fix: request JSON decisions and retain sealed acceptance history

docs/DEVELOPMENT.md

@@ -69,6 +69,8 @@ uv run pytest -q
 
 真实 UltiCode HTTP / 模型 e2e 均为显式 opt-in。`e2e_sourced_analysis.py` uses agent-authored synthetic Markdown—not submissions, DTOs, or licensed user material—and validates a read-only submission projection without a real model. U03 workflow model analysis additionally requires a valid U02 gate and active budget authorization; other evaluation scripts follow their own gates. Supply credentials through a secure environment/secret store, never command text or logs. Runner contracts live in source and Linear; keep per-run results out of core docs.
 
+`DeepseekModel` requests provider JSON Output with `response_format={"type":"json_object"}` and an explicit JSON prompt. The guard accepts only that exact format when supplied; legacy requests may omit the field. Empty, truncated, malformed or schema-invalid decisions still fail strict validation after known usage is settled. JSON Output does not replace authorization, tool argument validation or citation verification. See [DeepSeek JSON Output](https://api-docs.deepseek.com/guides/json_mode/).
+
 只读工具模型遇到混合请求时拒绝越权部分，继续执行独立且已授权的部分；用户已明确要求的合法只读操作应直接调用工具，不再次征求确认或只提出执行建议。提交分析缺少明确 ID 或可靠会话选择时，不调用提交选择工具或用最近提交列表代替澄清；混合请求先执行独立授权的只读查询，再在最终答案中询问缺失的提交 ID。工具仍绑定当前服务端会话，不能因请求要求切换身份。隔离验收同时要求没有泄露和本人数据的正向工具对照，不能以整段拒绝冒充完整通过。
 
 授权周期的 `authorized_budget_period` 仍只保存生命周期元数据；其快照始终明确
@@ -494,6 +496,7 @@ The same explicitly authorized rollover contract applies to the later supported
 | `acceptance-revalidation-v12` | V11 |
 | `acceptance-revalidation-v13` | V12, including its retained unknown request |
 | `acceptance-revalidation-v14` | Settled V13 on the integrated U03 candidate |
+| `acceptance-revalidation-v15` | Settled V14, including its invalid-JSON failure |
 
 Before preparation, retain the predecessor ledger, binding and guard at their pinned
 fingerprints and preserve the entire earlier history chain. The selected policy's
@@ -526,6 +529,12 @@ commitments and inherited unknown liabilities remain cumulative. Failed protocol
 evidence stays in the predecessor, while the integrated candidate receives a fresh
 complete acceptance prefix without changing the existing purpose or token limits.
 
+V15 carries the permanently halted V14 ledger, binding and guard into the JSON-output
+candidate under the same cumulative USD200 authorization. Keep the invalid-JSON
+failure and all known costs in V14; do not resume it or reinterpret it as passing
+acceptance. The earlier three unknown liabilities and every purpose/token cap remain
+unchanged. A new complete prefix must establish acceptance for the changed adapter.
+
 Issue the evidence-bound U02 gate only after those inputs validate. Gate artifact references are
 relative to the private directory containing the gate; place the referenced artifacts there:
 

services/agent/data/repository_corpus_manifest.json

@@ -29,8 +29,8 @@
   },
   {
     "doc_id": "repository-development",
-    "version": "sha256-24c1cea137aab6834ef530d75fdf95984ebbe55bafc5322d6173a6c65e8a5a89",
-    "chunk_id": "repository-development:sha256-24c1cea137aab6834ef530d75fdf95984ebbe55bafc5322d6173a6c65e8a5a89:1",
+    "version": "sha256-f26b77d34a904e1fae301357a5569b657309e95ac1a84ec1e6b6a6d735f130c3",
+    "chunk_id": "repository-development:sha256-f26b77d34a904e1fae301357a5569b657309e95ac1a84ec1e6b6a6d735f130c3:1",
     "source_path": "docs/DEVELOPMENT.md",
     "access_scope": "repository-public",
     "sample_kind": "real",

services/agent/src/agent_service/gate.py

@@ -988,6 +988,8 @@ def _check_dav58_raw_judge(
     expected_body = {
         "model": "deepseek-flash",
         "messages": [{"role": "system", "content": system}, {"role": "user", "content": user_prompt}],
+        **({"response_format": {"type": "json_object"}}
+           if isinstance(request_body, dict) and "response_format" in request_body else {}),
         "temperature": 0, "max_tokens": 2000, "thinking": {"type": "disabled"},
     }
     body_sha = hashlib.sha256(json.dumps(

services/agent/src/authorized_budget_period.py

@@ -74,6 +74,7 @@
 REVALIDATION_V12_POLICY_ID = "acceptance-revalidation-v12"
 REVALIDATION_V13_POLICY_ID = "acceptance-revalidation-v13"
 REVALIDATION_V14_POLICY_ID = "acceptance-revalidation-v14"
+REVALIDATION_V15_POLICY_ID = "acceptance-revalidation-v15"
 REVALIDATION_POLICY_IDS = frozenset({REVALIDATION_POLICY_ID, REVALIDATION_V2_POLICY_ID, REVALIDATION_V3_POLICY_ID, REVALIDATION_V4_POLICY_ID, REVALIDATION_V5_POLICY_ID, REVALIDATION_V6_POLICY_ID, REVALIDATION_V7_POLICY_ID, REVALIDATION_V8_POLICY_ID, REVALIDATION_V9_POLICY_ID, REVALIDATION_V10_POLICY_ID, REVALIDATION_V11_POLICY_ID, REVALIDATION_V12_POLICY_ID})
 REVALIDATION_V2_HISTORY = MappingProxyType({
     "attempts": 164, "known_actual_micro_usd": 42_307,
@@ -227,6 +228,17 @@
 })
 REVALIDATION_POLICY_IDS = REVALIDATION_POLICY_IDS | {REVALIDATION_V14_POLICY_ID}
 
+REVALIDATION_V15_HISTORY = MappingProxyType({
+    "attempts": 1183, "known_actual_micro_usd": 330_069,
+    "known_committed_micro_usd": 10_440_914,
+    "unknown_attempts": 3, "unknown_encumbrance_micro_usd": 2_359_296,
+    "cumulative_attempt_limit": 1424, "cumulative_limit_micro_usd": 200_000_000,
+})
+REVALIDATION_V15_POLICY = MappingProxyType({
+    **REVALIDATION_V14_POLICY, "history": REVALIDATION_V15_HISTORY,
+})
+REVALIDATION_POLICY_IDS = REVALIDATION_POLICY_IDS | {REVALIDATION_V15_POLICY_ID}
+
 
 def policy_for(policy_id: str):
     if policy_id == POLICY_ID:
@@ -259,6 +271,8 @@ def policy_for(policy_id: str):
         return REVALIDATION_V13_POLICY
     if policy_id == REVALIDATION_V14_POLICY_ID:
         return REVALIDATION_V14_POLICY
+    if policy_id == REVALIDATION_V15_POLICY_ID:
+        return REVALIDATION_V15_POLICY
     raise PeriodError("unsupported policy")
 
 

services/agent/src/boundary_evaluation.py

@@ -584,6 +584,7 @@ def _request_body(
     body: dict[str, object] = {
         "model": getattr(model, "_model", None),
         "messages": api_messages(messages),
+        "response_format": {"type": "json_object"},
         "temperature": 0,
         "max_tokens": getattr(model, "_max_tokens", None),
     }

services/agent/src/dav58_live_guard.py

@@ -204,7 +204,8 @@ def begin(self, request: httpx.Request, lane: str, reservation=None):
                     raise ModelBudgetExceeded("guard and budget history disagree")
             body = json.loads(request.content)
             if (not isinstance(body, dict)
-                    or set(body) != {"model", "messages", "temperature", "max_tokens", "thinking"}
+                    or set(body) - {"response_format"} != {"model", "messages", "temperature", "max_tokens", "thinking"}
+                    or ("response_format" in body and body["response_format"] != {"type": "json_object"})
                     or body["model"] != "deepseek-flash"
                     or body["thinking"] != {"type": "disabled"}
                     or type(body["max_tokens"]) is not int

services/agent/src/deepseek_model.py

@@ -260,6 +260,7 @@ async def decide(self, messages: list[dict[str, object]]) -> ModelDecision:
         body: dict[str, object] = {
             "model": self._model,
             "messages": api_messages,
+            "response_format": {"type": "json_object"},
             "temperature": 0,
             "max_tokens": self._max_tokens,
         }

services/agent/src/revalidation_history.py

@@ -268,14 +268,30 @@ def _check_unknown_rollover(snapshot, guard, *, predecessor="v2"):
     "dav58-increment-10a5b715a8dc4a6b96056116a789abea.json": "a4f08105519da4c96c258625676b92d428028066da3e230c8a1e7bad2a8607a8",
 }
 
+V14_ROLLOVER_IDENTITY = {
+    "period_id": "revalidation-v14-20261011",
+    "identity": "38633ca751c44e94b0ae63d17e672bde",
+    "config_sha256": "a650011b552cde66b265fb030daaba824ee1f660d05f48326693ff4e8210d7c1",
+    "policy_id": "acceptance-revalidation-v14",
+}
+V14_ROLLOVER_SHA256 = {
+    "budget.sqlite3": "46bc407b736cae0ad7d9c271b028f6556c363eb7a5de5906699b7636c56a58c6",
+    "binding.json": "d8f7384635e3a7d2f103bda75cf39dd1027966fef34b6a6960cc93b7ee355bb2",
+    "dav58-increment-38633ca751c44e94b0ae63d17e672bde.json": "605aababbad617e0d25f2c26103e5ece9960f68536dd4374fc526d353fbada85",
+}
+
 
 def _validate_rollover_history(sources, *, settled_policy="v1"):
-    from authorized_budget_period import REVALIDATION_V14_HISTORY
+    from authorized_budget_period import REVALIDATION_V14_HISTORY, REVALIDATION_V15_HISTORY
     from authorized_budget_period import PeriodError, PeriodIdentity, REVALIDATION_V2_HISTORY, REVALIDATION_V4_HISTORY, REVALIDATION_V5_HISTORY, REVALIDATION_V6_HISTORY, REVALIDATION_V7_HISTORY, REVALIDATION_V8_HISTORY, REVALIDATION_V9_HISTORY, REVALIDATION_V10_HISTORY, REVALIDATION_V11_HISTORY, REVALIDATION_V12_HISTORY, _parent, _file
     from model_budget import ModelBudget, authorization_slot
     from dav58_live_guard import IncrementalGuard
 
-    if settled_policy == "v13":
+    if settled_policy == "v14":
+        original = _validate_rollover_history(sources, settled_policy="v13")
+        identity, fingerprints = V14_ROLLOVER_IDENTITY, V14_ROLLOVER_SHA256
+        totals, history = (97, 33_416, 872_400), REVALIDATION_V15_HISTORY
+    elif settled_policy == "v13":
         original = _validate_unknown_rollover_history(sources, predecessor="v12")
         identity, fingerprints = V13_ROLLOVER_IDENTITY, V13_ROLLOVER_SHA256
         totals, history = (98, 33_738, 877_200), REVALIDATION_V14_HISTORY
@@ -351,7 +367,7 @@ def _validate_rollover_history(sources, *, settled_policy="v1"):
     try:
         # Validate sealed receipts without making the permanently halted journal resumable.
         receipt_view = guard
-        if settled_policy == "v13":
+        if settled_policy in {"v13", "v14"}:
             if guard.get("halted") is not True:
                 raise PeriodError("previous guard must be permanently halted")
             receipt_view = {**guard, "halted": False}
@@ -364,12 +380,14 @@ def _validate_rollover_history(sources, *, settled_policy="v1"):
 
 
 def validate_history(sources, *, policy_id="acceptance-revalidation-v1"):
-    from authorized_budget_period import REVALIDATION_V13_POLICY_ID, REVALIDATION_V14_POLICY_ID
+    from authorized_budget_period import REVALIDATION_V13_POLICY_ID, REVALIDATION_V14_POLICY_ID, REVALIDATION_V15_POLICY_ID
     from authorized_budget_period import PeriodError, REVALIDATION_HISTORY, _parent, _file
     from authorized_budget_period import REVALIDATION_POLICY_ID, REVALIDATION_V2_POLICY_ID, REVALIDATION_V3_POLICY_ID, REVALIDATION_V4_POLICY_ID, REVALIDATION_V5_POLICY_ID, REVALIDATION_V6_POLICY_ID, REVALIDATION_V7_POLICY_ID, REVALIDATION_V8_POLICY_ID, REVALIDATION_V9_POLICY_ID, REVALIDATION_V10_POLICY_ID, REVALIDATION_V11_POLICY_ID, REVALIDATION_V12_POLICY_ID
 
     if policy_id == REVALIDATION_V14_POLICY_ID:
         return _validate_rollover_history(sources, settled_policy="v13")
+    if policy_id == REVALIDATION_V15_POLICY_ID:
+        return _validate_rollover_history(sources, settled_policy="v14")
     if policy_id == REVALIDATION_V13_POLICY_ID:
         return _validate_unknown_rollover_history(sources, predecessor="v12")
     if policy_id == REVALIDATION_V2_POLICY_ID:

services/agent/tests/test_dav58_live_guard.py

@@ -171,7 +171,9 @@ async def run():
 
 
 @pytest.mark.parametrize("changes", [{"model": "deepseek-v4-pro"}, {"thinking": {"type": "enabled"}},
-    {"max_tokens": 2001}, {"stream": True}, {"tools": []}])
+    {"max_tokens": 2001}, {"stream": True}, {"tools": []},
+    {"response_format": {"type": "text"}}, {"response_format": None},
+    {"response_format": {"type": "json_object", "extra": True}}])
 def test_unexpected_paid_payload_rejected_before_send(tmp_path, changes):
     guard = IncrementalGuard(tmp_path / "increment.json")
     body = {"model": "deepseek-flash", "messages": [{"role": "user", "content": "x"}], "max_tokens": 2000,
@@ -182,6 +184,19 @@ def test_unexpected_paid_payload_rejected_before_send(tmp_path, changes):
     assert guard.state["receipts"] == []
 
 
+def test_guard_accepts_provider_json_output(tmp_path):
+    guard = IncrementalGuard(tmp_path / "increment.json")
+    body = {"model": "deepseek-flash", "messages": [{"role": "system", "content": "Reply in JSON."}],
+            "max_tokens": 2000, "temperature": 0, "thinking": {"type": "disabled"},
+            "response_format": {"type": "json_object"}}
+    try:
+        entry = guard.begin(httpx.Request("POST", "https://api.deepseek.com/chat/completions", json=body),
+                            "dav58_loop")
+        assert entry["lane"] == "dav58_loop"
+    finally:
+        guard.close()
+
+
 def test_existing_journal_cannot_reset(tmp_path):
     path = tmp_path / "increment.json"
     IncrementalGuard(path)

services/agent/tests/test_deepseek_model.py

@@ -1,4 +1,5 @@
 import asyncio
+import json
 
 import deepseek_model
 import httpx
@@ -7,6 +8,25 @@
 from deepseek_model import DeepseekModel, ModelProtocolError, _parse_decision
 
 
+@pytest.mark.parametrize("tools,content", [
+    ({}, '{"answer":"ok"}'),
+    ({"get_problem": "args"}, '{"tool":"get_problem","args":{}}'),
+])
+def test_decision_requests_provider_json_output(tools, content):
+    async def handler(request):
+        payload = json.loads(request.content)
+        assert payload["response_format"] == {"type": "json_object"}
+        assert "json" in payload["messages"][0]["content"].lower()
+        return httpx.Response(200, json={"choices": [{"message": {"content": content}}]})
+
+    async def scenario():
+        async with DeepseekModel("test-key", tool_specs=tools,
+                                 transport=httpx.MockTransport(handler)) as model:
+            await model.decide([{"role": "user", "content": "question"}])
+
+    asyncio.run(scenario())
+
+
 @pytest.mark.parametrize(
     "response",
     [

services/agent/tests/test_v15_json_history.py

@@ -0,0 +1,18 @@
+from authorized_budget_period import policy_for
+
+
+def test_json_acceptance_retains_all_known_costs_and_unknown_liabilities():
+    old = policy_for("acceptance-revalidation-v14")
+    policy = policy_for("acceptance-revalidation-v15")
+    history = policy["history"]
+    assert history["attempts"] == old["history"]["attempts"] + 97 == 1183
+    assert history["known_actual_micro_usd"] == old["history"]["known_actual_micro_usd"] + 33416 == 330069
+    assert history["known_committed_micro_usd"] == old["history"]["known_committed_micro_usd"] + 872400 == 10440914
+    assert history["unknown_attempts"] == old["history"]["unknown_attempts"] == 3
+    assert history["unknown_encumbrance_micro_usd"] == old["history"]["unknown_encumbrance_micro_usd"] == 2359296
+    assert history["cumulative_limit_micro_usd"] == old["history"]["cumulative_limit_micro_usd"] == 200000000
+    assert history["cumulative_attempt_limit"] == history["attempts"] + policy["attempts"] == 1424
+    assert policy["lanes"] == old["lanes"]
+    assert policy["limit_micro_usd"] == old["limit_micro_usd"]
+    assert (history["known_committed_micro_usd"] + history["unknown_encumbrance_micro_usd"]
+            + policy["limit_micro_usd"] <= history["cumulative_limit_micro_usd"])