fix: preserve exploration tool history and cap context budget

apps/showcase_ui/src/app/components/agent-stream/agent-stream.component.html

@@ -401,7 +401,7 @@ limitations under the License.
                 @if (runInfoContextPercent() !== null) {
                   <span class="run-info-value">
                     {{ runInfoContextPercent() }}%
-                    <span class="run-info-sub-inline">/ {{ formatRunTokens(runUsage()?.operator_context_window_tokens) }}</span>
+                    <span class="run-info-sub-inline">/ {{ formatRunTokens(runInfoContextWindow()) }}</span>
                   </span>
                 } @else {
                   <span class="run-info-value is-muted">—</span>
@@ -425,7 +425,9 @@ limitations under the License.
                 <span class="run-info-label">屏幕识别</span>
                 <span class="run-info-value">{{ runTuningLabel('explore', runInfoTuning()?.explorer_mode) }}</span>
               </div>
-            } @else {
+            } @else if (runInfoMode() === 'explore') {
+              <div class="run-info-note">探索支持对话指导与人工示范，模型和时限使用独立的探索设置。</div>
+            } @else if (runInfoMode() === 'flash') {
               <div class="run-info-note">Flash 无需额外设置。</div>
             }
           </div>

apps/showcase_ui/src/app/components/agent-stream/agent-stream.component.ts

@@ -35,6 +35,8 @@ import { OverlayModule, ConnectedPosition } from '@angular/cdk/overlay';
 import { StepBlock, PhaseBlock, StepEvent, ActionParam, CheckerResult, StreamResetNotice, DEFAULT_STREAM_RESET_MESSAGE } from '../../core/models/stream.model';
 import {
   contextPercent,
+  contextWindow,
+  runMode,
   formatCompactTokens,
   formatElapsed,
   parseRunTuning,
@@ -297,7 +299,9 @@ export class AgentStreamComponent implements AfterViewInit {
   public runUsage = signal<SessionUsage | null>(null);
   private runInfoClock = signal<number>(Date.now());
 
-  public runInfoIsPro = computed(() => this.getModelClass(this.agentService.viewedModel()?.name) === 'is-pro');
+  public runInfoMode = computed(() => runMode(this.agentService.currentSession(), this.agentService.viewedModel()));
+  public runInfoIsPro = computed(() => this.runInfoMode() === 'pro');
+  public runInfoContextWindow = computed(() => contextWindow(this.runUsage()?.operator_context_window_tokens));
 
   public runInfoElapsed = computed(() => {
     const session = this.agentService.currentSession();

apps/showcase_ui/src/app/utils/run-info.util.spec.ts

@@ -1,5 +1,6 @@
 import {
   contextPercent,
+  runMode,
   formatCompactTokens,
   formatElapsed,
   parseRunTuning,
@@ -35,13 +36,14 @@ describe('run-info.util', () => {
 
   describe('contextPercent', () => {
     it('reports one-decimal share of the window and clamps at 100', () => {
-      expect(contextPercent(150_000, 1_000_000)).toBe(15);
+      expect(contextPercent(150_000, 300_000)).toBe(50);
-      expect(contextPercent(123_456, 1_000_000)).toBe(12.3);
+      expect(contextPercent(123_456, 300_000)).toBe(41.2);
       expect(contextPercent(2_000_000, 1_000_000)).toBe(100);
     });
 
-    it('falls back to the 1M window and is null when usage is unknown', () => {
+    it('falls back to the 300k window and is null when usage is unknown', () => {
-      expect(contextPercent(250_000, undefined)).toBe(25);
+      expect(contextPercent(150_000, undefined)).toBe(50);
+      expect(contextPercent(150_000, 1_000_000)).toBe(50);
       expect(contextPercent(null, 1_000_000)).toBeNull();
       expect(contextPercent(undefined, 1_000_000)).toBeNull();
     });
@@ -90,3 +92,23 @@ describe('run-info.util', () => {
     });
   });
 });
+
+
+describe('runMode', () => {
+  const model = { name: 'mimo-v2.6-pro', id: 'mimo-v2.6-pro', provider: 'openai', architecture: 'flash' };
+  it('uses recorded execution mode, never the API model name', () => {
+    expect(runMode({ profile: 'pro' }, model)).toBe('pro');
+    expect(runMode(null, model)).toBe('flash');
+    expect(runMode(null, { ...model, name: 'mimo-v2.6-pro', architecture: undefined })).toBe('unknown');
+  });
+  it('distinguishes exploration from a Flash or Pro validation rerun', () => {
+    expect(runMode({ profile: 'pro', exploration_kind: 'exploration' }, model)).toBe('explore');
+    expect(runMode({ profile: 'pro', exploration_kind: 'validation' }, model)).toBe('pro');
+    expect(runMode({ profile: 'flash', exploration_kind: 'validation' }, model)).toBe('flash');
+  });
+  it('supports older saved mode names without defaulting missing data to Flash', () => {
+    expect(runMode(null, { ...model, name: 'ARTEMIS Pro', architecture: undefined })).toBe('pro');
+    expect(runMode(null, { ...model, name: '探索', architecture: undefined })).toBe('explore');
+    expect(runMode(null, null)).toBe('unknown');
+  });
+});

apps/showcase_ui/src/app/utils/run-info.util.ts

@@ -27,10 +27,24 @@ import {
   DEFAULT_EXPLORER_MODE,
   DEFAULT_VERIFICATION_LEVEL
 } from '../core/models/pro-tuning.model';
-import { SessionRunTuning } from '../core/models/session.model';
+import { ModelInfo, Session, SessionRunTuning } from '../core/models/session.model';
 
 /** Denominator for "context used" when the backend does not report one. */
-export const OPERATOR_CONTEXT_WINDOW_TOKENS = 1_000_000;
+export const OPERATOR_CONTEXT_WINDOW_TOKENS = 300_000;
+
+/** Execution mode is independent of the API model name (e.g. MiMo Pro). */
+export function runMode(session?: Partial<Session> | null, model?: ModelInfo | null): 'flash' | 'pro' | 'explore' | 'unknown' {
+  if (session?.exploration_kind === 'exploration') return 'explore';
+  const recorded = session?.profile || session?.model_info?.architecture || model?.architecture;
+  const mode = (recorded || session?.model_info?.name || model?.name || '').toLowerCase().replace(/^artemis\s+/, '');
+  if (mode === '探索' || mode === 'explore') return 'explore';
+  return mode === 'flash' || mode === 'pro' ? mode : 'unknown';
+}
+
+export function contextWindow(windowTokens?: number | null): number {
+  const value = Number(windowTokens);
+  return value > 0 && Number.isFinite(value) ? Math.min(value, OPERATOR_CONTEXT_WINDOW_TOKENS) : OPERATOR_CONTEXT_WINDOW_TOKENS;
+}
 
 /** `42s`, `12m 05s`, `1h 02m 05s`. Negative or invalid input reads as `0s`. */
 export function formatElapsed(totalSeconds: number | null | undefined): string {
@@ -58,7 +72,7 @@ export function contextPercent(
   windowTokens: number | null | undefined
 ): number | null {
   const used = Number(usedTokens);
-  const window = Number(windowTokens) || OPERATOR_CONTEXT_WINDOW_TOKENS;
+  const window = contextWindow(windowTokens);
   if (!Number.isFinite(used) || used < 0 || usedTokens === null || usedTokens === undefined) {
     return null;
   }

artemis/agents/operator/operator.py

@@ -17,7 +17,7 @@ from collections.abc import Awaitable
 from typing import Any
 from uuid import uuid4
 
-from langchain_core.messages import SystemMessage, ToolMessage
+from langchain_core.messages import BaseMessage, SystemMessage, ToolMessage
 from langchain_core.tools import BaseTool
 
 from artemis.agents.validator.tool_declarations import normalize_coordinate_target
@@ -146,9 +146,9 @@ class OperatorNode:
                 from artemis.config.agent import MemoryTranscriptConfig
 
                 self._transcript_cfg = MemoryTranscriptConfig()
-        # Index into the rendered message list where this turn's tail begins;
+        # Keep the actual observation, not an index: plugin/exploration context
-        # set by the transcript build and consumed after the tool loop.
+        # may be prepended after render, shifting every message's position.
-        self._transcript_turn_base: int | None = None
+        self._transcript_turn_tail: BaseMessage | None = None
         self._turn_plan_baseline: str | None = None
         self._turn_ledger_streak: int = 0
         # Whether the previous turn's replies carried no visible reasoning text;
@@ -225,7 +225,7 @@ class OperatorNode:
         active_background_tasks: list[dict] = None,
         newly_finished_tasks: list[dict] = None,
     ) -> list:
-        self._transcript_turn_base = None
+        self._transcript_turn_tail = None
         if getattr(self._transcript_cfg, "enabled", False):
             try:
                 return await self._build_prompt_transcript(
@@ -243,7 +243,7 @@ class OperatorNode:
                     f"Transcript prompt path failed; falling back to the legacy"
                     f" 2-message build for this turn: {e}"
                 )
-                self._transcript_turn_base = None
+                self._transcript_turn_tail = None
 
         plan_and_history = self._build_plan_and_history(steps, task_plan)
 
@@ -468,18 +468,19 @@ class OperatorNode:
             mark_ephemeral(tail, builder.ephemeral_indices)
 
         messages = ledger.render([tail])
-        self._transcript_turn_base = len(messages) - 1
+        self._transcript_turn_tail = tail
         return messages
 
     def _stage_transcript_turn(self, messages: list) -> None:
         """Hand this turn's tail + tool-loop products to the ledger."""
-        base = self._transcript_turn_base
+        tail = self._transcript_turn_tail
-        self._transcript_turn_base = None
+        self._transcript_turn_tail = None
-        if base is None:
+        if tail is None:
             return
         try:
             ledger = getattr(self.ctx, "transcript_ledger", None)
             if ledger is not None:
+                base = next(i for i, message in enumerate(messages) if message is tail)
                 ledger.stage_turn(messages[base:])
         except Exception as e:
             logger.error(f"Failed to stage transcript turn: {e}")

artemis/config/agent.py

@@ -23,6 +23,7 @@ from pydantic import BaseModel, Field, model_validator
 from artemis.config.constants import (
     AGENT_CONFIG_FILENAME,
     DEFAULT_EXPLORER_VERSION,
+    MAX_CONTEXT_WINDOW_TOKENS,
     ExplorerVersion,
 )
 from artemis.config.paths import ROOT_DIR, get_config_path
@@ -483,6 +484,7 @@ class MemoryTranscriptConfig(BaseModel):
     context_budget_tokens: int = Field(
         default=80000,
         ge=1024,
+        le=MAX_CONTEXT_WINDOW_TOKENS,
         description=(
             "Measured-token budget for the transcript context (provisional"
             " default pending the on-device baseline; §3.4). The soft/hard"

artemis/config/constants.py

@@ -16,6 +16,9 @@
 
 from typing import Literal
 
+# Application context ceiling; independent from output and session token totals.
+MAX_CONTEXT_WINDOW_TOKENS = 300_000
+
 # ==============================================================================
 # Environment Variable Keys
 # ==============================================================================

artemis/llm/reliability.py

@@ -81,6 +81,11 @@ _AUTH_MARKERS = ("unauthorized", "forbidden", "api key", "credential")
 # "deadline exceeded"/"deadline_exceeded" cover google-genai SDK timeouts that
 # surface without an HTTP status code attached.
 _TIMEOUT_MARKERS = ("timed out", "timeout", "deadline exceeded", "deadline_exceeded")
+_TOOL_HISTORY_ERRORS = (
+    "no tool call found for function call output",
+    "no tool output found for function call",
+    "tool_call_id not found",
+)
 
 
 def classify_failure(error: BaseException) -> Failure:
@@ -91,6 +96,12 @@ def classify_failure(error: BaseException) -> Failure:
     code = extract_status_code(error)
     message = str(error).lower()
 
+    # Compatible Responses endpoints can report input validation inside an
+    # HTTP 200 stream as APIError without a status code. Retrying the same
+    # broken call/output history cannot repair it, even on a fallback model.
+    if any(marker in message for marker in _TOOL_HISTORY_ERRORS):
+        return Failure(FailureCategory.BAD_REQUEST, False, False)
+
     if code == 429 or any(marker in message for marker in _RATE_LIMIT_MARKERS):
         return Failure(FailureCategory.RATE_LIMIT, True, True)
     if code in {500, 502, 503, 504} or any(marker in message for marker in _UNAVAILABLE_MARKERS):

artemis/resources/config/artemis.jsonc

@@ -94,6 +94,7 @@
       },
       "transcript": {
         "enabled": true,
+        // 历史压缩预算，允许范围 1024–300000；上下文参照上限固定为 300k。
         "context_budget_tokens": 80000,
         "start_ratio": 0.35,
         "soft_ratio": 0.7,

artemis/services/token_meter.py

@@ -40,6 +40,7 @@ import threading
 from typing import Any
 
 from artemis.data_engine.context_vars import CURRENT_NODE_NAME
+from artemis.config.constants import MAX_CONTEXT_WINDOW_TOKENS
 from artemis.llm.usage import usage_from_message
 from artemis.utils.logger import get_logger
 
@@ -52,7 +53,7 @@ _METERS: dict[str, "SessionTokenMeter"] = {}
 # measured against when the UI reports "context used". The prompt size of the
 # last executor call (``context_base_tokens`` on its ``llm_usage`` trace) is
 # the numerator; this is the denominator.
-OPERATOR_CONTEXT_WINDOW_TOKENS = 1_000_000
+OPERATOR_CONTEXT_WINDOW_TOKENS = MAX_CONTEXT_WINDOW_TOKENS
 
 # Traced node names whose LLM calls carry the live executor context. Both the
 # Pro Operator and the Flash runner are "the operator" from the user's view.

config/artemis.jsonc

@@ -94,6 +94,7 @@
       },
       "transcript": {
         "enabled": true,
+        // 历史压缩预算，允许范围 1024–300000；上下文参照上限固定为 300k。
         "context_budget_tokens": 80000,
         "start_ratio": 0.35,
         "soft_ratio": 0.7,

docs/ai-automation.md

@@ -80,3 +80,5 @@ Flash 和 Pro 默认复用只读 Checker，以任务提示词和用户指导中
 2.0.0 插件协议与可运行示例见 [开发规则](plugin-development.md)。单任务和批次全部资料包含各来源截图/录像、插件附件及包内证据索引。
 
 通过 `mobile_config(section="capabilities")` 可读取 `recording` 参数；用户授权后可保存全局默认。单任务用 `runtime_overrides={"recording":{"codec":"h264","computer_scope":"window","max_long_edge":1280,"fps":15}}`，批次在 `task_config.recording` 提供同样对象。JSON 配置的 `agent.recording` 覆盖环境默认；不要把录像参数当作模型截图参数。录屏开关继续独立且默认关闭。
+
+运行信息的上下文参照上限固定为 300,000 tokens，前后端使用相同口径。历史压缩预算默认 80,000，JSON 的 `agent.memory.transcript.context_budget_tokens` 可调但不能超过 300,000；这是压缩触发预算，不是单次生成的输出 token 数，也不是整项任务累计 token 的限额。

tests/conftest.py

@@ -46,6 +46,25 @@ def mock_driver():
     return MockDeviceDriver(device_id="fixture-mock-device", width=1080, height=2400)
 
 
+@pytest.fixture(autouse=True)
+def no_host_data_engine_ipc(monkeypatch, tmp_path, request):
+    """Unit-test events must not reach a user's running desktop queue.
+
+    The reconnect path also reads a port file, so overriding the inherited
+    environment port alone does not isolate tests. IPC tests can explicitly
+    override these discovery functions with their own endpoints or mocks.
+    """
+    if any(
+        request.node.get_closest_marker(name)
+        for name in ("android", "integration", "e2e", "manual")
+    ):
+        return
+    monkeypatch.setattr("artemis.data_engine.engine.read_ipc_port", lambda: None)
+    monkeypatch.setattr(
+        "artemis.data_engine.engine.get_ipc_port_file", lambda: tmp_path / "no-host-ipc"
+    )
+
+
 @pytest.fixture(autouse=True)
 def isolated_plugin_registry(monkeypatch, tmp_path, request):
     """Unit tests never load or mutate the user's installed plugins and credentials."""

tests/unit/admin_console/test_session_usage_endpoint.py

@@ -80,7 +80,7 @@ def test_usage_sums_all_agents_and_tracks_latest_operator_context(tmp_path):
     assert usage["cached_tokens"] == 90_000
     assert usage["operator_context_tokens"] == 150_000
     assert usage["operator_context_updated_at"] == 13.0
-    assert usage["operator_context_window_tokens"] == 1_000_000
+    assert usage["operator_context_window_tokens"] == 300_000
     assert usage["profile"] == "pro"
     assert usage["run_tuning"] == {"verification_level": "checkpoints", "explorer_mode": "pro"}
 

tests/unit/agents/test_operator_transcript.py

@@ -30,7 +30,7 @@ from types import SimpleNamespace
 from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
-from langchain_core.messages import HumanMessage, SystemMessage
+from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 
 from artemis.agents.operator.operator import OperatorNode
 from artemis.agents.operator.prompts import load_operator_prompts
@@ -141,6 +141,67 @@ def _transcript_state(**overrides):
     return state
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "exploration,plugins", [(False, False), (True, False), (False, True), (True, True)]
+)
+async def test_injected_prefix_and_compression_preserve_tool_pairs(
+    monkeypatch, exploration, plugins
+):
+    from langchain_openai import ChatOpenAI
+
+    ctx = _transcript_ctx()
+    ctx.exploration_id = "test-exploration" if exploration else None
+    ctx.plugins = SimpleNamespace(prompt=lambda: "Plugin instructions") if plugins else None
+    monkeypatch.setattr(
+        "artemis.exploration.runtime.exploration_messages",
+        lambda _: [
+            SystemMessage(content="Exploration instructions"),
+            HumanMessage(content="User guidance"),
+        ],
+    )
+    monkeypatch.setattr("artemis.agents.operator.operator.get_llm", lambda **_: MagicMock())
+    node = OperatorNode(ctx, transcript_config=MemoryTranscriptConfig(enabled=True))
+    requests = []
+    wire_model = ChatOpenAI(model="test", api_key="test", use_responses_api=True, store=False)
+
+    async def tool_loop(current_messages, **_kwargs):
+        messages = current_messages
+        payload = wire_model._get_request_payload(messages)
+        requests.append(payload["input"])
+        call_id = f"call_{len(requests)}"
+        messages.extend(
+            [
+                AIMessage(
+                    content="", tool_calls=[{"name": "click", "args": {"target": 1}, "id": call_id}]
+                ),
+                ToolMessage(content="Action Recorded", tool_call_id=call_id),
+            ]
+        )
+        return [{"action": "tap"}], None, None, False
+
+    monkeypatch.setattr(node, "_invoke_llm_loop", tool_loop)
+    for turn in range(8):
+        await node(_transcript_state(current_step_id=f"step-{turn}"))
+        if turn == 4:
+            ctx.transcript_ledger.freeze_turns(2, [HumanMessage(content="Summarized first turns")])
+
+    for items in requests:
+        pending, seen = set(), set()
+        for item in items:
+            if item.get("type") == "function_call":
+                assert item["call_id"] not in seen
+                pending.add(item["call_id"])
+                seen.add(item["call_id"])
+            elif item.get("type") == "function_call_output":
+                assert item["call_id"] in pending
+                pending.remove(item["call_id"])
+        assert not pending
+    assert {item["call_id"] for item in requests[-1] if item.get("type") == "function_call"} == {
+        f"call_{n}" for n in range(3, 8)
+    }
+
+
 def _action_llm(captured: list):
     mock_llm = MagicMock()
     mock_response = MagicMock()

tests/unit/test_llm_gateway.py

@@ -215,6 +215,40 @@ async def test_with_fallback_refuses_to_hide_bad_requests():
     assert calls["fallback"] == 0
 
 
+@pytest.mark.asyncio
+@pytest.mark.parametrize("method", ["ainvoke", "complete"])
+async def test_streamed_tool_pairing_error_does_not_retry_pause_or_fallback(monkeypatch, method):
+    import httpx
+    from openai import APIError
+    from unittest.mock import AsyncMock
+
+    error = APIError(
+        "No tool call found for function call output with call_id call_test.",
+        request=httpx.Request("POST", "https://example.test/responses"),
+        body=None,
+    )
+    calls = []
+
+    class BrokenHistory:
+        async def ainvoke(self, *_args, **_kwargs):
+            calls.append("invoke")
+            raise error
+
+        async def astream(self, *_args, **_kwargs):
+            calls.append("stream")
+            raise error
+            yield  # Make this a stream whose error arrives before any chunks.
+
+    pause, fallback = AsyncMock(), AsyncMock()
+    monkeypatch.setattr(llm_service, "_wait_for_resume", pause)
+    wrapper = RobustChatModelWrapper(BrokenHistory())
+    with pytest.raises(LLMPermanentError):
+        await with_fallback(lambda: getattr(wrapper, method)([]), fallback)
+    assert len(calls) == 1
+    pause.assert_not_called()
+    fallback.assert_not_called()
+
+
 @pytest.mark.asyncio
 async def test_acomplete_structured_corrective_reask():
     class CorrectsItself:

tests/unit/test_memory_config.py

@@ -19,6 +19,16 @@ import pytest
 from artemis.config import AgentGlobalConfig, load_agent_config
 
 
+def test_transcript_budget_cannot_exceed_300k_context_ceiling():
+    from pydantic import ValidationError
+    from artemis.config.agent import MemoryTranscriptConfig
+
+    assert MemoryTranscriptConfig().context_budget_tokens == 80_000
+    assert MemoryTranscriptConfig(context_budget_tokens=300_000).context_budget_tokens == 300_000
+    with pytest.raises(ValidationError):
+        MemoryTranscriptConfig(context_budget_tokens=300_001)
+
+
 def test_memory_defaults():
     cfg = AgentGlobalConfig.model_validate({})
     assert cfg.flash.step_summarizer.timeout_seconds == 120.0