-prompt decision probe composes its realistic external-wait +case through the existing complete-source TypeScript owner. Thin and brief modes +receive one checked source instance per probe batch, including complete task and +stop text; the scenario validator reads the source portfolio directly instead +of constructing an old compact packet. The four small independent decision +controls remain synthetic. The frozen raw/v0 model corpus and supported history +readers are unchanged. This migrates qualification input, not the ordinary CLI +writer or installed Host: provider execution remains explicitly release opt-in, +and scripted probe passes do not qualify model behavior, settlement, accepted +outcomes or complete cost. + ### Differential context Reuse the existing context-delivery owner and immutable capture mechanism. diff --git a/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.zh-CN.md b/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.zh-CN.md index efc1e6ee5d..2e73976144 100644 --- a/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.zh-CN.md +++ b/docs/architecture/rfcs/controller-executor-unified-contract-v0/contract-and-context.zh-CN.md @@ -117,6 +117,13 @@ settlement plan 在投影中保留,改变其 Turn 身份会改变 action signa recommendation 属于诊断,未知字段是合成输入,完整来源承诺仍能区分两份输入。 没有测量模型收益。 +发布用宿主提示决策探针的真实 external-wait 场景,经既有完整来源 TypeScript owner +组合。Thin 与 brief 在每批探针中接收同一份已检查来源,保留完整任务和停止正文; +场景校验器直接读取来源 portfolio,不再构造旧压缩包。四个独立的小型决策对照仍为 +合成输入,冻结的 raw/v0 模型语料和受支持历史 reader 不变。这只迁移资格验证输入, +不切换普通 CLI writer,也不证明已安装 Host 采用;provider 执行仍须显式 release +opt-in,脚本探针通过不代表模型行为、结算、验收成果或完整成本已经取得资格。 + ### 差分上下文 复用现有 context-delivery owner 与不可变 capture 机制。目标包含两个正交层, diff --git a/loopx/control_plane/testing/action_portfolio_scenarios.py b/loopx/control_plane/testing/action_portfolio_scenarios.py index c178b2205f..6910e3e555 100644 --- a/loopx/control_plane/testing/action_portfolio_scenarios.py +++ b/loopx/control_plane/testing/action_portfolio_scenarios.py @@ -3,7 +3,6 @@ from collections.abc import Mapping from typing import Any -from ..quota.cli_projection import compact_quota_should_run_cli_payload from ..quota.effective_action import EffectiveAction from ..quota.should_run import build_quota_should_run from ..work_items.interaction_contract import build_interaction_contract @@ -369,9 +368,7 @@ def validate_external_wait_fallback_scenario( resume_blocked = list(summary.get("resume_blocked_items") or []) first_resume_blocked = dict(resume_blocked[0]) if resume_blocked else {} condition = dict(first_resume_blocked.get("resume_condition") or {}) - compact = compact_quota_should_run_cli_payload(dict(source_packet)) - compact_portfolio = dict(compact.get("action_portfolio") or {}) - suggested = list(compact_portfolio.get("suggested_actions") or []) + suggested = list(portfolio.get("suggested_actions") or []) first_suggested = dict(suggested[0]) if suggested else {} if not ( selected.get("todo_id") == "todo_external_wait_fallback" @@ -391,7 +388,7 @@ def validate_external_wait_fallback_scenario( ): raise ValueError( "external-wait scenario must expose the pending P0 condition while " - "executing the bounded fallback from the compact default packet" + "executing the bounded fallback from the complete quota source" ) diff --git a/loopx/control_plane/testing/host_prompt_behavior.py b/loopx/control_plane/testing/host_prompt_behavior.py index f0fea8e121..64af2785fa 100644 --- a/loopx/control_plane/testing/host_prompt_behavior.py +++ b/loopx/control_plane/testing/host_prompt_behavior.py @@ -6,8 +6,8 @@ from pathlib import Path from ...heartbeat_prompt import build_heartbeat_prompt -from ..quota.cli_projection import compact_quota_should_run_cli_payload from ..quota.effective_action import EffectiveAction +from ..quota.turn_envelope import compose_turn_decision from .action_portfolio_scenarios import external_wait_fallback_scenario_source from .model_tool_behavior import DoubaoExecToolClient @@ -47,7 +47,9 @@ def cases() -> list[dict]: } for name, work, notify, replan, action in rows] probes.append({ "id": "wait_fallback_work", - "packet": compact_quota_should_run_cli_payload(external_wait_fallback_scenario_source()), + "packet": compose_turn_decision( + external_wait_fallback_scenario_source(), composed_settlement=True, + ), "expected": {"action": "work", "notify": True, "finish_goal": False}, }) return probes @@ -83,9 +85,10 @@ def probe_messages(mode: str, packet: dict) -> list[dict]: def run_probe(client: DoubaoExecToolClient, *, repeats: int = 2) -> dict: if not 1 <= repeats <= 5: raise ValueError("repeats must be between 1 and 5") + probes = cases() results = [] for mode in ("thin", "brief"): - for case in cases(): + for case in probes: messages = probe_messages(mode, case["packet"]) for attempt in range(repeats): # Each repetition is independent; failed attempts are not diff --git a/tests/control_plane/test_host_prompt_behavior.py b/tests/control_plane/test_host_prompt_behavior.py index 19c8c2cd79..d50e55f8b5 100644 --- a/tests/control_plane/test_host_prompt_behavior.py +++ b/tests/control_plane/test_host_prompt_behavior.py @@ -5,6 +5,11 @@ import pytest +from loopx.control_plane.effect_runtime import EffectRuntimeRejected +from loopx.control_plane.quota.turn_envelope import assert_complete_turn_decision +from loopx.control_plane.testing.action_portfolio_scenarios import ( + external_wait_fallback_scenario_source, +) from loopx.control_plane.testing.host_prompt_behavior import cases, probe_messages, run_probe @@ -78,6 +83,57 @@ def test_oracle_boundary_allows_version_preconditions_but_rejects_answer_leakage ScriptedClient().next_final_content(probe_messages("thin", cases()[0])) +def test_fallback_probe_delivers_one_complete_source_to_both_prompt_modes(monkeypatch): + source = external_wait_fallback_scenario_source() + stop = "保留完整材料和原权限。" * 60 + "\nStop when the original authority is revoked." + task = source["selected_todo"]["text"] + "\n" + "不得遗漏验收要求。" * 60 + source["goal_boundary"]["stop_condition"] = stop + source["selected_todo"]["text"] = task + monkeypatch.setattr( + "loopx.control_plane.testing.host_prompt_behavior.external_wait_fallback_scenario_source", + lambda: source, + ) + + class CompleteSourceClient(ScriptedClient): + def __init__(self): + super().__init__() + self.fallbacks = [] + + def next_final_content(self, messages): + packet = sent_packet(messages) + if packet.get("selected_todo", {}).get("todo_id") == "todo_external_wait_fallback": + assert_complete_turn_decision(packet) + assert packet["goal_boundary"]["stop_condition"] == stop + assert packet["selected_todo"]["text"] == task + self.fallbacks.append(packet) + return super().next_final_content(messages) + + client = CompleteSourceClient() + report = run_probe(client, repeats=1) + assert report["qualification_passed"] + assert client.count == report["provider_call_count"] == 10 + assert len(client.fallbacks) == 2 + assert client.fallbacks[0] == client.fallbacks[1] + assert source["goal_boundary"]["stop_condition"] == stop + assert source["selected_todo"]["text"] == task + + +@pytest.mark.parametrize("reserved", [ + {"schema_version": "unknown_source_version"}, {"contract_binding": {}}, +]) +def test_unknown_source_binding_is_rejected_before_provider_call(monkeypatch, reserved): + source = external_wait_fallback_scenario_source() + source.update(reserved) + monkeypatch.setattr( + "loopx.control_plane.testing.host_prompt_behavior.external_wait_fallback_scenario_source", + lambda: source, + ) + client = ScriptedClient() + with pytest.raises(EffectRuntimeRejected, match="reserved version/binding"): + run_probe(client, repeats=1) + assert client.count == 0 + + @pytest.mark.parametrize("mutation", [ {"action": "wait"}, {"notify": True}, {"finish_goal": True}, {"notify": 0}, ]) From f9cb71d1b8a6053d25931a0f3f7195537b8b6ff7 Mon Sep 17 00:00:00 2001 From: LoopX Agent <337587101+loopx-agent@users.noreply.github.com> Date: Wed, 18 Nov 2026 11:33:48 +0800 Subject: [PATCH 2/2] docs: align release probe guidance with complete source input Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com> --- docs/development/testing-and-quality.md | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-) diff --git a/docs/development/testing-and-quality.md b/docs/development/testing-and-quality.md index 2439ec2b9a..f1181fe903 100644 --- a/docs/development/testing-and-quality.md +++ b/docs/development/testing-and-quality.md @@ -2301,8 +2301,9 @@ explicit release qualification. It defaults to no calls; missing credentials report `skipped`, not a live pass. With securely injected `ARK_API_KEY`, it uses Doubao evolving for two independent repetitions of quiet-work, notifying-wait, quiet-wait, required-vision-replan and typed external-wait fallback cases in -each mode. The fallback case uses the real compact quota projection: its wait -transition already exists, so the host must advance the selected independent +each mode. The fallback case composes the complete quota source through the +shared TypeScript owner and reuses that checked instance in both modes. Its +wait transition already exists, so the host must advance the selected independent successor and notify rather than authoring another transition. Expected decisions remain outside model input. All attempts must pass; no answer correction or retry-until-pass is used. Ordinary pytest only checks the probe @@ -2313,8 +2314,9 @@ as separate evidence. Only hashes and pass/fail receipts are emitted, not raw prompts/responses. Model transport failures fail qualification rather than becoming environment skips. -仅发布前显式执行,普通 CI 不调用模型。检查静默不等于空转、等待不能擅自执行、 -vision replan 未关闭时不能提前结束 Goal; +回退场景使用完整 C1 来源,thin 与 brief 每批共享同一份已检查契约;原四个小型 +决策对照仍为合成输入。仅发布前显式执行,普通 CI 不调用模型。检查静默不等于空转、等待不能擅自执行、 +vision replan 未关闭时不能提前结束 Goal;这不是完整 Claude/Codex 行为验收的替代。 ```bash # No provider call by default. Explicit release opt-in uses ARK_API_KEY from the environment.