From 8d91d472d5dc5ef9860ae895a1ac821fbf116057 Mon Sep 17 00:00:00 2001
From: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
Date: Thu, 29 Oct 2026 19:33:22 +0800
Subject: [PATCH] test(prompts): align consumer checks with retained input
 contracts

Signed-off-by: LoopX Agent <337587101+loopx-agent@users.noreply.github.com>
---
 docs/development/testing-and-quality.md       | 18 ++++++++
 .../test_heartbeat_brief_budget.py            | 42 +++++++++----------
 tests/test_chat_turn_steering.py              |  2 +-
 tests/test_host_loop_activation.py            |  9 ++--
 4 files changed, 45 insertions(+), 26 deletions(-)

diff --git a/docs/development/testing-and-quality.md b/docs/development/testing-and-quality.md
index b4ffef0eab..615e1c5277 100644
--- a/docs/development/testing-and-quality.md
+++ b/docs/development/testing-and-quality.md
@@ -1313,6 +1313,24 @@ fields; this does not introduce a new telemetry system or receipt gate.
 
 ### Budget Failure Decisions
 
+The fixed visible-Goal host fixture's 3,200-character regression guard predates
+`1d5d2d0f6`, which adds the known Todo choice to the first current guard. On the
+same four host inputs at `b201da8ba`, substituting only the pre-change dispatch
+rule gives 3,199 / 3,177 / 2,966 / 2,895 normalized body characters; the complete
+current rule gives 3,262 / 3,240 / 3,029 / 2,958. The 63-character increase teaches
+an already-known choice before admission, while retaining selection/reentry,
+ordered settlement, exact identities, actual outcomes and ambiguous-effect
+recovery. Preserve the original 3,200-ceiling failure; calibrate only this test's
+regression allowance to 3,300, leaving 38 characters above the largest fixed
+input. Production prompt bytes, the 4,000-character host limit, longer-route
+checks, execution authority and frozen qualification results remain unchanged.
+This is no output reduction or model-compliance qualification.
+
+相同四个宿主输入中，首次 guard 的已知 Todo 选择指引增加 63 个归一化字符，
+最大值从 3,199 变为 3,262。保留原 3,200 超限失败，仅将该固定测试的回归预算
+校准为 3,300，留 38 字符余量；不删指令，不改生产正文、4,000 宿主硬边界、
+长路径检查、执行权限或冻结资格结果，也不据此声称输出节省或模型遵从。
+
 The actual-default model portfolio retains the complete capability candidate
 context required by the shared diagnostic projection. On the unchanged three
 noisy fixtures, `74b3d526ab1c1fd3d5f96308bef2482fcad7e636` emitted 4,944 /
diff --git a/tests/control_plane/test_heartbeat_brief_budget.py b/tests/control_plane/test_heartbeat_brief_budget.py
index f750e11e22..c2814f8183 100644
--- a/tests/control_plane/test_heartbeat_brief_budget.py
+++ b/tests/control_plane/test_heartbeat_brief_budget.py
@@ -2,7 +2,6 @@
 from __future__ import annotations
 
 import contextlib
-import hashlib
 import io
 import json
 import shlex
@@ -93,26 +92,27 @@ def test_brief_keeps_complete_routes_with_128_character_windows_root() -> None:
     assert json.loads(encoded) == payload
     root, _, _, _ = _windows_routes()
     assert encoded.count(json.dumps(root)[1:-1]) == 20
-    # The unchanged fixture on main b5a1a816a measures 10,522 JSON chars and
-    # 8,579 Markdown. The complete outcome paragraph adds 523 body/Markdown
-    # chars; omitting absent brief metadata saves 35 JSON chars: 11,010/9,102.
-    # The later diagnostic-only renderer change retains that full body and JSON,
-    # prints an identical active/resolved path once and omits absent pre-quota
-    # metadata: Markdown is 8,867 chars. The expected output was derived from
-    # b16aeb3's frozen 9,102-char rendering using only those two substitutions.
-    # These hashes retain every prior clause and route, plus the reviewed
-    # instruction transition. This is a synthetic spelling, not a Windows run.
-    assert hashlib.sha256(payload["task_body"].encode()).hexdigest() == (
-        "cfc65091d2710ef298174e6a5280edf55f6cc7acdce8e5688a2ae213b5f4b4ac"
-    )
-    assert hashlib.sha256(markdown.encode()).hexdigest() == (
-        "e5c10d37ed7cef08e8fa56a450f3c119de92d1ef99e3516ff6b106a8853c8186"
-    )
-    # fd1b23b59 raised the existing body allowance, preserving every clause.
-    # Only that diagnostic number differs from the earlier frozen rendering.
-    assert hashlib.sha256(markdown.replace("max_chars=`5400`", "max_chars=`4800`").encode()).hexdigest() == (
-        "29604e41d97963976ab8911e69c6328421aeaf63053cdb12c197dbe9b5291b9b"
-    )
+    # Assert the decision obligations rather than a hash of dated prose.
+    # The full body must reach Markdown, with its executable guard only once.
+    body = payload["task_body"]
+    assert body in markdown
+    assert markdown.count(payload["quota_guard_command"]) == 1
+    for obligation in (
+        "true=work+writeback; only false permits no-op",
+        "Read agent_channel.work_context and required_reads (or envelope) before work",
+        "retain original and cumulative outcome requirements",
+        "reuse inline goal_acceptance, inspect only if absent, then refine",
+        "Known new-Turn choice: add `--todo-id <id>` to first guard; otherwise",
+        "use selection_command when required; re-enter as instructed, do/verify authorized work",
+        "Settle with current settlement_plan.ordered_steps, else next_cli_actions",
+        "preserve exact identities/effect ids and actual outcomes",
+        "Never reconstruct static refresh/spend; ambiguous effects need readback/recovery, not blind retry",
+        "Todo done≠Turn settled; no invented successor/--no-follow-up",
+        "No wait spend",
+        "Finish Goal only on terminal no-follow-up",
+        "Destructive Git/production requires explicit authorization",
+    ):
+        assert obligation in body, obligation
     spec = CLI_OUTPUT_MODE_VARIANT_BY_ID["heartbeat_prompt_brief"]
     assert spec.max_chars == {"json": 11_400, "markdown": 9_600}
     for output_format, text in (("json", encoded), ("markdown", markdown)):
diff --git a/tests/test_chat_turn_steering.py b/tests/test_chat_turn_steering.py
index e6f8ba11be..0d9e7b69d1 100644
--- a/tests/test_chat_turn_steering.py
+++ b/tests/test_chat_turn_steering.py
@@ -397,7 +397,7 @@ def record_write(self, payload):
     initial_text = starts[0]["input"][0]["text"]
     assert "LoopX context (supporting context only):" in initial_text
     assert "Inspect a bounded task." in initial_text
-    assert initial_text.endswith("Current user message: wait for steer")
+    assert initial_text.endswith("Current user message:\nwait for steer")
     updates = [row["params"] for row in writes if row.get("method") == "turn/steer"]
     assert updates == [{
         "threadId": session.thread_id,
diff --git a/tests/test_host_loop_activation.py b/tests/test_host_loop_activation.py
index e28f7a80d0..5ee6ddd239 100644
--- a/tests/test_host_loop_activation.py
+++ b/tests/test_host_loop_activation.py
@@ -450,10 +450,11 @@ def test_goal_hosts_reuse_thin_dispatch_and_stay_compact() -> None:
             "preserve exact identities/effect ids",
         ):
             assert obligation in payload["task_body"]
-        # Current authority/continuation clauses fit 3,200 normalized
-        # chars on this fixed fixture. Keep a tight regression guard and the
-        # native host's separate hard cap, without shortening those obligations.
-        assert payload["interface_budget"]["budget_char_count"] <= 3_200
+        # The known first-guard choice adds 63 useful normalized characters:
+        # the largest fixed host fixture is 3,262, leaving 38 under this
+        # regression allowance. Keep the native host's hard cap separate.
+        assert "Known new-Turn choice: add `--todo-id <id>` to first guard" in payload["task_body"]
+        assert payload["interface_budget"]["budget_char_count"] <= 3_300
         assert payload["interface_budget"]["max_chars"] == 4_000
         assert payload["interface_budget"]["within_budget"] is True