fix: disable chat checkpoint resume

config.example.yaml

@@ -304,7 +304,7 @@ multi_agent:
     plan_execute_executed_steps_budget_ratio: 0.2 # plan_execute 中 executed_steps 预算比例
     plan_execute_max_step_result_runes: 4000 # plan_execute 每步结果最大字符数（超出截断）
     plan_execute_keep_last_steps: 8 # plan_execute 仅保留最近 N 步正文，早期步骤折叠为标题
-    checkpoint_dir: data/eino-checkpoints # P0：进程崩溃/OOM 后同会话自动 ADK Resume；正常结束会删 .ckpt；与「中断并继续」(last_react_*) 是两套机制
+    checkpoint_dir: "" # 聊天链路不再使用 ADK checkpoint；跨轮模型态统一走 conversations.last_react_*，便于排查 stale context
     model_retry_max_retries: 0 # Eino 原生 ChatModel retry；408/409/425/429/5xx/网络抖动/空流式输出会重试；0=默认 4（永久性 4xx 不重试）
     model_retry_max_backoff_sec: 0 # Eino 原生 ChatModel retry 单次退避上限秒数；0=默认 30
     model_failover_channels: [] # Eino 原生 ChatModel failover；填写 ai.channels ID，例如 [qwen-plus]；retry 耗尽后按顺序切换

internal/config/config.go

@@ -298,7 +298,8 @@ type MultiAgentEinoMiddlewareConfig struct {
 	PlanExecuteMaxStepResultRunes int `yaml:"plan_execute_max_step_result_runes,omitempty" json:"plan_execute_max_step_result_runes,omitempty"`
 	// PlanExecuteKeepLastSteps keeps only the tail steps in prompt view (default 8).
 	PlanExecuteKeepLastSteps int `yaml:"plan_execute_keep_last_steps,omitempty" json:"plan_execute_keep_last_steps,omitempty"`
-	// CheckpointDir when non-empty enables adk.Runner CheckPointStore (file-backed) for interrupt/resume persistence.
+	// CheckpointDir is retained for config compatibility. Chat agent runs do
+	// not consume it; cross-turn recovery is centralized in conversations.last_react_*.
 	CheckpointDir string `yaml:"checkpoint_dir,omitempty" json:"checkpoint_dir,omitempty"`
 	// DeepOutputKey passed to deep.Config OutputKey (session final text); empty = off.
 	DeepOutputKey string `yaml:"deep_output_key,omitempty" json:"deep_output_key,omitempty"`

internal/multiagent/eino_single_runner.go

@@ -203,15 +203,18 @@ func RunEinoSingleChatModelAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                "eino_single",
-		OrchestratorName:        einoSingleAgentName,
-		ConversationID:          conversationID,
-		Progress:                progress,
-		Logger:                  logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		OrchMode:             "eino_single",
+		OrchestratorName:     einoSingleAgentName,
+		ConversationID:       conversationID,
+		Progress:             progress,
+		Logger:               logger,
+		SnapshotMCPIDs:       snapshotMCPIDs,
+		StreamsMainAssistant: streamsMainAssistant,
+		EinoRoleTag:          einoRoleTag,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,

internal/multiagent/runner.go

@@ -614,15 +614,18 @@ func RunDeepAgent(
 	}
 
 	return runEinoADKAgentLoop(ctx, &einoADKRunLoopArgs{
-		OrchMode:                orchMode,
-		OrchestratorName:        orchestratorName,
-		ConversationID:          conversationID,
-		Progress:                progress,
-		Logger:                  logger,
-		SnapshotMCPIDs:          snapshotMCPIDs,
-		StreamsMainAssistant:    streamsMainAssistant,
-		EinoRoleTag:             einoRoleTag,
-		CheckpointDir:           ma.EinoMiddleware.CheckpointDir,
+		OrchMode:             orchMode,
+		OrchestratorName:     orchestratorName,
+		ConversationID:       conversationID,
+		Progress:             progress,
+		Logger:               logger,
+		SnapshotMCPIDs:       snapshotMCPIDs,
+		StreamsMainAssistant: streamsMainAssistant,
+		EinoRoleTag:          einoRoleTag,
+		// Chat history recovery is intentionally centralized in last_react_*.
+		// ADK checkpoints are a second persisted model-state channel and make
+		// stale-context bugs hard to reason about across user turns.
+		CheckpointDir:           "",
 		RunRetryMaxAttempts:     RunRetryMaxAttemptsFromConfig(&ma.EinoMiddleware),
 		RunRetryMaxBackoffSec:   int(einoRunRetryMaxBackoffFromConfig(&ma.EinoMiddleware).Seconds()),
 		McpIDsMu:                &mcpIDsMu,