Keep Runtime Tool deadlines outside provider budgets

backend/app/services/agent_runtime/node_executor.py

@@ -778,6 +778,8 @@ async def _model(
                 repair_limit = (
                     WRITE_FILE_PROTOCOL_REPAIR_LIMIT
                     if is_write_file_repair
+                    else 10
+                    if repair_code == "invalid_tool_call"
                     else 1
                 )
                 repair_counter_key = (

backend/app/services/agent_runtime/tool_contracts.py

@@ -66,7 +66,13 @@ def __post_init__(self) -> None:
         "runtime_default", 60.0, 300.0, "stop_waiting_only"
     ),
     "network_read": ToolDeadlinePolicy(
-        "network_read", 30.0, 60.0, "stop_waiting_only"
+        "network_read", 60.0, 60.0, "stop_waiting_only"
+    ),
+    "image_generation": ToolDeadlinePolicy(
+        "image_generation", 120.0, 120.0, "stop_waiting_only"
+    ),
+    "custom_image_generation": ToolDeadlinePolicy(
+        "custom_image_generation", 600.0, 600.0, "stop_waiting_only"
     ),
     "local_code": ToolDeadlinePolicy(
         "local_code", 30.0, 3600.0, "cooperative"
@@ -93,6 +99,14 @@ def deadline_policy_for_tool(tool_name: str) -> ToolDeadlinePolicy:
         return _DEADLINE_POLICIES["agentbay_read"]
     if tool_name in {"read_emails", "read_webpage", "jina_read"}:
         return _DEADLINE_POLICIES["network_read"]
+    if tool_name == "generate_image_custom":
+        return _DEADLINE_POLICIES["custom_image_generation"]
+    if tool_name in {
+        "generate_image_siliconflow",
+        "generate_image_openai",
+        "generate_image_google",
+    }:
+        return _DEADLINE_POLICIES["image_generation"]
     return _DEADLINE_POLICIES["runtime_default"]
 
 

backend/app/services/agent_runtime/tool_execution.py

@@ -44,7 +44,7 @@
     "reconcile",
 ]
 ToolSideEffectState = Literal["none", "confirmed", "possible", "unknown"]
-SAFE_READ_MAX_ATTEMPTS = 3
+SAFE_READ_MAX_ATTEMPTS = 10
 
 # These tools dispatch an external image-generation request and can therefore
 # leave the provider outcome uncertain after a response timeout.  Direct Chat

backend/app/services/agent_runtime/tool_repair_budget.py

@@ -10,7 +10,7 @@
 from app.services.agent_runtime.state import JsonObject
 
 SAME_FINGERPRINT_FAILURE_LIMIT = 10
-TOOL_EPISODE_FAILURE_LIMIT = 20
+TOOL_EPISODE_FAILURE_LIMIT = 10
 _REPAIRABLE_MODEL_ACTIONS = frozenset(
     {"repair_arguments", "choose_other_tool"}
 )

backend/app/services/builtin_tool_definitions.py

@@ -3837,7 +3837,7 @@
     "generate_image_siliconflow": 120,
     "generate_image_openai": 120,
     "generate_image_google": 120,
-    "generate_image_custom": 120,
+    "generate_image_custom": 600,
 }
 
 

backend/app/services/llm/caller.py

@@ -66,7 +66,7 @@ async def execute_tool(*args, **kwargs):
     "send_message_to_agent", "send_feishu_message", "send_email"
 })
 
-WRITE_FILE_PROTOCOL_REPAIR_LIMIT = 3
+WRITE_FILE_PROTOCOL_REPAIR_LIMIT = 10
 WRITE_FILE_PROTOCOL_REPAIR_COUNTER_KEY = "invalid_tool_call:write_file"
 WRITE_FILE_PROTOCOL_REPAIR_INSTRUCTION = (
     "Your previous `write_file` call was not executed because `function.arguments` "
@@ -788,7 +788,7 @@ async def _buffer_chunk(_text: str) -> None:
             repair_limit = (
                 WRITE_FILE_PROTOCOL_REPAIR_LIMIT
                 if retry_tool_name == "write_file"
-                else 1
+                else 10
             )
             repair_counter_key = (
                 WRITE_FILE_PROTOCOL_REPAIR_COUNTER_KEY

backend/tests/test_agent_runtime_model_step_service.py

@@ -598,7 +598,7 @@ async def complete(model_arg, _messages, **_kwargs):
 
 
 @pytest.mark.asyncio
-async def test_invalid_write_file_arguments_request_three_protocol_repairs() -> None:
+async def test_invalid_write_file_arguments_request_ten_protocol_repairs() -> None:
     tenant_id = uuid.uuid4()
     model = _model(tenant_id)
     agent = _agent(tenant_id)

backend/tests/test_agent_runtime_node_executor.py

@@ -1351,15 +1351,16 @@ async def test_empty_output_is_repaired_once_then_fails_explicitly() -> None:
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
-    ("repair_code", "instruction"),
+    ("repair_code", "instruction", "repair_limit"),
     [
-        ("invalid_finish", "Retry finish with valid content."),
-        ("invalid_tool_call", "Retry with valid JSON tool arguments."),
+        ("invalid_finish", "Retry finish with valid content.", 1),
+        ("invalid_tool_call", "Retry with valid JSON tool arguments.", 10),
     ],
 )
 async def test_repeated_model_tool_protocol_repair_code_fails_explicitly(
     repair_code: str,
     instruction: str,
+    repair_limit: int,
 ) -> None:
     run_id = uuid.uuid4()
     repair = ModelStepResult(
@@ -1368,7 +1369,7 @@ async def test_repeated_model_tool_protocol_repair_code_fails_explicitly(
         repair_instruction=instruction,
         repair_code=repair_code,
     )
-    model = ModelService(repair, repair)
+    model = ModelService(*([repair] * (repair_limit + 1)))
     executor = _executor(model)
 
     result = await _invoke(run_id, executor, model_turn_limit=50)
@@ -1377,13 +1378,13 @@ async def test_repeated_model_tool_protocol_repair_code_fails_explicitly(
     assert lifecycle["status"] == "failed"
     assert lifecycle["reason"] == "model_tool_protocol_violation"
     assert lifecycle["error"]["code"] == "model_tool_protocol_violation"
-    assert lifecycle["model_protocol_repairs"] == {repair_code: 1}
-    assert lifecycle["model_step_count"] == 2
-    assert model.calls == 2
+    assert lifecycle["model_protocol_repairs"] == {repair_code: repair_limit}
+    assert lifecycle["model_step_count"] == repair_limit + 1
+    assert model.calls == repair_limit + 1
 
 
 @pytest.mark.asyncio
-async def test_write_file_protocol_repair_uses_three_attempts_then_guides_user() -> None:
+async def test_write_file_protocol_repair_uses_ten_attempts_then_guides_user() -> None:
     run_id = uuid.uuid4()
     repair = ModelStepResult(
         intent="text",
@@ -1392,7 +1393,7 @@ async def test_write_file_protocol_repair_uses_three_attempts_then_guides_user()
         repair_code="invalid_tool_call",
         repair_tool_name="write_file",
     )
-    model = ModelService(repair, repair, repair, repair)
+    model = ModelService(*([repair] * 11))
     executor = _executor(model)
 
     result = await _invoke(run_id, executor, model_turn_limit=50)
@@ -1408,14 +1409,14 @@ async def test_write_file_protocol_repair_uses_three_attempts_then_guides_user()
         ),
     }
     assert lifecycle["model_protocol_repairs"] == {
-        "invalid_tool_call:write_file": 3,
+        "invalid_tool_call:write_file": 10,
     }
-    assert lifecycle["model_step_count"] == 4
-    assert model.calls == 4
+    assert lifecycle["model_step_count"] == 11
+    assert model.calls == 11
 
 
 @pytest.mark.asyncio
-async def test_write_file_protocol_can_recover_on_the_third_repair() -> None:
+async def test_write_file_protocol_can_recover_on_the_tenth_repair() -> None:
     run_id = uuid.uuid4()
     repair = ModelStepResult(
         intent="text",
@@ -1424,9 +1425,7 @@ async def test_write_file_protocol_can_recover_on_the_third_repair() -> None:
         repair_tool_name="write_file",
     )
     model = ModelService(
-        repair,
-        repair,
-        repair,
+        *([repair] * 10),
         ModelStepResult(intent="finish", finish_content="Recovered"),
     )
     executor = _executor(model)
@@ -1435,9 +1434,9 @@ async def test_write_file_protocol_can_recover_on_the_third_repair() -> None:
 
     assert result["lifecycle"]["status"] == "completed"
     assert result["lifecycle"]["model_protocol_repairs"] == {
-        "invalid_tool_call:write_file": 3,
+        "invalid_tool_call:write_file": 10,
     }
-    assert model.calls == 4
+    assert model.calls == 11
 
 
 @pytest.mark.asyncio

backend/tests/test_agent_runtime_tool_contracts.py

@@ -8,9 +8,34 @@
     ToolContractError,
     ToolExecutionBinding,
     ToolWorksetEntry,
+    deadline_policy_for_tool,
+    resolve_tool_deadline_seconds,
     parse_step_tool_context,
     workset_version,
 )
+from app.services.builtin_tool_definitions import BUILTIN_TOOL_DEFINITIONS
+
+
+def test_runtime_deadlines_cover_declared_network_and_image_provider_budgets() -> None:
+    expected = {
+        "read_webpage": 60.0,
+        "jina_read": 60.0,
+        "generate_image_siliconflow": 120.0,
+        "generate_image_openai": 120.0,
+        "generate_image_google": 120.0,
+        "generate_image_custom": 600.0,
+    }
+
+    assert {
+        name: resolve_tool_deadline_seconds(deadline_policy_for_tool(name).name)
+        for name in expected
+    } == expected
+    declared = {
+        item["name"]: float(item["timeout_seconds"])
+        for item in BUILTIN_TOOL_DEFINITIONS
+        if item["name"] in expected
+    }
+    assert declared == expected
 
 
 def _entry() -> ToolWorksetEntry:

backend/tests/test_agent_runtime_tool_repair_budget.py

@@ -50,7 +50,7 @@ def test_tenth_consecutive_fingerprint_pauses_without_off_by_one() -> None:
     assert _episode(state)["total_failures"] == 10
 
 
-def test_twentieth_tool_failure_pauses_even_when_fingerprint_changes() -> None:
+def test_tenth_tool_failure_pauses_even_when_fingerprint_changes() -> None:
     state: dict = {}
     transition = None
     for model_step in range(1, TOOL_EPISODE_FAILURE_LIMIT + 1):
@@ -63,7 +63,7 @@ def test_twentieth_tool_failure_pauses_even_when_fingerprint_changes() -> None:
 
     assert transition is not None
     assert transition.pause_reason == "tool_repair_episode_limit_reached"
-    assert _episode(state)["total_failures"] == 20
+    assert _episode(state)["total_failures"] == 10
     assert _episode(state)["same_fingerprint_failures"] == 1
 
 

backend/tests/test_agent_runtime_tool_step_service.py

@@ -2977,7 +2977,7 @@ async def test_retryable_read_exhaustion_returns_one_non_retryable_result(
         "call-read-exhausted",
         "read_file",
     )
-    execution.attempt_count = 3
+    execution.attempt_count = 10
 
     async def reserve(db, **kwargs):
         del db
@@ -3017,7 +3017,7 @@ async def mark_failed(db, **kwargs):
     assert "Do not repeat the identical tool call unchanged" in result.messages[0][
         "content"
     ]
-    assert execution.result_metadata["runtime_attempt_count"] == 3
+    assert execution.result_metadata["runtime_attempt_count"] == 10
     assert execution.result_metadata["runtime_retry_exhausted"] is True
     assert execution.result_metadata["last_error_code"] == "temporary_read_failure"
 

backend/tests/test_finish_protocol.py

@@ -811,7 +811,7 @@ async def test_repeated_invalid_tool_json_is_bounded_by_protocol_code(monkeypatc
             }
         ],
     )
-    fake_client = FakeStreamClient([invalid, invalid])
+    fake_client = FakeStreamClient([invalid] * 11)
     monkeypatch.setattr(caller, "_get_agent_config", lambda _agent_id: _async_return((50, None)))
     monkeypatch.setattr(caller, "_get_user_name", lambda _user_id: _async_return("Ray"))
     monkeypatch.setattr(
@@ -841,12 +841,12 @@ async def test_repeated_invalid_tool_json_is_bounded_by_protocol_code(monkeypatc
     )
 
     assert result.startswith("[Error] invalid_tool_call_protocol_violation:")
-    assert len(fake_client.messages_seen) == 2
+    assert len(fake_client.messages_seen) == 11
     assert fake_client.closed is True
 
 
 @pytest.mark.asyncio
-async def test_invalid_write_file_json_gets_three_bounded_repairs(monkeypatch):
+async def test_invalid_write_file_json_gets_ten_bounded_repairs(monkeypatch):
     from app.services.llm import caller
     from app.services.llm.client import LLMResponse
 
@@ -863,7 +863,7 @@ async def test_invalid_write_file_json_gets_three_bounded_repairs(monkeypatch):
             }
         ],
     )
-    fake_client = FakeStreamClient([invalid, invalid, invalid, invalid])
+    fake_client = FakeStreamClient([invalid] * 11)
     monkeypatch.setattr(caller, "_get_agent_config", lambda _agent_id: _async_return((50, None)))
     monkeypatch.setattr(caller, "_get_user_name", lambda _user_id: _async_return("Ray"))
     monkeypatch.setattr(
@@ -897,7 +897,7 @@ async def test_invalid_write_file_json_gets_three_bounded_repairs(monkeypatch):
         "本次文件生成未完成：write_file 工具参数无效或被截断，连续重试后仍无法执行。"
         "请回复「重新生成」，我会基于当前对话重新尝试。"
     )
-    assert len(fake_client.messages_seen) == 4
+    assert len(fake_client.messages_seen) == 11
     assert fake_client.closed is True
 
 

backend/tests/test_tool_execution.py

@@ -881,7 +881,10 @@ async def test_expired_final_safe_read_attempt_closes_without_provider_replay():
     assert reservation.prior_failure is not None
     assert reservation.prior_failure.error_code == "tool_retry_exhausted"
     assert execution.status == "failed"
-    assert execution.result_metadata["runtime_attempt_count"] == 3
+    assert (
+        execution.result_metadata["runtime_attempt_count"]
+        == tool_execution.SAFE_READ_MAX_ATTEMPTS
+    )
     assert execution.result_metadata["runtime_retry_exhausted"] is True
     assert db.flush_count == 1
 

specs/002-tool-runtime-contract/checklists/requirements.md

@@ -33,4 +33,4 @@
 
 - 第一次校验即通过，无 `[NEEDS CLARIFICATION]` 项。
 - `Tool Call`、`Run`、`Receipt`、`checkpoint` 等词是本产品领域对象，不是具体实现方案；具体数据结构、文件和迁移步骤将在 Plan 阶段定义。
-- Spec 已覆盖用户确认的 10/20 repair budget、模型可见错误反馈、unknown write 禁止自动重放和旧 checkpoint 兼容边界。
+- Spec 已覆盖用户确认的 Tool repair/retry 上限统一为 10、模型可见错误反馈、unknown write 禁止自动重放和旧 checkpoint 兼容边界；计数结构统一重构已明确延期。

specs/002-tool-runtime-contract/contracts/repair-and-lifecycle.md

@@ -3,7 +3,8 @@
 ## Tool Repair Episode
 
 - `same_fingerprint_failures` reaches 10: pause immediately after recording the 10th failure; do not invoke model step 11 for that loop.
-- `total_failures` reaches 20 for the same Tool episode: pause immediately; do not invoke the next model step.
+- `total_failures` reaches 10 for the same Tool episode: pause immediately; do not invoke the next model step.
+- Generic Tool protocol repair, `write_file` protocol repair, and safe-read replay retain their current independent counters but each uses a limit of 10; counter unification is deferred.
 - Changing fingerprint resets only the consecutive counter.
 - Success of the same Tool, new Run, or explicit user correction resets the Tool episode.
 - Success of another Tool does not reset it.

specs/002-tool-runtime-contract/plan.md

@@ -5,7 +5,7 @@
 
 ## Summary
 
-在现有 Durable Runtime、`AgentToolExecution` Receipt、safe-read replay 和 unknown/reconcile 机制之上，增加一次 Model Step 固化、checkpoint 可恢复的 `StepToolContext`。新 Tool Step 只使用已接受的 Tool Contract/Execution Binding，不再调用 ToolProvider 重建 Workset；同时把 Provider Call ID、Runtime Call Instance 和 Execution Receipt 分离，统一 schema validation、authorization/approval、模型可见失败反馈及 10/20 repair budget。操作 deadline、取消传播和 Receipt lease 继续保持三个独立控制面。长期通过可渐进迁移的 RegisteredTool 收敛模型定义与执行能力，不一次性替换现有 Handler。
+在现有 Durable Runtime、`AgentToolExecution` Receipt、safe-read replay 和 unknown/reconcile 机制之上，增加一次 Model Step 固化、checkpoint 可恢复的 `StepToolContext`。新 Tool Step 只使用已接受的 Tool Contract/Execution Binding，不再调用 ToolProvider 重建 Workset；同时把 Provider Call ID、Runtime Call Instance 和 Execution Receipt 分离，统一 schema validation、authorization/approval、模型可见失败反馈，并将现有独立 Tool repair/retry 上限统一为 10。操作 deadline、取消传播和 Receipt lease 继续保持三个独立控制面。长期通过可渐进迁移的 RegisteredTool 收敛模型定义与执行能力，不一次性替换现有 Handler。
 
 ## Technical Context
 
@@ -50,7 +50,7 @@
 ### Phase C — Repair budgets
 
 1. checkpoint 保存 per-tool repair episode、连续 fingerprint 计数和总计数。
-2. 第 10 次连续相同失败或第 20 次同 Tool episode 失败后暂停，且不发起下一次模型调用。
+2. 第 10 次连续相同失败或第 10 次同 Tool episode 失败后暂停，且不发起下一次模型调用；普通 Tool JSON repair、`write_file` JSON repair 和 safe-read replay 也只把现有独立上限改为 10，不在本轮重构计数结构。
 3. Tool 成功、新 Run、用户明确纠正按 contract 重置；Provider retry、safe internal replay、permission/confirmation、pending、cancel、unknown 不计数。
 4. Verifier repair 改为当前 issue episode 计数，保留全局 `model_turn_limit` 独立语义。
 
@@ -122,7 +122,7 @@ backend/
 2. Runtime integration tests：Model Step → checkpoint → 新 Worker Tool Step；普通 availability 变化不影响已接受 Call；安全状态变化仍阻断。
 3. Receipt tests：replay 复用同一 execution；lease renewal/loss/fence；unknown write no replay；safe read bounded retry。
 4. Compatibility tests：旧 checkpoint 单次 resolver、新 checkpoint 禁止 resolver、mixed-version nullable fields。
-5. Lifecycle tests：10/20 off-by-one、reset/exclusion、operation deadline、cancel propagation。
+5. Lifecycle tests：统一上限 10 的 off-by-one、reset/exclusion、operation deadline、cancel propagation。
 6. Static gates：scoped Ruff、pytest、Alembic single head + upgrade/downgrade、`scripts/arch-guard.sh`。
 
 ## Complexity Tracking

specs/002-tool-runtime-contract/quickstart.md

@@ -26,7 +26,7 @@ Expected branch: `002-tool-runtime-contract`; base contains `upstream/main@251ae
 3. Remove ToolProvider access from new-format Tool Step; add legacy batch resolver.
 4. Add DB columns/migration and projection metadata.
 5. Add shared validation/authorization/failure envelope.
-6. Add repair episode state and 10/20 gates.
+6. Add repair episode state and uniform Tool repair/retry limit 10 gates.
 7. Harden operation deadlines/cancel/lease tests.
 8. Add RegisteredTool boundary and migrate representative tools only.
 
@@ -66,7 +66,7 @@ cd backend
 - checkpoint restart on another Worker uses the same binding and execution row;
 - repeated Provider-local ID in another Assistant Turn does not collide;
 - schema failure returns exactly one sanitized Tool Result;
-- failure 10 and 20 pause before the next model invocation;
+- the 10th repair failure pauses before the next model invocation;
 - provider retry, safe replay, pending, cancel and unknown do not increment repair budget;
 - lease loss blocks stale settlement; uncertain write is never auto-replayed;
 - legacy checkpoint resolves once per pending batch, new checkpoint never uses legacy fallback.

specs/002-tool-runtime-contract/research.md

@@ -56,7 +56,7 @@
 
 ### D7. Repair budget 是 Tool episode，不是 Provider/Receipt retry
 
-**Decision**: 连续同 fingerprint 第 10 次、同 Tool episode 第 20 次暂停；只计模型可见、可修复失败。
+**Decision**: 连续同 fingerprint 第 10 次、同 Tool episode 第 10 次暂停；只计模型可见、可修复失败。普通 Tool protocol repair、`write_file` protocol repair 和 safe-read replay 继续使用各自现有计数入口，但上限统一为 10，状态结构后续再整体重构。
 
 **Rationale**: Provider transport retry 和 Receipt safe replay 都不代表模型做了错误决策；混计会过早停机或掩盖循环。