fix: recover Work runs after failed verification

agent_platform/runtime/app.py

@@ -22,6 +22,18 @@ from agent_platform.runtime.schemas import ChatCompletionRequest
 from agent_platform.runtime.tools import ToolRegistry
 from agent_platform.store import RuntimeStore
 
+OPENWEBUI_BACKGROUND_TASKS = {
+    "title_generation",
+    "follow_up_generation",
+    "tags_generation",
+    "emoji_generation",
+    "query_generation",
+    "image_prompt_generation",
+    "autocomplete_generation",
+    "function_calling",
+    "moa_response_generation",
+}
+
 
 def _sse_chunk(model: str, content: str = "", finish_reason: str | None = None) -> bytes:
     payload = {
@@ -145,15 +157,7 @@ def create_app(
         except ValueError as exc:
             raise HTTPException(status_code=404, detail=str(exc)) from exc
 
-        stable_chat_id = (chat_id or message_id or f"ephemeral-{uuid.uuid4().hex}").strip()
+        async def forward_direct():
-        selected_mode = await app.state.store.select_mode(identity.user_id, chat_id or "", spec.mode)
-        if selected_mode == "work" and spec.mode == "chat":
-            raise HTTPException(
-                status_code=status.HTTP_409_CONFLICT,
-                detail="This conversation has been upgraded to Work and cannot return to Chat.",
-            )
-
-        if spec.mode == "chat":
             payload = body.model_dump(exclude_none=True)
             payload["model"] = spec.provider_model
             response = await app.state.provider.forward(payload)
@@ -183,6 +187,21 @@ def create_app(
                 headers=headers,
             )
 
+        background_task = str((body.metadata or {}).get("task", "")).strip()
+        if background_task in OPENWEBUI_BACKGROUND_TASKS:
+            return await forward_direct()
+
+        stable_chat_id = (chat_id or message_id or f"ephemeral-{uuid.uuid4().hex}").strip()
+        selected_mode = await app.state.store.select_mode(identity.user_id, chat_id or "", spec.mode)
+        if selected_mode == "work" and spec.mode == "chat":
+            raise HTTPException(
+                status_code=status.HTTP_409_CONFLICT,
+                detail="This conversation has been upgraded to Work and cannot return to Chat.",
+            )
+
+        if spec.mode == "chat":
+            return await forward_direct()
+
         messages = [message.model_dump(exclude_none=True) for message in body.messages]
         if not body.stream:
 

agent_platform/runtime/loop.py

@@ -32,6 +32,9 @@ relevant verification, and report concrete results. Use tools instead of inventi
 command output. A claim that a file was created, changed, executed, or tested must be backed by a
 successful tool result from this run. For scripts, reports, code, and other deliverables, create real
 files under /workspace, inspect them after writing, run relevant checks, and cite their exact paths.
+Write source files with real line breaks rather than literal `\\n` escape sequences. If any command
+fails, inspect the failure, repair the underlying problem, and successfully rerun the relevant check;
+never replace failed output with invented numbers, placeholders, or a hand-written success report.
 Never use an interactive command such as bare `python3`, `bash`, or `node`; provide a complete
 non-interactive command. Do not use `echo` merely to simulate progress. Keep the plan current. Use
 background processes for servers or long jobs. Delegate bounded independent investigations when it
@@ -59,8 +62,15 @@ ACTION_REQUEST = re.compile(
     r"\b(?:run|execute|test|inspect|analy[sz]e|compare|debug|deploy)\b)",
     re.IGNORECASE,
 )
+EXECUTION_REQUEST = re.compile(
+    r"(运行|执行|测试|脚本|程序|代码|比较|对比|基准|"
+    r"\b(?:run|execute|test|script|program|code|compare|benchmark)\b)",
+    re.IGNORECASE,
+)
 MUTATION_TOOLS = {"write_file", "apply_patch"}
-VERIFICATION_TOOLS = {"read_file", "list_files", "search_files", "git_status", "git_diff", "exec", "poll_process"}
+VERIFICATION_TOOLS = {"read_file", "git_status", "git_diff", "exec", "poll_process"}
+EXECUTION_TOOLS = {"exec", "start_process", "poll_process"}
+MAX_CONSECUTIVE_CHECKPOINT_REJECTIONS = 3
 
 
 def latest_user_text(messages: list[dict[str, Any]]) -> str:
@@ -78,19 +88,49 @@ def latest_user_text(messages: list[dict[str, Any]]) -> str:
     return ""
 
 
-def completion_failure(artifact_required: bool, mutation_count: int, verified_mutation: int) -> str | None:
+def completion_failure(
+    artifact_required: bool,
+    mutation_count: int,
+    verified_mutation: int,
+    action_required: bool = False,
+    execution_required: bool = False,
+    successful_executions: int = 0,
+    unresolved_execution_failure: str | None = None,
+) -> str | None:
+    reasons: list[str] = []
     if artifact_required and mutation_count == 0:
-        return (
+        reasons.append(
             "The user requested a concrete deliverable, but this run has no successful file-writing tool call. "
-            "Do not claim completion. Create the requested files with write_file or apply_patch, then inspect "
+            "Create the requested files with write_file or apply_patch."
-            "and verify them with tools."
         )
-    if artifact_required and verified_mutation < mutation_count:
+    elif artifact_required and verified_mutation < mutation_count:
-        return (
+        reasons.append(
             "Files changed, but no successful verification occurred after the latest change. Inspect the actual "
-            "files and run the relevant check before claiming completion."
+            "files and run the relevant check."
         )
-    return None
+    if action_required and unresolved_execution_failure:
+        reasons.append(
+            "A command or process check failed and has not been followed by a successful execution. "
+            f"Repair and rerun it. Last failure: {unresolved_execution_failure}"
+        )
+    elif execution_required and successful_executions == 0:
+        reasons.append(
+            "The task requires a script, code, test, benchmark, or comparison, but no command or process "
+            "completed successfully. Run the relevant verification and use its real output."
+        )
+    if not reasons:
+        return None
+    return " Do not claim completion. ".join(reasons)
+
+
+def recovery_message(failure: str) -> str:
+    return (
+        "Your attempted checkpoint was rejected by the execution policy. Continue the original task now. "
+        "You MUST call the appropriate tools in this response instead of repeating or paraphrasing a final answer. "
+        f"Evidence gap: {failure} "
+        "If a command failed, inspect and repair the real files, rerun the command successfully, then inspect the "
+        "resulting deliverables before finishing."
+    )
 
 
 @dataclass(slots=True)
@@ -147,7 +187,7 @@ class AgentLoop:
             "run.created",
             {
                 "model_tier": spec.strength,
-                "strategy_version": "work-loop-v1",
+                "strategy_version": "work-loop-v2",
                 "scheduler_version": "safe-parallel-v1",
                 "context_policy": "recent-visible-v1",
             },
@@ -200,9 +240,13 @@ class AgentLoop:
         request_text = latest_user_text(messages)
         artifact_required = depth == 0 and bool(ARTIFACT_REQUEST.search(request_text))
         action_required = depth == 0 and bool(ACTION_REQUEST.search(request_text))
+        execution_required = depth == 0 and bool(EXECUTION_REQUEST.search(request_text))
         mutation_count = 0
         verified_mutation = -1
         successful_tools = 0
+        successful_executions = 0
+        unresolved_execution_failure: str | None = None
+        consecutive_checkpoint_rejections = 0
         for iteration in range(max_iterations):
             await recorder.emit(
                 "model.requested",
@@ -228,13 +272,26 @@ class AgentLoop:
             tool_calls = message.get("tool_calls") or []
             if not tool_calls:
                 candidate = str(message.get("content") or "").strip()
-                failure = completion_failure(artifact_required, mutation_count, verified_mutation)
+                failure = completion_failure(
+                    artifact_required,
+                    mutation_count,
+                    verified_mutation,
+                    action_required,
+                    execution_required,
+                    successful_executions,
+                    unresolved_execution_failure,
+                )
                 if not failure and action_required and successful_tools == 0:
                     failure = (
                         "The user requested an executed or inspected result, but no tool completed successfully. "
                         "Use the appropriate workspace tools and return evidence instead of an unverified answer."
                     )
-                if failure and iteration + 1 < max_iterations:
+                consecutive_checkpoint_rejections += 1
+                if (
+                    failure
+                    and iteration + 1 < max_iterations
+                    and consecutive_checkpoint_rejections < MAX_CONSECUTIVE_CHECKPOINT_REJECTIONS
+                ):
                     await recorder.emit(
                         "completion.rejected",
                         {"reason": failure, "iteration": iteration + 1, "depth": depth},
@@ -242,7 +299,7 @@ class AgentLoop:
                     messages.extend(
                         [
                             {"role": "assistant", "content": candidate},
-                            {"role": "system", "content": failure},
+                            {"role": "user", "content": recovery_message(failure)},
                         ]
                     )
                     continue
@@ -257,6 +314,7 @@ class AgentLoop:
                     )
                 return candidate
 
+            consecutive_checkpoint_rejections = 0
             assistant_message = {
                 "role": "assistant",
                 "content": message.get("content"),
@@ -279,6 +337,11 @@ class AgentLoop:
                         mutation_count += 1
                     elif name in VERIFICATION_TOOLS and mutation_count:
                         verified_mutation = mutation_count
+                    if name in EXECUTION_TOOLS:
+                        successful_executions += 1
+                        unresolved_execution_failure = None
+                elif name in EXECUTION_TOOLS:
+                    unresolved_execution_failure = f"{name}: {public_tool_summary(result, 1000)}"
                 messages.append(
                     {
                         "role": "tool",
@@ -299,7 +362,15 @@ class AgentLoop:
         )
         response = await self.provider.complete(model=spec.provider_model, messages=messages, tools=None)
         answer = str(response["choices"][0].get("message", {}).get("content") or "").strip()
-        if completion_failure(artifact_required, mutation_count, verified_mutation):
+        if completion_failure(
+            artifact_required,
+            mutation_count,
+            verified_mutation,
+            action_required,
+            execution_required,
+            successful_executions,
+            unresolved_execution_failure,
+        ):
             return (
                 "未完成：工具迭代已用尽，但交付物没有形成完整的“写入后验证”证据链。"
                 "我不会声称文件已经完成；工作区中已有内容可从“工作区文件”查看。"

agent_platform/runtime/schemas.py

@@ -23,6 +23,7 @@ class ChatCompletionRequest(BaseModel):
     stream: bool = False
     tools: list[dict[str, Any]] | None = None
     tool_choice: Any = None
+    metadata: dict[str, Any] | None = None
 
 
 class PlanItem(BaseModel):

compose.yaml

@@ -39,6 +39,11 @@ services:
       TASK_MODEL: chat-light
       ENABLE_TITLE_GENERATION: "false"
       ENABLE_FOLLOW_UP_GENERATION: "false"
+      ENABLE_TAGS_GENERATION: "false"
+      ENABLE_SEARCH_QUERY_GENERATION: "false"
+      ENABLE_RETRIEVAL_QUERY_GENERATION: "false"
+      ENABLE_AUTOCOMPLETE_GENERATION: "false"
+      ENABLE_IMAGE_PROMPT_GENERATION: "false"
       ENABLE_OLLAMA_API: "false"
       ENABLE_DIRECT_CONNECTIONS: "false"
       ENABLE_PERSISTENT_CONFIG: "false"

deploy/docker-compose.yml

@@ -39,6 +39,11 @@ services:
       TASK_MODEL: chat-light
       ENABLE_TITLE_GENERATION: "false"
       ENABLE_FOLLOW_UP_GENERATION: "false"
+      ENABLE_TAGS_GENERATION: "false"
+      ENABLE_SEARCH_QUERY_GENERATION: "false"
+      ENABLE_RETRIEVAL_QUERY_GENERATION: "false"
+      ENABLE_AUTOCOMPLETE_GENERATION: "false"
+      ENABLE_IMAGE_PROMPT_GENERATION: "false"
       ENABLE_OLLAMA_API: "false"
       ENABLE_DIRECT_CONNECTIONS: "false"
       ENABLE_PERSISTENT_CONFIG: "false"

tests/test_agent_loop.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import asyncio
+import copy
 import json
 from typing import Any
 
@@ -25,7 +26,7 @@ class ScriptedProvider:
         self.requests = []
 
     async def complete(self, **kwargs):
-        self.requests.append(kwargs)
+        self.requests.append(copy.deepcopy(kwargs))
         return self.responses.pop(0)
 
 
@@ -203,7 +204,7 @@ async def test_artifact_completion_is_rejected_until_written_and_verified(settin
     try:
         answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
             spec=get_model_spec("work-light"),
-            messages=[{"role": "user", "content": "写一个脚本并生成报告"}],
+            messages=[{"role": "user", "content": "写一个报告"}],
             identity=UserIdentity("u1", "", "", "user"),
             raw_user_jwt="jwt",
             chat_id="c1",
@@ -212,6 +213,191 @@ async def test_artifact_completion_is_rejected_until_written_and_verified(settin
         assert answer == "已完成 report.md"
         assert [name for name, _ in registry.calls] == ["write_file", "read_file"]
         assert "completion.rejected" in events
+        assert provider.requests[1]["messages"][-1]["role"] == "user"
+        assert "MUST call" in provider.requests[1]["messages"][-1]["content"]
+    finally:
+        await store.close()
+
+
+class FailingExecutionRegistry(RecordingRegistry):
+    def __init__(self) -> None:
+        super().__init__()
+        self.exec_calls = 0
+
+    async def execute(self, name, arguments, context):
+        self.calls.append((name, arguments))
+        if name == "exec":
+            self.exec_calls += 1
+            if self.exec_calls == 1:
+                return {"ok": False, "error": "SyntaxError: invalid source"}
+        return {"ok": True, "output": f"{name} complete"}
+
+
+async def test_failed_execution_must_be_repaired_before_completion(settings) -> None:
+    write = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("1", "write_file", {"path": "sort.py", "content": "broken"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    read = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("2", "read_file", {"path": "sort.py"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    failed_exec = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("3", "exec", {"command": "python3 sort.py"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    premature = {"choices": [{"message": {"role": "assistant", "content": "已完成"}, "finish_reason": "stop"}]}
+    repaired_exec = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("4", "exec", {"command": "python3 sort.py"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {
+        "choices": [{"message": {"role": "assistant", "content": "脚本运行和报告验证完成"}, "finish_reason": "stop"}]
+    }
+    provider = ScriptedProvider([write, read, failed_exec, premature, repaired_exec, final])
+    registry = FailingExecutionRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+    events = []
+
+    async def callback(event_type, payload):
+        events.append(event_type)
+
+    try:
+        answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "写脚本对比排序算法并生成报告"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer == "脚本运行和报告验证完成"
+        assert registry.exec_calls == 2
+        assert events.count("completion.rejected") == 1
+        recovery = provider.requests[4]["messages"][-1]
+        assert recovery["role"] == "user"
+        assert "SyntaxError" in recovery["content"]
+    finally:
+        await store.close()
+
+
+async def test_script_delivery_requires_successful_execution(settings) -> None:
+    write = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("1", "write_file", {"path": "script.py", "content": "print('ok')"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    read = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("2", "read_file", {"path": "script.py"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    premature = {"choices": [{"message": {"role": "assistant", "content": "脚本完成"}, "finish_reason": "stop"}]}
+    execute = {
+        "choices": [
+            {
+                "message": {
+                    "role": "assistant",
+                    "tool_calls": [tool_call("3", "exec", {"command": "python3 script.py"})],
+                },
+                "finish_reason": "tool_calls",
+            }
+        ]
+    }
+    final = {"choices": [{"message": {"role": "assistant", "content": "脚本执行验证完成"}, "finish_reason": "stop"}]}
+    provider = ScriptedProvider([write, read, premature, execute, final])
+    registry = RecordingRegistry()
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+
+    async def callback(event_type, payload):
+        return None
+
+    try:
+        answer = await AgentLoop(provider, registry, store, max_tool_output_chars=10_000).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "写一个 Python 脚本"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer == "脚本执行验证完成"
+        assert [name for name, _ in registry.calls] == ["write_file", "read_file", "exec"]
+        assert "no command or process completed successfully" in provider.requests[3]["messages"][-1]["content"]
+    finally:
+        await store.close()
+
+
+async def test_rejected_checkpoint_does_not_spin_without_tool_calls(settings) -> None:
+    stopped = {"choices": [{"message": {"role": "assistant", "content": "已完成"}, "finish_reason": "stop"}]}
+    provider = ScriptedProvider([stopped, stopped, stopped])
+    store = RuntimeStore(settings.database_url)
+    await store.initialize()
+    events = []
+
+    async def callback(event_type, payload):
+        events.append(event_type)
+
+    try:
+        answer = await AgentLoop(
+            provider,
+            RecordingRegistry(),
+            store,
+            max_tool_output_chars=10_000,
+        ).run(
+            spec=get_model_spec("work-light"),
+            messages=[{"role": "user", "content": "写一个报告"}],
+            identity=UserIdentity("u1", "", "", "user"),
+            raw_user_jwt="jwt",
+            chat_id="c1",
+            callback=callback,
+        )
+        assert answer.startswith("未完成：")
+        assert len(provider.requests) == 3
+        assert events.count("completion.rejected") == 2
+        assert events.count("completion.unverified") == 1
     finally:
         await store.close()
 

tests/test_runtime_api.py

@@ -136,6 +136,25 @@ def test_chat_stream_hides_provider_model_and_preserves_native_tool_calls(settin
     assert "data: [DONE]" in response.text
 
 
+def test_openwebui_background_task_bypasses_work_loop(settings, identity_jwt) -> None:
+    provider = FakeModelProvider()
+    app = create_app(settings, store=RuntimeStore(settings.database_url), provider=provider, tools=FakeTools())
+    with TestClient(app) as client:
+        response = client.post(
+            "/v1/chat/completions",
+            headers=headers(settings, identity_jwt, "background-task"),
+            json={
+                "model": "work-medium",
+                "messages": [{"role": "user", "content": "Generate tags"}],
+                "metadata": {"task": "tags_generation"},
+            },
+        )
+        assert response.status_code == 200
+        assert response.json()["model"] == "work-medium"
+        assert provider.completed == []
+        assert provider.forwarded[-1]["model"] == "ChatGPT-5.6:Terra"
+
+
 def test_work_stream_never_emits_delta_tool_calls(settings, identity_jwt) -> None:
     app = create_app(
         settings,