diff --git a/internal/engine/capacity_pause.go b/internal/engine/capacity_pause.go
index 8637db2..6ce7115 100644
--- a/internal/engine/capacity_pause.go
+++ b/internal/engine/capacity_pause.go
@@ -13,7 +13,7 @@ import (
 // capacity / session-limit exhaustion at a given pipeline stage. The "capacity
 // limit" phrasing routes pauseResumeHint to the correct operator guidance.
 func capacityPauseReason(stage string, err error) string {
-	return fmt.Sprintf("LLM capacity/session limit during %s — resume after reset: %v", stage, err)
+	return fmt.Sprintf("transient LLM capacity/network error during %s — resume after it clears: %v", stage, err)
 }
 
 // pauseIfCapacity inspects an LLM-call error. If it is a transient capacity
diff --git a/internal/llm/capacity_test.go b/internal/llm/capacity_test.go
index 2408746..6c613c5 100644
--- a/internal/llm/capacity_test.go
+++ b/internal/llm/capacity_test.go
@@ -26,6 +26,12 @@ func TestIsCapacityError(t *testing.T) {
 		{name: "rate limit string", err: fmt.Errorf("claude CLI error: rate limit exceeded"), expect: true},
 		{name: "too many requests string", err: fmt.Errorf("too many requests"), expect: true},
 		{name: "overloaded string", err: fmt.Errorf("the service is currently overloaded"), expect: true},
+		// Transient network/transport failures (api_error_status null) — must
+		// also classify as transient so they take the clean-pause path.
+		{name: "socket closed string", err: fmt.Errorf(`claude CLI error: exit status 1 (output: {"is_error":true,"api_error_status":null,"result":"API Error: The socket connection was closed unexpectedly"})`), expect: true},
+		{name: "connection reset", err: fmt.Errorf("read tcp: connection reset by peer"), expect: true},
+		{name: "i/o timeout", err: fmt.Errorf("dial tcp: i/o timeout"), expect: true},
+		{name: "503 service unavailable", err: fmt.Errorf("503 service unavailable"), expect: true},
 		{name: "api_error_status 429 embedded", err: fmt.Errorf(`output: {"api_error_status":429}`), expect: true},
 		{name: "api_error_status 529 embedded", err: fmt.Errorf(`output: {"api_error_status":529}`), expect: true},
 
diff --git a/internal/llm/errors.go b/internal/llm/errors.go
index 722340e..9439a52 100644
--- a/internal/llm/errors.go
+++ b/internal/llm/errors.go
@@ -77,6 +77,21 @@ var capacitySignatures = []string{
 	`"api_error_status":529`,
 	`"api_error_status": 429`,
 	`"api_error_status": 529`,
+
+	// Transient network/transport failures. Like a session limit, these are not
+	// a story-quality problem and succeed on retry/resume — so they must take
+	// the clean-pause path, not burn the escalation chain. Surfaced by the CLI
+	// as e.g. "API Error: The socket connection was closed unexpectedly".
+	"socket connection was closed",
+	"connection closed unexpectedly",
+	"connection reset",
+	"connection refused",
+	"i/o timeout",
+	"tls handshake timeout",
+	"unexpected eof",
+	"service unavailable",
+	"bad gateway",
+	"gateway timeout",
 }
 
 // ContainsCapacitySignature reports whether a raw string carries a capacity /