diff --git a/cmd/gateway_errors.go b/cmd/gateway_errors.go
index 76b676bc..29cd3639 100644
--- a/cmd/gateway_errors.go
+++ b/cmd/gateway_errors.go
@@ -70,6 +70,14 @@ func isContextOverflowError(lower string) bool {
 		"prompt is too long",
 		"exceeds model context window",
 		"request exceeds the maximum size",
+		// Issue 958: Additional patterns (sync with providers/error_classify.go)
+		"prompt exceeds max length", // ZAI/GLM-5
+		"input is too long",         // DashScope
+		"token limit",
+		"too many tokens",
+		"请求输入过长",       // Chinese generic
+		"超出最大长度限制",     // Chinese Qwen
+		"上下文长度",        // Chinese context length
 	) || (strings.Contains(lower, "context") &&
 		containsAny(lower, "overflow", "too large", "too long", "limit", "exceeded"))
 }
diff --git a/internal/agent/loop_pipeline_adapter.go b/internal/agent/loop_pipeline_adapter.go
index 5ae4703a..92f0fe6f 100644
--- a/internal/agent/loop_pipeline_adapter.go
+++ b/internal/agent/loop_pipeline_adapter.go
@@ -58,6 +58,7 @@ func (l *Loop) buildPipelineDeps(req *RunRequest, bridgeRS *runState) pipeline.P
 			CheckpointInterval: 5,
 			ContextWindow:      l.contextWindow,
 			MaxTokens:          l.effectiveMaxTokens(),
+			ReserveTokens:      l.resolveReserveTokens(),
 			Compaction:         l.compactionCfg,
 			// V3 memory/retrieval flags removed — always true at runtime.
 		},
diff --git a/internal/agent/loop_types.go b/internal/agent/loop_types.go
index a9e72e07..59a5dba8 100644
--- a/internal/agent/loop_types.go
+++ b/internal/agent/loop_types.go
@@ -452,6 +452,15 @@ func (l *Loop) effectiveMaxTokens() int {
 	return defaultMaxTokens
 }
 
+// resolveReserveTokens returns the reserve token buffer from compaction config.
+// Issue 958: Wire ReserveTokensFloor to prevent context overflow before compaction.
+func (l *Loop) resolveReserveTokens() int {
+	if l.compactionCfg != nil && l.compactionCfg.ReserveTokensFloor > 0 {
+		return l.compactionCfg.ReserveTokensFloor
+	}
+	return 0
+}
+
 func NewLoop(cfg LoopConfig) *Loop {
 	if cfg.MaxIterations <= 0 {
 		cfg.MaxIterations = config.DefaultMaxIterations
diff --git a/internal/channels/errors.go b/internal/channels/errors.go
new file mode 100644
index 00000000..b33fe7f9
--- /dev/null
+++ b/internal/channels/errors.go
@@ -0,0 +1,45 @@
+package channels
+
+import (
+	"strings"
+
+	"github.com/nextlevelbuilder/goclaw/internal/providers"
+)
+
+// FormatAgentError converts internal error to user-friendly message.
+// Issue 958: Send user-friendly error on RunFailed instead of silent "...".
+func FormatAgentError(errStr string) string {
+	if errStr == "" {
+		return ""
+	}
+
+	lower := strings.ToLower(errStr)
+
+	// Context overflow (highest priority — specific actionable message)
+	if providers.IsContextOverflowMessage(lower) {
+		return "⚠️ The conversation has grown too long. Please start a new chat or ask me to summarize."
+	}
+
+	// Rate limit
+	if strings.Contains(lower, "rate limit") || strings.Contains(lower, "too many requests") || strings.Contains(lower, "429") {
+		return "⏳ Too many requests. Please wait a moment and try again."
+	}
+
+	// Auth errors
+	if strings.Contains(lower, "unauthorized") || strings.Contains(lower, "invalid api key") || strings.Contains(lower, "401") || strings.Contains(lower, "403") {
+		return "🔑 Authentication error. Please check your API configuration."
+	}
+
+	// Timeout
+	if strings.Contains(lower, "timeout") || strings.Contains(lower, "deadline exceeded") {
+		return "⏱️ Request timed out. Please try again."
+	}
+
+	// Overloaded
+	if strings.Contains(lower, "overload") {
+		return "🔄 Service is busy. Please try again in a moment."
+	}
+
+	// Generic fallback (don't expose internal error details)
+	return "❌ Something went wrong. Please try again."
+}
diff --git a/internal/channels/errors_test.go b/internal/channels/errors_test.go
new file mode 100644
index 00000000..84d14a9f
--- /dev/null
+++ b/internal/channels/errors_test.go
@@ -0,0 +1,65 @@
+package channels
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestFormatAgentError_ContextOverflow(t *testing.T) {
+	t.Parallel()
+	testCases := []string{
+		"context length exceeded",
+		"Prompt exceeds max length",
+		"request_too_large: payload too big",
+		"Input is too long for this model",
+		"token limit exceeded",
+		"请求输入过长",
+	}
+
+	for _, tc := range testCases {
+		result := FormatAgentError(tc)
+		if !strings.Contains(result, "conversation has grown too long") {
+			t.Errorf("expected context overflow message for %q, got %q", tc, result)
+		}
+	}
+}
+
+func TestFormatAgentError_RateLimit(t *testing.T) {
+	t.Parallel()
+	result := FormatAgentError("rate limit exceeded")
+	if !strings.Contains(result, "Too many requests") {
+		t.Errorf("unexpected rate limit message: %s", result)
+	}
+}
+
+func TestFormatAgentError_Auth(t *testing.T) {
+	t.Parallel()
+	result := FormatAgentError("unauthorized access")
+	if !strings.Contains(result, "Authentication error") {
+		t.Errorf("unexpected auth message: %s", result)
+	}
+}
+
+func TestFormatAgentError_Timeout(t *testing.T) {
+	t.Parallel()
+	result := FormatAgentError("request timeout")
+	if !strings.Contains(result, "timed out") {
+		t.Errorf("unexpected timeout message: %s", result)
+	}
+}
+
+func TestFormatAgentError_Generic(t *testing.T) {
+	t.Parallel()
+	result := FormatAgentError("some unknown error")
+	if !strings.Contains(result, "Something went wrong") {
+		t.Errorf("unexpected generic message: %s", result)
+	}
+}
+
+func TestFormatAgentError_Empty(t *testing.T) {
+	t.Parallel()
+	result := FormatAgentError("")
+	if result != "" {
+		t.Errorf("expected empty string for empty error, got %q", result)
+	}
+}
diff --git a/internal/channels/events.go b/internal/channels/events.go
index 5f5fe005..841b0fed 100644
--- a/internal/channels/events.go
+++ b/internal/channels/events.go
@@ -234,8 +234,27 @@ func (m *Manager) HandleAgentEvent(eventType, runID string, payload any) {
 				}
 				sc.FinalizeStream(ctx, rc.ChatID, currentStream)
 			}
-		case protocol.AgentEventRunFailed, protocol.AgentEventRunCancelled:
-			// Clean up streaming state on failure or cancellation
+		case protocol.AgentEventRunFailed:
+			// Clean up streaming state on failure
+			rc.mu.Lock()
+			currentStream := rc.stream
+			rc.stream = nil
+			rc.mu.Unlock()
+			if currentStream != nil {
+				_ = currentStream.Stop(ctx)
+			}
+			// Issue 958: Send user-friendly error message instead of silent "..."
+			errStr := extractPayloadString(payload, "error")
+			if friendlyMsg := FormatAgentError(errStr); friendlyMsg != "" {
+				m.bus.PublishOutbound(bus.OutboundMessage{
+					Channel:  rc.ChannelName,
+					ChatID:   rc.ChatID,
+					Content:  friendlyMsg,
+					TenantID: rc.TenantID,
+				})
+			}
+		case protocol.AgentEventRunCancelled:
+			// Clean up streaming state on cancellation
 			rc.mu.Lock()
 			currentStream := rc.stream
 			rc.stream = nil
diff --git a/internal/gateway/methods/sessions.go b/internal/gateway/methods/sessions.go
index 2f66b286..8f0bb321 100644
--- a/internal/gateway/methods/sessions.go
+++ b/internal/gateway/methods/sessions.go
@@ -30,6 +30,7 @@ func (m *SessionsMethods) Register(router *gateway.MethodRouter) {
 	router.Register(protocol.MethodSessionsPatch, m.handlePatch)
 	router.Register(protocol.MethodSessionsDelete, m.handleDelete)
 	router.Register(protocol.MethodSessionsReset, m.handleReset)
+	router.Register(protocol.MethodSessionsCompact, m.handleCompact)
 }
 
 type sessionsListParams struct {
@@ -230,3 +231,65 @@ func (m *SessionsMethods) handleReset(ctx context.Context, client *gateway.Clien
 	}))
 	emitAudit(m.eventBus, client, "session.reset", "session", params.Key)
 }
+
+type sessionCompactParams struct {
+	Key      string `json:"key"`
+	KeepLast int    `json:"keepLast,omitempty"` // default 4
+}
+
+// handleCompact truncates session history to the last N messages.
+// Issue 958: Manual session compaction API (truncate-only, no LLM summarization).
+func (m *SessionsMethods) handleCompact(ctx context.Context, client *gateway.Client, req *protocol.RequestFrame) {
+	locale := store.LocaleFromContext(ctx)
+	var params sessionCompactParams
+	if err := json.Unmarshal(req.Params, &params); err != nil {
+		client.SendResponse(protocol.NewErrorResponse(req.ID, protocol.ErrInvalidRequest, i18n.T(locale, i18n.MsgInvalidJSON)))
+		return
+	}
+
+	if params.Key == "" {
+		client.SendResponse(protocol.NewErrorResponse(req.ID, protocol.ErrInvalidRequest, "key is required"))
+		return
+	}
+
+	keepLast := params.KeepLast
+	if keepLast <= 0 {
+		keepLast = 4 // default: keep last 2 exchanges
+	}
+
+	// Auth check
+	sess := m.sessions.Get(ctx, params.Key)
+	if sess == nil {
+		client.SendResponse(protocol.NewErrorResponse(req.ID, protocol.ErrNotFound, i18n.T(locale, i18n.MsgNotFound, "session", params.Key)))
+		return
+	}
+	if !canSeeAll(client.Role(), m.cfg.Gateway.OwnerIDs, client.UserID()) {
+		if sess.UserID != client.UserID() {
+			client.SendResponse(protocol.NewErrorResponse(req.ID, protocol.ErrUnauthorized, i18n.T(locale, i18n.MsgPermissionDenied, "session")))
+			return
+		}
+	}
+
+	history := m.sessions.GetHistory(ctx, params.Key)
+	originalLen := len(history)
+	if originalLen < 6 {
+		client.SendResponse(protocol.NewOKResponse(req.ID, map[string]any{
+			"ok":      true,
+			"message": "session too short to compact",
+			"kept":    originalLen,
+		}))
+		return
+	}
+
+	// Truncate history to last N messages
+	m.sessions.TruncateHistory(ctx, params.Key, keepLast)
+	m.sessions.IncrementCompaction(ctx, params.Key)
+	m.sessions.Save(ctx, params.Key)
+
+	client.SendResponse(protocol.NewOKResponse(req.ID, map[string]any{
+		"ok":       true,
+		"original": originalLen,
+		"kept":     keepLast,
+	}))
+	emitAudit(m.eventBus, client, "session.compacted", "session", params.Key)
+}
diff --git a/internal/pipeline/stages_test.go b/internal/pipeline/stages_test.go
index 89aab466..94c25fda 100644
--- a/internal/pipeline/stages_test.go
+++ b/internal/pipeline/stages_test.go
@@ -336,6 +336,118 @@ func TestThinkStage_LLMError_Propagates(t *testing.T) {
 	}
 }
 
+// Issue 958: Context overflow triggers emergency compaction + retry
+
+func TestThinkStage_ContextOverflow_TriggersCompaction(t *testing.T) {
+	t.Parallel()
+	callCount := 0
+	compacted := false
+
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxIterations: 10, MaxTokens: 1000},
+		CallLLM: func(_ context.Context, _ *RunState, _ providers.ChatRequest) (*providers.ChatResponse, error) {
+			callCount++
+			if callCount == 1 {
+				return nil, &providers.HTTPError{Status: 400, Body: "Prompt exceeds max length"}
+			}
+			return &providers.ChatResponse{Content: "success after compact", FinishReason: "stop"}, nil
+		},
+		CompactMessages: func(_ context.Context, msgs []providers.Message, _ string) ([]providers.Message, error) {
+			compacted = true
+			return []providers.Message{{Role: "user", Content: "[Summary]"}}, nil
+		},
+	}
+
+	stage := NewThinkStage(deps)
+	state := defaultState()
+	state.Messages.SetHistory([]providers.Message{{Role: "user", Content: "test"}})
+
+	// First call: overflow → compact → retry
+	err := stage.Execute(context.Background(), state)
+	if err != nil {
+		t.Fatalf("first Execute() should trigger retry, got error: %v", err)
+	}
+	if !compacted {
+		t.Error("expected compaction to be triggered")
+	}
+	if state.Think.OverflowRetries != 1 {
+		t.Errorf("expected OverflowRetries=1, got %d", state.Think.OverflowRetries)
+	}
+	// Stage returns Continue (nil error) to signal retry this iteration
+	if stage.Result() != Continue {
+		t.Errorf("Result() = %v after compaction, want Continue", stage.Result())
+	}
+}
+
+func TestThinkStage_ContextOverflow_FailsAfterOneRetry(t *testing.T) {
+	t.Parallel()
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxIterations: 10, MaxTokens: 1000},
+		CallLLM: func(_ context.Context, _ *RunState, _ providers.ChatRequest) (*providers.ChatResponse, error) {
+			return nil, &providers.HTTPError{Status: 400, Body: "Prompt exceeds max length"}
+		},
+		CompactMessages: func(_ context.Context, _ []providers.Message, _ string) ([]providers.Message, error) {
+			return []providers.Message{{Role: "user", Content: "[Summary]"}}, nil
+		},
+	}
+
+	stage := NewThinkStage(deps)
+	state := defaultState()
+	state.Think.OverflowRetries = 1 // Already retried once
+
+	err := stage.Execute(context.Background(), state)
+	if err == nil {
+		t.Error("expected error after second overflow")
+	}
+	if !strings.Contains(err.Error(), "context overflow after compaction") {
+		t.Errorf("expected 'context overflow after compaction' message, got %v", err)
+	}
+}
+
+func TestThinkStage_ContextOverflow_NoCompactCallback_FailsGracefully(t *testing.T) {
+	t.Parallel()
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxIterations: 10, MaxTokens: 1000},
+		CallLLM: func(_ context.Context, _ *RunState, _ providers.ChatRequest) (*providers.ChatResponse, error) {
+			return nil, &providers.HTTPError{Status: 400, Body: "Prompt exceeds max length"}
+		},
+		CompactMessages: nil, // No compaction available
+	}
+
+	stage := NewThinkStage(deps)
+	state := defaultState()
+
+	err := stage.Execute(context.Background(), state)
+	if err == nil {
+		t.Error("expected error when no compaction available")
+	}
+}
+
+func TestThinkStage_ContextOverflow_CompactionFails_ReturnsOriginalError(t *testing.T) {
+	t.Parallel()
+	deps := &PipelineDeps{
+		Config: PipelineConfig{MaxIterations: 10, MaxTokens: 1000},
+		CallLLM: func(_ context.Context, _ *RunState, _ providers.ChatRequest) (*providers.ChatResponse, error) {
+			return nil, &providers.HTTPError{Status: 400, Body: "Prompt exceeds max length"}
+		},
+		CompactMessages: func(_ context.Context, _ []providers.Message, _ string) ([]providers.Message, error) {
+			return nil, errors.New("compaction failed")
+		},
+	}
+
+	stage := NewThinkStage(deps)
+	state := defaultState()
+
+	err := stage.Execute(context.Background(), state)
+	if err == nil {
+		t.Error("expected error when compaction fails")
+	}
+	// Should return LLM error wrapped, not compaction error
+	if !strings.Contains(err.Error(), "llm call") {
+		t.Errorf("expected 'llm call' in error message, got %v", err)
+	}
+}
+
 // --- PruneStage tests ---
 
 func TestPruneStage_UnderBudget_NoOp(t *testing.T) {
diff --git a/internal/pipeline/substates.go b/internal/pipeline/substates.go
index 6bc3f0e5..a0e56a91 100644
--- a/internal/pipeline/substates.go
+++ b/internal/pipeline/substates.go
@@ -32,6 +32,7 @@ type ThinkState struct {
 	LastResponse    *providers.ChatResponse
 	TotalUsage      providers.Usage
 	TruncRetries    int  // consecutive truncation retries (max 3)
+	OverflowRetries int  // context overflow compact+retry attempts (max 1)
 	StreamingActive bool // true during active stream
 }
 
diff --git a/internal/pipeline/think_stage.go b/internal/pipeline/think_stage.go
index f80afb1f..582a79a2 100644
--- a/internal/pipeline/think_stage.go
+++ b/internal/pipeline/think_stage.go
@@ -3,6 +3,8 @@ package pipeline
 import (
 	"context"
 	"fmt"
+	"log/slog"
+	"strings"
 
 	"github.com/nextlevelbuilder/goclaw/internal/providers"
 )
@@ -57,6 +59,28 @@ func (s *ThinkStage) Execute(ctx context.Context, state *RunState) error {
 	}
 	resp, err := s.deps.CallLLM(ctx, state, req)
 	if err != nil {
+		// Issue 958: Check for context overflow — attempt emergency compaction + retry
+		if isContextOverflowErr(err) {
+			if state.Think.OverflowRetries > 0 {
+				return fmt.Errorf("context overflow after compaction: %w", err)
+			}
+			state.Think.OverflowRetries++
+			// Attempt emergency compaction
+			if s.deps.CompactMessages != nil {
+				originalLen := len(state.Messages.History())
+				compacted, compactErr := s.deps.CompactMessages(ctx, state.Messages.History(), state.Model)
+				if compactErr == nil {
+					state.Messages.ReplaceHistory(compacted)
+					slog.Info("emergency_compaction_triggered",
+						"run_id", state.RunID,
+						"original_msgs", originalLen,
+						"compacted_msgs", len(compacted),
+					)
+					return nil // Retry this iteration (Continue result)
+				}
+				slog.Warn("emergency_compaction_failed", "error", compactErr)
+			}
+		}
 		return fmt.Errorf("llm call: %w", err)
 	}
 	state.Think.LastResponse = resp
@@ -92,7 +116,8 @@ func (s *ThinkStage) Execute(ctx context.Context, state *RunState) error {
 		state.Messages.AppendPending(providers.Message{Role: "user", Content: hint})
 		return nil // Continue to next iteration for retry
 	}
-	state.Think.TruncRetries = 0 // reset on success
+	state.Think.TruncRetries = 0    // reset on success
+	state.Think.OverflowRetries = 0 // reset on success
 
 	// 7. Uniquify tool call IDs (OpenAI returns 400 on duplicates across iterations).
 	// Skip if raw content present (Anthropic thinking passback) to avoid desync.
@@ -192,3 +217,13 @@ func toolCallsHaveMissingRequiredArgs(calls []providers.ToolCall) bool {
 	}
 	return false
 }
+
+// isContextOverflowErr checks if an error indicates context window overflow.
+// Uses the exported helper from providers package for pattern matching.
+func isContextOverflowErr(err error) bool {
+	if err == nil {
+		return false
+	}
+	lower := strings.ToLower(err.Error())
+	return providers.IsContextOverflowMessage(lower)
+}
diff --git a/internal/providers/error_classify.go b/internal/providers/error_classify.go
index 88bdf652..ec73e9ed 100644
--- a/internal/providers/error_classify.go
+++ b/internal/providers/error_classify.go
@@ -155,9 +155,19 @@ func isContextOverflow(lower string) bool {
 		// Chinese patterns (Qwen/DashScope)
 		"超出最大长度限制",
 		"上下文长度",
+		// Issue 958: Additional patterns
+		"prompt exceeds max length", // ZAI/GLM-5
+		"request_too_large",         // Generic
+		"input is too long",         // DashScope
+		"请求输入过长",                    // Chinese generic
 	)
 }
 
+// IsContextOverflowMessage exports overflow detection for use by pipeline.
+func IsContextOverflowMessage(lower string) bool {
+	return isContextOverflow(lower)
+}
+
 // isNetworkError checks if an error is a network-level failure.
 func isNetworkError(err error) bool {
 	if err == nil {
diff --git a/internal/providers/error_classify_test.go b/internal/providers/error_classify_test.go
index 2de37022..7ec52120 100644
--- a/internal/providers/error_classify_test.go
+++ b/internal/providers/error_classify_test.go
@@ -324,3 +324,37 @@ func TestClassifyUnknownError(t *testing.T) {
 		t.Errorf("expected FailoverUnknown, got %s", result.Reason)
 	}
 }
+
+// Issue 958: New context overflow patterns for ZAI/GLM, DashScope, generic
+
+func TestClassifyPromptExceedsMaxLength(t *testing.T) {
+	classifier := NewDefaultClassifier()
+	result := classifier.Classify(nil, 400, `{"error":{"code":"1261","message":"Prompt exceeds max length"}}`)
+	if result.Kind != "context_overflow" {
+		t.Errorf("expected context_overflow, got %s (reason: %s)", result.Kind, result.Reason)
+	}
+}
+
+func TestClassifyInputTooLong(t *testing.T) {
+	classifier := NewDefaultClassifier()
+	result := classifier.Classify(nil, 400, "Input is too long for this model")
+	if result.Kind != "context_overflow" {
+		t.Errorf("expected context_overflow, got %s (reason: %s)", result.Kind, result.Reason)
+	}
+}
+
+func TestClassifyRequestTooLarge(t *testing.T) {
+	classifier := NewDefaultClassifier()
+	result := classifier.Classify(nil, 400, "request_too_large: payload exceeds limit")
+	if result.Kind != "context_overflow" {
+		t.Errorf("expected context_overflow, got %s (reason: %s)", result.Kind, result.Reason)
+	}
+}
+
+func TestClassifyChineseInputTooLong(t *testing.T) {
+	classifier := NewDefaultClassifier()
+	result := classifier.Classify(nil, 400, "请求输入过长")
+	if result.Kind != "context_overflow" {
+		t.Errorf("expected context_overflow, got %s (reason: %s)", result.Kind, result.Reason)
+	}
+}
diff --git a/pkg/protocol/methods.go b/pkg/protocol/methods.go
index c4a2a787..c57e35f6 100644
--- a/pkg/protocol/methods.go
+++ b/pkg/protocol/methods.go
@@ -39,6 +39,7 @@ const (
 	MethodSessionsPatch   = "sessions.patch"
 	MethodSessionsDelete  = "sessions.delete"
 	MethodSessionsReset   = "sessions.reset"
+	MethodSessionsCompact = "sessions.compact"
 
 	// System
 	MethodConnect = "connect"