Snapshot MaxTurns in checkpoint and apply it on restore

MaxTurns is the only agent bound compared against a counter that is
serialised in the checkpoint (Turns). When config drifts between save
and restore -- typically because a deploy changed WithMaxTurns or a
different build of the agent is registered by name -- cp.Turns can
exceed agent.maxTurns on the resumed run, which previously surfaced
as a warning log and then a MaxTurnsExceededError on the first
iteration of the resumed coreLoop.

Capture MaxTurns in the new AgentConfig on every save, and on
restore clone the registry-resolved agent with WithMaxTurns applied
from the snapshot. The override flows through the outer Restore path
and through both inner-agent resolution sites in
restoreNestedSuspended and restoreAwaitingApproval, so nested
runs get the same treatment. Other loop bounds
(maxEmptyOutputRetries, maxToolDepth) reset per turn / per tool
depth and stay intentionally live so deploys can tune them without
invalidating in-flight checkpoints. Live references (tools, hooks,
LLM client, approval callbacks, guardrails) are not snapshotted for
the same reason.

With the snapshot in place, the "restored agent run has already
reached max turns" warning at the top of continueFromMessages is
structurally unreachable -- the live agent's bound is now the same
value cp.Turns was bounded by at save time -- and is removed.

Signed-off-by: Aurélien Sibiril <81782+aureliensibiril@users.noreply.github.com>

pkg/agent/checkpoint.go

@@ -23,9 +23,25 @@ import (
 type (
 	AgentStatus string
 
+	// AgentConfig captures the subset of agent options that must remain
+	// stable across a suspend/restore cycle to keep the run coherent.
+	// Currently that is only MaxTurns, because Checkpoint.Turns is a
+	// counter compared against it — if the live agent's bound were
+	// lowered below the saved counter we would either short-circuit the
+	// restored run or fail the warning at restoreSuspended. Other loop
+	// bounds (maxEmptyOutputRetries, maxToolDepth) reset per turn and
+	// are safe to change mid-suspension. Live references (tools,
+	// handoffs, hooks, LLM client, approval callbacks, guardrails) are
+	// intentionally not snapshotted so deploys can update behavior
+	// while runs are paused.
+	AgentConfig struct {
+		MaxTurns int
+	}
+
 	Checkpoint struct {
 		Status        AgentStatus
 		AgentName     string
+		Config        AgentConfig
 		Messages      []llm.Message
 		Usage         llm.Usage
 		Turns         int

pkg/agent/restore.go

@@ -20,7 +20,6 @@ import (
 	"fmt"
 	"sync"
 
-	"go.gearno.de/kit/log"
 	"go.probo.inc/probo/pkg/llm"
 )
 
@@ -44,10 +43,23 @@ func Restore(
 	if err != nil {
 		return nil, fmt.Errorf("cannot resolve agent %q: %w", cp.AgentName, err)
 	}
+	agent = applyCheckpointConfig(agent, cp.Config)
 
 	return restoreCheckpoint(ctx, agent, cp, store, runID, registry)
 }
 
+// applyCheckpointConfig returns a clone of agent with the bounds from
+// the checkpoint snapshot overriding the live values. Zero values in
+// cfg fall through to the live agent so older checkpoints written
+// before the Config field existed, or test-constructed Checkpoint
+// literals that omit Config, still resume correctly.
+func applyCheckpointConfig(agent *Agent, cfg AgentConfig) *Agent {
+	if cfg.MaxTurns <= 0 {
+		return agent
+	}
+	return agent.Clone(WithMaxTurns(cfg.MaxTurns))
+}
+
 func restoreCheckpoint(
 	ctx context.Context,
 	agent *Agent,
@@ -96,16 +108,6 @@ func continueFromMessages(
 	messagesCopy := make([]llm.Message, len(messages))
 	copy(messagesCopy, messages)
 
-	if cp.Turns >= agent.maxTurns {
-		agent.logger.WarnCtx(
-			ctx,
-			"restored agent run has already reached max turns",
-			log.String("agent", agent.name),
-			log.Int("turns", cp.Turns),
-			log.Int("max_turns", agent.maxTurns),
-		)
-	}
-
 	return coreLoop(
 		ctx,
 		agent,
@@ -169,6 +171,7 @@ func restoreNestedSuspended(
 			entries[i].err = fmt.Errorf("cannot resolve inner agent %q: %w", innerCP.AgentName, err)
 			continue
 		}
+		innerAgent = applyCheckpointConfig(innerAgent, innerCP.Config)
 
 		wg.Add(1)
 		go func(i int, tc llm.ToolCall, innerAgent *Agent, innerCP *Checkpoint) {
@@ -297,6 +300,7 @@ func restoreAwaitingApproval(
 			if err != nil {
 				return nil, fmt.Errorf("cannot resolve inner agent %q: %w", innerCP.AgentName, err)
 			}
+			innerAgent = applyCheckpointConfig(innerAgent, innerCP.Config)
 
 			innerIE := &InterruptedError{
 				ToolCalls:        innerCP.PendingToolCalls,

pkg/agent/restore_test.go

@@ -461,4 +461,100 @@ func TestRestore(t *testing.T) {
 			assert.Contains(t, err.Error(), "unknown checkpoint status")
 		},
 	)
+
+	t.Run(
+		"buildCheckpoint captures MaxTurns in config snapshot",
+		func(t *testing.T) {
+			t.Parallel()
+
+			ag := agent.New(
+				"producer-agent",
+				newTestClient(&mockProvider{}),
+				agent.WithModel("test-model"),
+				agent.WithMaxTurns(7),
+			)
+
+			store := newMemoryCheckpointer()
+
+			stopCh := make(chan struct{})
+			close(stopCh)
+			ctx := agent.WithStopSignal(context.Background(), stopCh)
+
+			_, err := ag.RunWithOpts(
+				ctx,
+				[]llm.Message{
+					{
+						Role:  llm.RoleUser,
+						Parts: []llm.Part{llm.TextPart{Text: "begin"}},
+					},
+				},
+				agent.WithCheckpointer(store, "run-save-side"),
+			)
+
+			var se *agent.SuspendedError
+			require.ErrorAs(t, err, &se)
+
+			cp, loadErr := store.Load(context.Background(), "run-save-side")
+			require.NoError(t, loadErr)
+			require.NotNil(t, cp)
+			assert.Equal(t, 7, cp.Config.MaxTurns)
+		},
+	)
+
+	t.Run(
+		"checkpoint config supersedes live agent config on restore",
+		func(t *testing.T) {
+			t.Parallel()
+
+			provider := &mockProvider{
+				responses: []*llm.ChatCompletionResponse{
+					stopResponse("Completed after resume."),
+				},
+			}
+
+			// Agent registered at restore time has a bound tighter
+			// than the count of turns already taken. Without the
+			// checkpoint config snapshot, coreLoop would immediately
+			// trip MaxTurnsExceededError on its first iteration.
+			restoreAgent := agent.New(
+				"test-agent",
+				newTestClient(provider),
+				agent.WithInstructions("Test."),
+				agent.WithModel("test-model"),
+				agent.WithMaxTurns(5),
+			)
+
+			store := newMemoryCheckpointer()
+			err := store.Save(context.Background(), "run-config", &agent.Checkpoint{
+				Status:    agent.AgentStatusSuspended,
+				AgentName: "test-agent",
+				Config:    agent.AgentConfig{MaxTurns: 20},
+				Messages: []llm.Message{
+					{
+						Role:  llm.RoleUser,
+						Parts: []llm.Part{llm.TextPart{Text: "hi"}},
+					},
+				},
+				Turns: 15,
+			})
+			require.NoError(t, err)
+
+			registry := &simpleRegistry{
+				agents: map[string]*agent.Agent{
+					"test-agent": restoreAgent,
+				},
+			}
+
+			result, err := agent.Restore(
+				context.Background(),
+				store,
+				"run-config",
+				registry,
+			)
+
+			require.NoError(t, err)
+			require.NotNil(t, result)
+			assert.Equal(t, "Completed after resume.", result.FinalMessage().Text())
+		},
+	)
 }

pkg/agent/run.go

@@ -227,8 +227,11 @@ func (s *loopState) buildCheckpoint(status AgentStatus) *Checkpoint {
 	copy(msgsCopy, s.messages)
 
 	return &Checkpoint{
-		Status:        status,
-		AgentName:     s.agent.name,
+		Status:    status,
+		AgentName: s.agent.name,
+		Config: AgentConfig{
+			MaxTurns: s.agent.maxTurns,
+		},
 		Messages:      msgsCopy,
 		Usage:         s.totalUsage,
 		Turns:         s.turns,